axon

package
v1.8.20 Latest Latest
Warning

This package is not in the latest version of its module.

 Go to latest Published: Oct 25, 2023 License: BSD-3-Clause Imports: 71 Imported by: 34

Documentation

Overview

Package axon provides the basic reference axon implementation, for rate-coded activations and standard error-driven learning. Other packages provide spiking or deep axon, PVLV, PBWM, etc.

The overall design seeks an "optimal" tradeoff between simplicity, transparency, ability to flexibly recombine and extend elements, and avoiding having to rewrite a bunch of stuff.

The *Stru elements handle the core structural components of the network, and hold emer.* interface pointers to elements such as emer.Layer, which provides a very minimal interface for these elements. Interfaces are automatically pointers, so think of these as generic pointers to your specific Layers etc.

This design means the same *Stru infrastructure can be re-used across different variants of the algorithm. Because we're keeping this infrastructure minimal and algorithm-free it should be much less confusing than dealing with the multiple levels of inheritance in C++ emergent. The actual algorithm-specific code is now fully self-contained, and largely orthogonalized from the infrastructure.

One specific cost of this is the need to cast the emer.* interface pointers into the specific types of interest, when accessing via the *Stru infrastructure.

The *Params elements contain all the (meta)parameters and associated methods for computing various functions. They are the equivalent of Specs from original emergent, but unlike specs they are local to each place they are used, and styling is used to apply common parameters across multiple layers etc. Params seems like a more explicit, recognizable name compared to specs, and this also helps avoid confusion about their different nature than old specs. Pars is shorter but confusable with "Parents" so "Params" is more unambiguous.

Params are organized into four major categories, which are more clearly functionally labeled as opposed to just structurally so, to keep things clearer and better organized overall: * ActParams -- activation params, at the Neuron level (in act.go) * InhibParams -- inhibition params, at the Layer / Pool level (in inhib.go) * LearnNeurParams -- learning parameters at the Neuron level (running-averages that drive learning) * LearnSynParams -- learning parameters at the Synapse level (both in learn.go)

The levels of structure and state are: * Network * .Layers * .Pools: pooled inhibition state -- 1 for layer plus 1 for each sub-pool (unit group) with inhibition * .RecvPrjns: receiving projections from other sending layers * .SendPrjns: sending projections from other receiving layers * .Neurons: neuron state variables

There are methods on the Network that perform initialization and overall computation, by iterating over layers and calling methods there. This is typically how most users will run their models.

Parallel computation across multiple CPU cores (threading) is achieved through persistent worker go routines that listen for functions to run on thread-specific channels. Each layer has a designated thread number, so you can experiment with different ways of dividing up the computation. Timing data is kept for per-thread time use -- see TimeReport() on the network.

The Layer methods directly iterate over Neurons, Pools, and Prjns, and there is no finer-grained level of computation (e.g., at the individual Neuron level), except for the *Params methods that directly compute relevant functions. Thus, looking directly at the layer.go code should provide a clear sense of exactly how everything is computed -- you may need to the refer to act.go, learn.go etc to see the relevant details but at least the overall organization should be clear in layer.go.

Computational methods are generally named: VarFmVar to specifically name what variable is being computed from what other input variables. e.g., SpikeFmG computes activation from conductances G.

The Pools (type Pool, in pool.go) hold state used for computing pooled inhibition, but also are used to hold overall aggregate pooled state variables -- the first element in Pools applies to the layer itself, and subsequent ones are for each sub-pool (4D layers). These pools play the same role as the AxonUnGpState structures in C++ emergent.

Prjns directly support all synapse-level computation, and hold the LearnSynParams and iterate directly over all of their synapses. It is the exact same Prjn object that lives in the RecvPrjns of the receiver-side, and the SendPrjns of the sender-side, and it maintains and coordinates both sides of the state. This clarifies and simplifies a lot of code. There is no separate equivalent of AxonConSpec / AxonConState at the level of connection groups per unit per projection.

The pattern of connectivity between units is specified by the prjn.Pattern interface and all the different standard options are avail in that prjn package. The Pattern code generates a full tensor bitmap of binary 1's and 0's for connected (1's) and not (0's) units, and can use any method to do so. This full lookup-table approach is not the most memory-efficient, but it is fully general and shouldn't be too-bad memory-wise overall (fully bit-packed arrays are used, and these bitmaps don't need to be retained once connections have been established). This approach allows patterns to just focus on patterns, and they don't care at all how they are used to allocate actual connections.

Index

Constants

View Source 
const (
	Version     = "v1.8.20"
	GitCommit   = "03242218"         // the commit JUST BEFORE the release
	VersionDate = "2023-10-25 20:13" // UTC
)
View Source 
const CyclesN = 10

CyclesN is the number of cycles to run as a group for ra25, 10 = ~50 msec / trial, 25 = ~48, all 150 / 50 minus / plus = ~44 10 is good enough and unlikely to mess with anything else..

Variables

View Source 
var (
	AvgMaxFloatFromIntErr   func()
	AvgMaxFloatFromIntErrMu sync.Mutex
)

AvgMaxFloatFromIntErr is called when there is an overflow error in AvgMaxI32 FloatFromInt

View Source 
var (
	NeuronVarNames []string
	NeuronVarsMap  map[string]int
)
View Source 
var (
	NeuronLayerVars  = []string{"DA", "ACh", "NE", "Ser", "Gated"}
	NNeuronLayerVars = len(NeuronLayerVars)
)

NeuronLayerVars are layer-level variables displayed as neuron layers.

View Source 
var (
	SynapseVarNames []string
	SynapseVarsMap  map[string]int
)
View Source 
var KiT_DAModTypes = kit.Enums.AddEnum(DAModTypesN, kit.NotBitFlag, nil)
View Source 
var KiT_GPLayerTypes = kit.Enums.AddEnum(GPLayerTypesN, kit.NotBitFlag, nil)
View Source 
var KiT_GlobalVars = kit.Enums.AddEnum(GlobalVarsN, kit.NotBitFlag, nil)
View Source 
var KiT_Layer = kit.Types.AddType(&Layer{}, LayerProps)
View Source 
var KiT_LayerTypes = kit.Enums.AddEnum(LayerTypesN, kit.NotBitFlag, nil)
View Source 
var KiT_Network = kit.Types.AddType(&Network{}, NetworkProps)
View Source 
var KiT_NeuronAvgVars = kit.Enums.AddEnum(NeuronAvgVarsN, kit.NotBitFlag, nil)
View Source 
var KiT_NeuronIdxs = kit.Enums.AddEnum(NeuronIdxsN, kit.NotBitFlag, nil)
View Source 
var KiT_NeuronVars = kit.Enums.AddEnum(NeuronVarsN, kit.NotBitFlag, nil)
View Source 
var KiT_Prjn = kit.Types.AddType(&Prjn{}, PrjnProps)
View Source 
var KiT_PrjnGTypes = kit.Enums.AddEnum(PrjnGTypesN, kit.NotBitFlag, nil)
View Source 
var KiT_PrjnTypes = kit.Enums.AddEnum(PrjnTypesN, kit.NotBitFlag, nil)
View Source 
var KiT_SynapseCaVars = kit.Enums.AddEnum(SynapseCaVarsN, kit.NotBitFlag, nil)
View Source 
var KiT_SynapseIdxs = kit.Enums.AddEnum(SynapseIdxsN, kit.NotBitFlag, nil)
View Source 
var KiT_SynapseVars = kit.Enums.AddEnum(SynapseVarsN, kit.NotBitFlag, nil)
View Source 
var KiT_ValenceTypes = kit.Enums.AddEnum(ValenceTypesN, kit.NotBitFlag, nil)
View Source 
var LayerProps = ki.Props{
	"EnumType:Typ": KiT_LayerTypes,
	"ToolBar": ki.PropSlice{
		{"Defaults", ki.Props{
			"icon": "reset",
			"desc": "return all parameters to their intial default values",
		}},
		{"InitWts", ki.Props{
			"icon": "update",
			"desc": "initialize the layer's weight values according to prjn parameters, for all *sending* projections out of this layer",
		}},
		{"InitActs", ki.Props{
			"icon": "update",
			"desc": "initialize the layer's activation values",
		}},
		{"sep-act", ki.BlankProp{}},
		{"LesionNeurons", ki.Props{
			"icon": "close",
			"desc": "Lesion (set the Off flag) for given proportion of neurons in the layer (number must be 0 -- 1, NOT percent!)",
			"Args": ki.PropSlice{
				{"Proportion", ki.Props{
					"desc": "proportion (0 -- 1) of neurons to lesion",
				}},
			},
		}},
		{"UnLesionNeurons", ki.Props{
			"icon": "reset",
			"desc": "Un-Lesion (reset the Off flag) for all neurons in the layer",
		}},
	},
}
View Source 
var NetworkProps = ki.Props{
	"ToolBar": ki.PropSlice{
		{"SaveWtsJSON", ki.Props{
			"label": "Save Wts...",
			"icon":  "file-save",
			"desc":  "Save json-formatted weights",
			"Args": ki.PropSlice{
				{"Weights File Name", ki.Props{
					"default-field": "WtsFile",
					"ext":           ".wts,.wts.gz",
				}},
			},
		}},
		{"OpenWtsJSON", ki.Props{
			"label": "Open Wts...",
			"icon":  "file-open",
			"desc":  "Open json-formatted weights",
			"Args": ki.PropSlice{
				{"Weights File Name", ki.Props{
					"default-field": "WtsFile",
					"ext":           ".wts,.wts.gz",
				}},
			},
		}},
		{"sep-file", ki.BlankProp{}},
		{"Build", ki.Props{
			"icon": "update",
			"desc": "build the network's neurons and synapses according to current params",
		}},
		{"InitWts", ki.Props{
			"icon": "update",
			"desc": "initialize the network weight values according to prjn parameters",
		}},
		{"InitActs", ki.Props{
			"icon": "update",
			"desc": "initialize the network activation values",
		}},
		{"sep-act", ki.BlankProp{}},
		{"AddLayer", ki.Props{
			"label": "Add Layer...",
			"icon":  "new",
			"desc":  "add a new layer to network",
			"Args": ki.PropSlice{
				{"Layer Name", ki.Props{}},
				{"Layer Shape", ki.Props{
					"desc": "shape of layer, typically 2D (Y, X) or 4D (Pools Y, Pools X, Units Y, Units X)",
				}},
				{"Layer Type", ki.Props{
					"desc": "type of layer -- used for determining how inputs are applied",
				}},
			},
		}},
		{"ConnectLayerNames", ki.Props{
			"label": "Connect Layers...",
			"icon":  "new",
			"desc":  "add a new connection between layers in the network",
			"Args": ki.PropSlice{
				{"Send Layer Name", ki.Props{}},
				{"Recv Layer Name", ki.Props{}},
				{"Pattern", ki.Props{
					"desc": "pattern to connect with",
				}},
				{"Prjn Type", ki.Props{
					"desc": "type of projection -- direction, or other more specialized factors",
				}},
			},
		}},
		{"AllGlobals", ki.Props{
			"icon":        "file-sheet",
			"desc":        "Shows the values of all network Global variables, for debugging purposes",
			"show-return": true,
		}},
		{"AllPrjnScales", ki.Props{
			"icon":        "file-sheet",
			"desc":        "AllPrjnScales returns a listing of all PrjnScale parameters in the Network in all Layers, Recv projections.  These are among the most important and numerous of parameters (in larger networks) -- this helps keep track of what they all are set to.",
			"show-return": true,
		}},
	},
}
View Source 
var NeuronVarProps = map[string]string{

	"Spike":  `desc:"whether neuron has spiked or not on this cycle (0 or 1)"`,
	"Spiked": `desc:"1 if neuron has spiked within the last 10 cycles (msecs), corresponding to a nominal max spiking rate of 100 Hz, 0 otherwise -- useful for visualization and computing activity levels in terms of average spiked levels."`,
	"Act":    `desc:"rate-coded activation value reflecting instantaneous estimated rate of spiking, based on 1 / ISIAvg.  This drives feedback inhibition in the FFFB function (todo: this will change when better inhibition is implemented), and is integrated over time for ActInt which is then used for performance statistics and layer average activations, etc.  Should not be used for learning or other computations."`,
	"ActInt": `desc:"integrated running-average activation value computed from Act with time constant Act.Dt.IntTau, to produce a longer-term integrated value reflecting the overall activation state across the ThetaCycle time scale, as the overall response of network to current input state -- this is copied to ActM and ActP at the ends of the minus and plus phases, respectively, and used in computing performance-level statistics (which are typically based on ActM).  Should not be used for learning or other computations."`,
	"ActM":   `desc:"ActInt activation state at end of third quarter, representing the posterior-cortical minus phase activation -- used for statistics and monitoring network performance. Should not be used for learning or other computations."`,
	"ActP":   `desc:"ActInt activation state at end of fourth quarter, representing the posterior-cortical plus_phase activation -- used for statistics and monitoring network performance.  Should not be used for learning or other computations."`,
	"Ext":    `desc:"external input: drives activation of unit from outside influences (e.g., sensory input)"`,
	"Target": `desc:"target value: drives learning to produce this activation value"`,

	"Ge":     `range:"2" desc:"total excitatory conductance, including all forms of excitation (e.g., NMDA) -- does *not* include Gbar.E"`,
	"Gi":     `auto-scale:"+" desc:"total inhibitory synaptic conductance -- the net inhibitory input to the neuron -- does *not* include Gbar.I"`,
	"Gk":     `auto-scale:"+" desc:"total potassium conductance, typically reflecting sodium-gated potassium currents involved in adaptation effects -- does *not* include Gbar.K"`,
	"Inet":   `desc:"net current produced by all channels -- drives update of Vm"`,
	"Vm":     `min:"0" max:"1" desc:"membrane potential -- integrates Inet current over time"`,
	"VmDend": `min:"0" max:"1" desc:"dendritic membrane potential -- has a slower time constant, is not subject to the VmR reset after spiking"`,
	"ISI":    `auto-scale:"+" desc:"current inter-spike-interval -- counts up since last spike.  Starts at -1 when initialized."`,
	"ISIAvg": `auto-scale:"+" desc:"average inter-spike-interval -- average time interval between spikes, integrated with ISITau rate constant (relatively fast) to capture something close to an instantaneous spiking rate.  Starts at -1 when initialized, and goes to -2 after first spike, and is only valid after the second spike post-initialization."`,

	"CaSyn":   `desc:"spike-driven calcium trace for synapse-level Ca-driven learning: exponential integration of SpikeG * Spike at SynTau time constant (typically 30).  Synapses integrate send.CaSyn * recv.CaSyn across M, P, D time integrals for the synaptic trace driving credit assignment in learning. Time constant reflects binding time of Glu to NMDA and Ca buffering postsynaptically, and determines time window where pre * post spiking must overlap to drive learning."`,
	"CaSpkM":  `desc:"spike-driven calcium trace used as a neuron-level proxy for synpatic credit assignment factor based on continuous time-integrated spiking: exponential integration of SpikeG * Spike at MTau time constant (typically 5).  Simulates a calmodulin (CaM) like signal at the most abstract level."`,
	"CaSpkP":  `desc:"continuous cascaded integration of CaSpkM at PTau time constant (typically 40), representing neuron-level purely spiking version of plus, LTP direction of weight change and capturing the function of CaMKII in the Kinase learning rule. Used for specialized learning and computational functions, statistics, instead of Act."`,
	"CaSpkD":  `desc:"continuous cascaded integration CaSpkP at DTau time constant (typically 40), representing neuron-level purely spiking version of minus, LTD direction of weight change and capturing the function of DAPK1 in the Kinase learning rule. Used for specialized learning and computational functions, statistics, instead of Act."`,
	"CaSpkPM": `desc:"minus-phase snapshot of the CaSpkP value -- similar to ActM but using a more directly spike-integrated value."`,
	"CaLrn":   `desc:"recv neuron calcium signal used to drive temporal error difference component of standard learning rule, combining NMDA (NmdaCa) and spiking-driven VGCC (VgccCaInt) calcium sources (vs. CaSpk* which only reflects spiking component).  This is integrated into CaM, CaP, CaD, and temporal derivative is CaP - CaD (CaMKII - DAPK1).  This approximates the backprop error derivative on net input, but VGCC component adds a proportion of recv activation delta as well -- a balance of both works best.  The synaptic-level trace multiplier provides the credit assignment factor, reflecting coincident activity and potentially integrated over longer multi-trial timescales."`,
	"NrnCaM":  `desc:"integrated CaLrn at MTau timescale (typically 5), simulating a calmodulin (CaM) like signal, which then drives CaP, CaD for delta signal driving error-driven learning."`,
	"NrnCaP":  `desc:"cascaded integration of CaM at PTau time constant (typically 40), representing the plus, LTP direction of weight change and capturing the function of CaMKII in the Kinase learning rule."`,
	"NrnCaD":  `desc:"cascaded integratoin of CaP at DTau time constant (typically 40), representing the minus, LTD direction of weight change and capturing the function of DAPK1 in the Kinase learning rule."`,
	"CaDiff":  `desc:"difference between CaP - CaD -- this is the error signal that drives error-driven learning."`,
	"RLRate":  `auto-scale:"+" desc:"recv-unit based learning rate multiplier, reflecting the sigmoid derivative computed from the CaSpkD of recv unit, and the normalized difference CaSpkP - CaSpkD / MAX(CaSpkP - CaSpkD)."`,
	"Attn":    `desc:"Attentional modulation factor, which can be set by special layers such as the TRC -- multiplies Ge"`,

	"SpkMaxCa": `desc:"Ca integrated like CaSpkP but only starting at MaxCycStart cycle, to prevent inclusion of carryover spiking from prior theta cycle trial -- the PTau time constant otherwise results in significant carryover.  This is the input to SpkMax"`,
	"SpkMax":   `desc:"maximum CaSpkP across one theta cycle time window (max of SpkMaxCa) -- used for specialized algorithms that have more phasic behavior within a single trial, e.g., BG Matrix layer gating.  Also useful for visualization of peak activity of neurons."`,
	"SpkPrv":   `desc:"final CaSpkD activation state at end of previous theta cycle.  used for specialized learning mechanisms that operate on delayed sending activations."`,
	"SpkSt1":   `desc:"the activation state at specific time point within current state processing window (e.g., 50 msec for beta cycle within standard theta cycle), as saved by SpkSt1() function.  Used for example in hippocampus for CA3, CA1 learning"`,
	"SpkSt2":   `desc:"the activation state at specific time point within current state processing window (e.g., 100 msec for beta cycle within standard theta cycle), as saved by SpkSt2() function.  Used for example in hippocampus for CA3, CA1 learning"`,
	"DASign":   `desc:"sign of dopamine-based learning effects for this neuron -- 1 = D1, -1 = D2"`,

	"GeNoiseP": `desc:"accumulating poisson probability factor for driving excitatory noise spiking -- multiply times uniform random deviate at each time step, until it gets below the target threshold based on lambda."`,
	"GeNoise":  `desc:"integrated noise excitatory conductance, added into Ge"`,
	"GiNoiseP": `desc:"accumulating poisson probability factor for driving inhibitory noise spiking -- multiply times uniform random deviate at each time step, until it gets below the target threshold based on lambda."`,
	"GiNoise":  `desc:"integrated noise inhibotyr conductance, added into Gi"`,

	"GeExt":     `desc:"extra excitatory conductance added to Ge -- from Ext input, GeCtxt etc"`,
	"GeRaw":     `desc:"raw excitatory conductance (net input) received from senders = current raw spiking drive"`,
	"GeSyn":     `range:"2" desc:"time-integrated total excitatory synaptic conductance, with an instantaneous rise time from each spike (in GeRaw) and exponential decay with Dt.GeTau, aggregated over projections -- does *not* include Gbar.E"`,
	"GiRaw":     `desc:"raw inhibitory conductance (net input) received from senders  = current raw spiking drive"`,
	"GiSyn":     `desc:"time-integrated total inhibitory synaptic conductance, with an instantaneous rise time from each spike (in GiRaw) and exponential decay with Dt.GiTau, aggregated over projections -- does *not* include Gbar.I.  This is added with computed FFFB inhibition to get the full inhibition in Gi"`,
	"GeInt":     `range:"2" desc:"integrated running-average activation value computed from Ge with time constant Act.Dt.IntTau, to produce a longer-term integrated value reflecting the overall Ge level across the ThetaCycle time scale (Ge itself fluctuates considerably) -- useful for stats to set strength of connections etc to get neurons into right range of overall excitatory drive"`,
	"GeIntNorm": `range:"1" desc:"GeIntNorm is normalized GeInt value (divided by the layer maximum) -- this is used for learning in layers that require learning on subthreshold activity."`,
	"GiInt":     `range:"2" desc:"integrated running-average activation value computed from GiSyn with time constant Act.Dt.IntTau, to produce a longer-term integrated value reflecting the overall synaptic Gi level across the ThetaCycle time scale (Gi itself fluctuates considerably) -- useful for stats to set strength of connections etc to get neurons into right range of overall inhibitory drive"`,
	"GModRaw":   `desc:"raw modulatory conductance, received from GType = ModulatoryG projections"`,
	"GModSyn":   `desc:"syn integrated modulatory conductance, received from GType = ModulatoryG projections"`,
	"GMaintRaw": `desc:"raw maintenance conductance, received from GType = MaintG projections"`,
	"GMaintSyn": `desc:"syn integrated maintenance conductance, integrated using MaintNMDA params."`,

	"SSGi":     `auto-scale:"+" desc:"SST+ somatostatin positive slow spiking inhibition"`,
	"SSGiDend": `auto-scale:"+" desc:"amount of SST+ somatostatin positive slow spiking inhibition applied to dendritic Vm (VmDend)"`,
	"Gak":      `auto-scale:"+" desc:"conductance of A-type K potassium channels"`,

	"MahpN":    `auto-scale:"+" desc:"accumulating voltage-gated gating value for the medium time scale AHP"`,
	"SahpCa":   `desc:"slowly accumulating calcium value that drives the slow AHP"`,
	"SahpN":    `desc:"sAHP gating value"`,
	"GknaMed":  `auto-scale:"+" desc:"conductance of sodium-gated potassium channel (KNa) medium dynamics (Slick) -- produces accommodation / adaptation of firing"`,
	"GknaSlow": `auto-scale:"+" desc:"conductance of sodium-gated potassium channel (KNa) slow dynamics (Slack) -- produces accommodation / adaptation of firing"`,

	"GnmdaSyn":   `auto-scale:"+" desc:"integrated NMDA recv synaptic current -- adds GeRaw and decays with time constant"`,
	"Gnmda":      `auto-scale:"+" desc:"net postsynaptic (recv) NMDA conductance, after Mg V-gating and Gbar -- added directly to Ge as it has the same reversal potential"`,
	"GnmdaMaint": `auto-scale:"+" desc:"net postsynaptic maintenance NMDA conductance, computed from GMaintSyn and GMaintRaw, after Mg V-gating and Gbar -- added directly to Ge as it has the same reversal potential"`,
	"GnmdaLrn":   `auto-scale:"+" desc:"learning version of integrated NMDA recv synaptic current -- adds GeRaw and decays with time constant -- drives NmdaCa that then drives CaM for learning"`,
	"NmdaCa":     `auto-scale:"+" desc:"NMDA calcium computed from GnmdaLrn, drives learning via CaM"`,

	"GgabaB": `auto-scale:"+" desc:"net GABA-B conductance, after Vm gating and Gbar + Gbase -- applies to Gk, not Gi, for GIRK, with .1 reversal potential."`,
	"GABAB":  `auto-scale:"+" desc:"GABA-B / GIRK activation -- time-integrated value with rise and decay time constants"`,
	"GABABx": `auto-scale:"+" desc:"GABA-B / GIRK internal drive variable -- gets the raw activation and decays"`,

	"Gvgcc":     `auto-scale:"+" desc:"conductance (via Ca) for VGCC voltage gated calcium channels"`,
	"VgccM":     `desc:"activation gate of VGCC channels"`,
	"VgccH":     `desc:"inactivation gate of VGCC channels"`,
	"VgccCa":    `auto-scale:"+" desc:"instantaneous VGCC calcium flux -- can be driven by spiking or directly from Gvgcc"`,
	"VgccCaInt": `auto-scale:"+" desc:"time-integrated VGCC calcium flux -- this is actually what drives learning"`,

	"SKCaIn": `desc:"intracellular calcium store level, available to be released with spiking as SKCaR, which can bind to SKCa receptors and drive K current. replenishment is a function of spiking activity being below a threshold"`,
	"SKCaR":  `desc:"released amount of intracellular calcium, from SKCaIn, as a function of spiking events.  this can bind to SKCa channels and drive K currents."`,
	"SKCaM":  `desc:"Calcium-gated potassium channel gating factor, driven by SKCaR via a Hill equation as in chans.SKPCaParams."`,
	"Gsk":    `desc:"Calcium-gated potassium channel conductance as a function of Gbar * SKCaM."`,

	"Burst":      `desc:"5IB bursting activation value, computed by thresholding regular CaSpkP value in Super superficial layers"`,
	"BurstPrv":   `desc:"previous Burst bursting activation from prior time step -- used for context-based learning"`,
	"CtxtGe":     `desc:"context (temporally delayed) excitatory conductance, driven by deep bursting at end of the plus phase, for CT layers."`,
	"CtxtGeRawa": `desc:"raw update of context (temporally delayed) excitatory conductance, driven by deep bursting at end of the plus phase, for CT layers."`,
	"CtxtGeOrig": `desc:"original CtxtGe value prior to any decay factor -- updates at end of plus phase."`,

	"NrnFlags": `view:"-" desc:"bit flags for external input and other neuron status state"`,

	"ActAvg":  `desc:"average activation (of minus phase activation state) over long time intervals (time constant = Dt.LongAvgTau) -- useful for finding hog units and seeing overall distribution of activation"`,
	"AvgPct":  `range:"2" desc:"ActAvg as a proportion of overall layer activation -- this is used for synaptic scaling to match TrgAvg activation -- updated at SlowInterval intervals"`,
	"TrgAvg":  `range:"2" desc:"neuron's target average activation as a proportion of overall layer activation, assigned during weight initialization, driving synaptic scaling relative to AvgPct"`,
	"DTrgAvg": `auto-scale:"+" desc:"change in neuron's target average activation as a result of unit-wise error gradient -- acts like a bias weight.  MPI needs to share these across processors."`,
	"AvgDif":  `desc:"AvgPct - TrgAvg -- i.e., the error in overall activity level relative to set point for this neuron, which drives synaptic scaling -- updated at SlowInterval intervals"`,
	"GeBase":  `desc:"baseline level of Ge, added to GeRaw, for intrinsic excitability"`,
	"GiBase":  `desc:"baseline level of Gi, added to GiRaw, for intrinsic excitability"`,
}

NeuronVarProps has all of the display properties for neuron variables, including desc tooltips

View Source 
var PrjnProps = ki.Props{
	"EnumType:Typ": KiT_PrjnTypes,
}
View Source 
var SynapseVarProps = map[string]string{
	"Wt ":   `desc:"effective synaptic weight value, determining how much conductance one spike drives on the receiving neuron, representing the actual number of effective AMPA receptors in the synapse.  Wt = SWt * WtSig(LWt), where WtSig produces values between 0-2 based on LWt, centered on 1."`,
	"LWt":   `desc:"rapidly learning, linear weight value -- learns according to the lrate specified in the connection spec.  Biologically, this represents the internal biochemical processes that drive the trafficking of AMPA receptors in the synaptic density.  Initially all LWt are .5, which gives 1 from WtSig function."`,
	"SWt":   `desc:"slowly adapting structural weight value, which acts as a multiplicative scaling factor on synaptic efficacy: biologically represents the physical size and efficacy of the dendritic spine.  SWt values adapt in an outer loop along with synaptic scaling, with constraints to prevent runaway positive feedback loops and maintain variance and further capacity to learn.  Initial variance is all in SWt, with LWt set to .5, and scaling absorbs some of LWt into SWt."`,
	"DWt":   `auto-scale:"+" desc:"delta (change in) synaptic weight, from learning -- updates LWt which then updates Wt."`,
	"DSWt":  `auto-scale:"+" desc:"change in SWt slow synaptic weight -- accumulates DWt"`,
	"CaM":   `auto-scale:"+" desc:"first stage running average (mean) Ca calcium level (like CaM = calmodulin), feeds into CaP"`,
	"CaP":   `auto-scale:"+"desc:"shorter timescale integrated CaM value, representing the plus, LTP direction of weight change and capturing the function of CaMKII in the Kinase learning rule"`,
	"CaD":   `auto-scale:"+" desc:"longer timescale integrated CaP value, representing the minus, LTD direction of weight change and capturing the function of DAPK1 in the Kinase learning rule"`,
	"Tr":    `auto-scale:"+" desc:"trace of synaptic activity over time -- used for credit assignment in learning.  In MatrixPrjn this is a tag that is then updated later when US occurs."`,
	"DTr":   `auto-scale:"+" desc:"delta (change in) Tr trace of synaptic activity over time"`,
	"DiDWt": `auto-scale:"+" desc:"delta weight for each data parallel index (Di) -- this is directly computed from the Ca values (in cortical version) and then aggregated into the overall DWt (which may be further integrated across MPI nodes), which then drives changes in Wt values"`,
}

SynapseVarProps has all of the display properties for synapse variables, including desc tooltips

View Source 
var TheGPU *vgpu.GPU

TheGPU is the gpu device, shared across all networks

Functions

func AddGlbUSneg added in v1.8.16 

func AddGlbUSneg(ctx *Context, di uint32, gvar GlobalVars, negIdx uint32, val float32)

AddGlbUSneg is the CPU version of the global USneg variable addor

func AddGlbUSposV added in v1.8.16 

func AddGlbUSposV(ctx *Context, di uint32, gvar GlobalVars, posIdx uint32, val float32)

AddGlbUSposV is the CPU version of the global Drive, USpos variable adder

func AddGlbV added in v1.8.0 

func AddGlbV(ctx *Context, di uint32, gvar GlobalVars, val float32)

AddGlbV is the CPU version of the global variable addor

func AddNrnAvgV added in v1.8.0 

func AddNrnAvgV(ctx *Context, ni uint32, nvar NeuronAvgVars, val float32)

AddNrnAvgV is the CPU version of the neuron variable addor

func AddNrnV added in v1.8.0 

func AddNrnV(ctx *Context, ni, di uint32, nvar NeuronVars, val float32)

AddNrnV is the CPU version of the neuron variable addor

func AddSynCaV added in v1.8.0 

func AddSynCaV(ctx *Context, syni, di uint32, svar SynapseCaVars, val float32)

AddSynCaV is the CPU version of the synapse variable addor

func AddSynV added in v1.8.0 

func AddSynV(ctx *Context, syni uint32, svar SynapseVars, val float32)

AddSynV is the CPU version of the synapse variable addor

func DecaySynCa added in v1.3.21 

func DecaySynCa(ctx *Context, syni, di uint32, decay float32)

DecaySynCa decays synaptic calcium by given factor (between trials) Not used by default.

func GetRandomNumber added in v1.7.7 

func GetRandomNumber(index uint32, counter slrand.Counter, funIdx RandFunIdx) float32

GetRandomNumber returns a random number that depends on the index, counter and function index. We increment the counter after each cycle, so that we get new random numbers. This whole scheme exists to ensure equal results under different multithreading settings.

func GlbUSneg added in v1.8.0 

func GlbUSneg(ctx *Context, di uint32, gvar GlobalVars, negIdx uint32) float32

GlbUSneg is the CPU version of the global USneg variable accessor

func GlbUSposV added in v1.8.16 

func GlbUSposV(ctx *Context, di uint32, gvar GlobalVars, posIdx uint32) float32

GlbUSposV is the CPU version of the global Drive, USpos variable accessor

func GlbV added in v1.8.0 

func GlbV(ctx *Context, di uint32, gvar GlobalVars) float32

GlbV is the CPU version of the global variable accessor

func GlobalSetRew added in v1.8.16 

func GlobalSetRew(ctx *Context, di uint32, rew float32, hasRew bool)

GlobalSetRew is a convenience function for setting the external reward state in Globals variables

func GlobalsReset added in v1.8.16 

func GlobalsReset(ctx *Context)

GlobalsReset resets all global values to 0, for all NData

func HashEncodeSlice added in v1.7.14 

func HashEncodeSlice(slice []float32) string

func InitSynCa added in v1.3.21 

func InitSynCa(ctx *Context, syni, di uint32)

InitSynCa initializes synaptic calcium state, including CaUpT

func IsExtLayerType added in v1.7.9 

func IsExtLayerType(lt LayerTypes) bool

IsExtLayerType returns true if the layer type deals with external input: Input, Target, Compare

func JsonToParams 

func JsonToParams(b []byte) string

JsonToParams reformates json output to suitable params display output

func LayerActsLog added in v1.7.11 

func LayerActsLog(net *Network, lg *elog.Logs, di int, gui *egui.GUI)

LayerActsLog records layer activity for tuning the network inhibition, nominal activity, relative scaling, etc. if gui is non-nil, plot is updated.

func LayerActsLogAvg added in v1.7.11 

func LayerActsLogAvg(net *Network, lg *elog.Logs, gui *egui.GUI, recReset bool)

LayerActsLogAvg computes average of LayerActsRec record of layer activity for tuning the network inhibition, nominal activity, relative scaling, etc. if gui is non-nil, plot is updated. if recReset is true, reset the recorded data after computing average.

func LayerActsLogConfig added in v1.7.11 

func LayerActsLogConfig(net *Network, lg *elog.Logs)

LayerActsLogConfig configures Tables to record layer activity for tuning the network inhibition, nominal activity, relative scaling, etc. in elog.MiscTables: LayerActs is current, LayerActsRec is record over trials, LayerActsAvg is average of recorded trials.

func LayerActsLogConfigGUI added in v1.7.11 

func LayerActsLogConfigGUI(lg *elog.Logs, gui *egui.GUI)

LayerActsLogConfigGUI configures GUI for LayerActsLog Plot and LayerActs Avg Plot

func LayerActsLogConfigMetaData added in v1.7.11 

func LayerActsLogConfigMetaData(dt *etable.Table)

LayerActsLogConfigMetaData configures meta data for LayerActs table

func LayerActsLogRecReset added in v1.7.11 

func LayerActsLogRecReset(lg *elog.Logs)

LayerActsLogRecReset resets the recorded LayerActsRec data used for computing averages

func LogAddCaLrnDiagnosticItems added in v1.5.3 

func LogAddCaLrnDiagnosticItems(lg *elog.Logs, mode etime.Modes, net *Network, times ...etime.Times)

LogAddCaLrnDiagnosticItems adds standard Axon diagnostic statistics to given logs, across two given time levels, in higher to lower order, e.g., Epoch, Trial These were useful for the development of the Ca-based "trace" learning rule that directly uses NMDA and VGCC-like spiking Ca

func LogAddDiagnosticItems added in v1.3.35 

func LogAddDiagnosticItems(lg *elog.Logs, layerNames []string, mode etime.Modes, times ...etime.Times)

LogAddDiagnosticItems adds standard Axon diagnostic statistics to given logs, across two given time levels, in higher to lower order, e.g., Epoch, Trial These are useful for tuning and diagnosing the behavior of the network.

func LogAddExtraDiagnosticItems added in v1.5.8 

func LogAddExtraDiagnosticItems(lg *elog.Logs, mode etime.Modes, net *Network, times ...etime.Times)

LogAddExtraDiagnosticItems adds extra Axon diagnostic statistics to given logs, across two given time levels, in higher to lower order, e.g., Epoch, Trial These are useful for tuning and diagnosing the behavior of the network.

func LogAddLayerGeActAvgItems added in v1.3.35 

func LogAddLayerGeActAvgItems(lg *elog.Logs, net *Network, mode etime.Modes, etm etime.Times)

LogAddLayerGeActAvgItems adds Ge and Act average items for Hidden and Target layers for given mode and time (e.g., Test, Cycle) These are useful for monitoring layer activity during testing.

func LogAddPCAItems added in v1.3.35 

func LogAddPCAItems(lg *elog.Logs, net *Network, mode etime.Modes, times ...etime.Times)

LogAddPCAItems adds PCA statistics to log for Hidden and Target layers across 3 given time levels, in higher to lower order, e.g., Run, Epoch, Trial These are useful for diagnosing the behavior of the network.

func LogAddPulvCorSimItems added in v1.7.0 

func LogAddPulvCorSimItems(lg *elog.Logs, net *Network, mode etime.Modes, times ...etime.Times)

LogAddPulvCorSimItems adds CorSim stats for Pulv / Pulvinar layers aggregated across three time scales, ordered from higher to lower, e.g., Run, Epoch, Trial.

func LogInputLayer added in v1.7.7 

func LogInputLayer(lg *elog.Logs, net *Network, mode etime.Modes)

func LogTestErrors added in v1.3.35 

func LogTestErrors(lg *elog.Logs)

LogTestErrors records all errors made across TestTrials, at Test Epoch scope

func LogisticFun added in v1.8.16 

func LogisticFun(v, gain float32) float32

LogisticFun is the sigmoid logistic function

func LooperResetLogBelow added in v1.3.35 

func LooperResetLogBelow(man *looper.Manager, logs *elog.Logs, except ...etime.Times)

LooperResetLogBelow adds a function in OnStart to all stacks and loops to reset the log at the level below each loop -- this is good default behavior. Exceptions can be passed to exclude specific levels -- e.g., if except is Epoch then Epoch does not reset the log below it

func LooperSimCycleAndLearn added in v1.3.35 

func LooperSimCycleAndLearn(man *looper.Manager, net *Network, ctx *Context, viewupdt *netview.ViewUpdt, trial ...etime.Times)

LooperSimCycleAndLearn adds Cycle and DWt, WtFmDWt functions to looper for given network, ctx, and netview update manager Can pass a trial-level time scale to use instead of the default etime.Trial

func LooperStdPhases added in v1.3.35 

func LooperStdPhases(man *looper.Manager, ctx *Context, net *Network, plusStart, plusEnd int, trial ...etime.Times)

LooperStdPhases adds the minus and plus phases of the theta cycle, along with embedded beta phases which just record St1 and St2 activity in this case. plusStart is start of plus phase, typically 150, and plusEnd is end of plus phase, typically 199 resets the state at start of trial. Can pass a trial-level time scale to use instead of the default etime.Trial

func LooperUpdtNetView added in v1.3.35 

func LooperUpdtNetView(man *looper.Manager, viewupdt *netview.ViewUpdt, net *Network, ctrUpdtFunc func(tm etime.Times))

LooperUpdtNetView adds netview update calls at each time level

func LooperUpdtPlots added in v1.3.35 

func LooperUpdtPlots(man *looper.Manager, gui *egui.GUI)

LooperUpdtPlots adds plot update calls at each time level

func MulNrnAvgV added in v1.8.0 

func MulNrnAvgV(ctx *Context, ni uint32, nvar NeuronAvgVars, val float32)

MulNrnAvgV is the CPU version of the neuron variable multiplier

func MulNrnV added in v1.8.0 

func MulNrnV(ctx *Context, ni, di uint32, nvar NeuronVars, val float32)

MulNrnV is the CPU version of the neuron variable multiplier

func MulSynCaV added in v1.8.0 

func MulSynCaV(ctx *Context, syni, di uint32, svar SynapseCaVars, val float32)

MulSynCaV is the CPU version of the synapse variable multiplier

func MulSynV added in v1.8.0 

func MulSynV(ctx *Context, syni uint32, svar SynapseVars, val float32)

MulSynV is the CPU version of the synapse variable multiplier

func NeuronVarIdxByName 

func NeuronVarIdxByName(varNm string) (int, error)

NeuronVarIdxByName returns the index of the variable in the Neuron, or error

func NrnAvgV added in v1.8.0 

func NrnAvgV(ctx *Context, ni uint32, nvar NeuronAvgVars) float32

NrnAvgV is the CPU version of the neuron variable accessor

func NrnClearFlag added in v1.8.0 

func NrnClearFlag(ctx *Context, ni, di uint32, flag NeuronFlags)

func NrnHasFlag added in v1.8.0 

func NrnHasFlag(ctx *Context, ni, di uint32, flag NeuronFlags) bool

func NrnI added in v1.8.0 

func NrnI(ctx *Context, ni uint32, idx NeuronIdxs) uint32

NrnI is the CPU version of the neuron idx accessor

func NrnIsOff added in v1.8.0 

func NrnIsOff(ctx *Context, ni uint32) bool

NrnIsOff returns true if the neuron has been turned off (lesioned) Only checks the first data item -- all should be consistent.

func NrnSetFlag added in v1.8.0 

func NrnSetFlag(ctx *Context, ni, di uint32, flag NeuronFlags)

func NrnV added in v1.8.0 

func NrnV(ctx *Context, ni, di uint32, nvar NeuronVars) float32

NrnV is the CPU version of the neuron variable accessor

func PCAStats added in v1.3.35 

func PCAStats(net *Network, lg *elog.Logs, stats *estats.Stats)

PCAStats computes PCA statistics on recorded hidden activation patterns from Analyze, Trial log data

func PVLVNormFun added in v1.8.16 

func PVLVNormFun(raw float32) float32

PVLVLNormFun is the normalizing function applied to the sum of all weighted raw values: 1 - (1 / (1 + usRaw.Sum()))

func PVLVUSStimVal added in v1.8.0 

func PVLVUSStimVal(ctx *Context, di uint32, usIdx uint32, valence ValenceTypes) float32

PVLVUSStimVal returns stimulus value for US at given index and valence. If US > 0.01, a full 1 US activation is returned.

func ParallelChunkRun added in v1.7.24 

func ParallelChunkRun(fun func(st, ed int), total int, nThreads int)

Maps the given function across the [0, total) range of items, using nThreads goroutines, in smaller-sized chunks for better load balancing. this may be better for larger number of threads, but is not better for small N

func ParallelRun added in v1.7.24 

func ParallelRun(fun func(st, ed uint32), total uint32, nThreads int)

Maps the given function across the [0, total) range of items, using nThreads goroutines.

func SaveWeights added in v1.3.29 

func SaveWeights(net *Network, ctrString, runName string) string

SaveWeights saves network weights to filename with WeightsFileName information to identify the weights. only for 0 rank MPI if running mpi Returns the name of the file saved to, or empty if not saved.

func SaveWeightsIfArgSet added in v1.3.35 

func SaveWeightsIfArgSet(net *Network, args *ecmd.Args, ctrString, runName string) string

SaveWeightsIfArgSet saves network weights if the "wts" arg has been set to true. uses WeightsFileName information to identify the weights. only for 0 rank MPI if running mpi Returns the name of the file saved to, or empty if not saved.

func SaveWeightsIfConfigSet added in v1.8.4 

func SaveWeightsIfConfigSet(net *Network, cfgWts bool, ctrString, runName string) string

SaveWeightsIfConfigSet saves network weights if the given config bool value has been set to true. uses WeightsFileName information to identify the weights. only for 0 rank MPI if running mpi Returns the name of the file saved to, or empty if not saved.

func SetAvgMaxFloatFromIntErr added in v1.8.0 

func SetAvgMaxFloatFromIntErr(fun func())

func SetGlbUSneg added in v1.8.0 

func SetGlbUSneg(ctx *Context, di uint32, gvar GlobalVars, negIdx uint32, val float32)

SetGlbUSneg is the CPU version of the global USneg variable settor

func SetGlbUSposV added in v1.8.16 

func SetGlbUSposV(ctx *Context, di uint32, gvar GlobalVars, posIdx uint32, val float32)

SetGlbUSposV is the CPU version of the global Drive, USpos variable settor

func SetGlbV added in v1.8.0 

func SetGlbV(ctx *Context, di uint32, gvar GlobalVars, val float32)

SetGlbV is the CPU version of the global variable settor

func SetNeuronExtPosNeg added in v1.7.0 

func SetNeuronExtPosNeg(ctx *Context, ni, di uint32, val float32)

SetNeuronExtPosNeg sets neuron Ext value based on neuron index with positive values going in first unit, negative values rectified to positive in 2nd unit

func SetNrnAvgV added in v1.8.0 

func SetNrnAvgV(ctx *Context, ni uint32, nvar NeuronAvgVars, val float32)

SetNrnAvgV is the CPU version of the neuron variable settor

func SetNrnI added in v1.8.0 

func SetNrnI(ctx *Context, ni uint32, idx NeuronIdxs, val uint32)

SetNrnI is the CPU version of the neuron idx settor

func SetNrnV added in v1.8.0 

func SetNrnV(ctx *Context, ni, di uint32, nvar NeuronVars, val float32)

SetNrnV is the CPU version of the neuron variable settor

func SetSynCaV added in v1.8.0 

func SetSynCaV(ctx *Context, syni, di uint32, svar SynapseCaVars, val float32)

SetSynCaV is the CPU version of the synapse variable settor

func SetSynI added in v1.8.0 

func SetSynI(ctx *Context, syni uint32, idx SynapseIdxs, val uint32)

SetSynI is the CPU version of the synapse idx settor

func SetSynV added in v1.8.0 

func SetSynV(ctx *Context, syni uint32, svar SynapseVars, val float32)

SetSynV is the CPU version of the synapse variable settor

func SigFun 

func SigFun(w, gain, off float32) float32

SigFun is the sigmoid function for value w in 0-1 range, with gain and offset params

func SigFun61 

func SigFun61(w float32) float32

SigFun61 is the sigmoid function for value w in 0-1 range, with default gain = 6, offset = 1 params

func SigInvFun 

func SigInvFun(w, gain, off float32) float32

SigInvFun is the inverse of the sigmoid function

func SigInvFun61 

func SigInvFun61(w float32) float32

SigInvFun61 is the inverse of the sigmoid function, with default gain = 6, offset = 1 params

func SynCaV added in v1.8.0 

func SynCaV(ctx *Context, syni, di uint32, svar SynapseCaVars) float32

SynCaV is the CPU version of the synapse variable accessor

func SynI added in v1.8.0 

func SynI(ctx *Context, syni uint32, idx SynapseIdxs) uint32

SynI is the CPU version of the synapse idx accessor

func SynV added in v1.8.0 

func SynV(ctx *Context, syni uint32, svar SynapseVars) float32

SynV is the CPU version of the synapse variable accessor

func SynapseVarByName 

func SynapseVarByName(varNm string) (int, error)

SynapseVarByName returns the index of the variable in the Synapse, or error

func ToggleLayersOff added in v1.3.29 

func ToggleLayersOff(net *Network, layerNames []string, off bool)

ToggleLayersOff can be used to disable layers in a Network, for example if you are doing an ablation study.

func WeightsFileName added in v1.3.35 

func WeightsFileName(net *Network, ctrString, runName string) string

WeightsFileName returns default current weights file name, using train run and epoch counters from looper and the RunName string identifying tag, parameters and starting run,

Types

type ActAvgParams 

type ActAvgParams struct {

	// [min: 0] [step: 0.01] [typically 0.01 - 0.2] nominal estimated average activity level in the layer, which is used in computing the scaling factor on sending projections from this layer.  In general it should roughly match the layer ActAvg.ActMAvg value, which can be logged using the axon.LogAddDiagnosticItems function.  If layers receiving from this layer are not getting enough Ge excitation, then this Nominal level can be lowered to increase projection strength (fewer active neurons means each one contributes more, so scaling factor goes as the inverse of activity level), or vice-versa if Ge is too high.  It is also the basis for the target activity level used for the AdaptGi option -- see the Offset which is added to this value.
	Nominal float32 `` /* 745-byte string literal not displayed */

	// enable adapting of layer inhibition Gi multiplier factor (stored in layer GiMult value) to maintain a Target layer level of ActAvg.ActMAvg.  This generally works well and improves the long-term stability of the models.  It is not enabled by default because it depends on having established a reasonable Nominal + Offset target activity level.
	AdaptGi slbool.Bool `` /* 349-byte string literal not displayed */

	// [def: 0] [viewif: AdaptGi] [min: 0] [step: 0.01] offset to add to Nominal for the target average activity that drives adaptation of Gi for this layer.  Typically the Nominal level is good, but sometimes Nominal must be adjusted up or down to achieve desired Ge scaling, so this Offset can compensate accordingly.
	Offset float32 `` /* 315-byte string literal not displayed */

	// [def: 0] [viewif: AdaptGi] tolerance for higher than Target target average activation as a proportion of that target value (0 = exactly the target, 0.2 = 20% higher than target) -- only once activations move outside this tolerance are inhibitory values adapted.
	HiTol float32 `` /* 266-byte string literal not displayed */

	// [def: 0.8] [viewif: AdaptGi] tolerance for lower than Target target average activation as a proportion of that target value (0 = exactly the target, 0.5 = 50% lower than target) -- only once activations move outside this tolerance are inhibitory values adapted.
	LoTol float32 `` /* 266-byte string literal not displayed */

	// [def: 0.1] [viewif: AdaptGi] rate of Gi adaptation as function of AdaptRate * (Target - ActMAvg) / Target -- occurs at spaced intervals determined by Network.SlowInterval value -- slower values such as 0.01 may be needed for large networks and sparse layers.
	AdaptRate float32 `` /* 263-byte string literal not displayed */
	// contains filtered or unexported fields
}

ActAvgParams represents the nominal average activity levels in the layer and parameters for adapting the computed Gi inhibition levels to maintain average activity within a target range.

func (*ActAvgParams) Adapt added in v1.2.37 

func (aa *ActAvgParams) Adapt(gimult *float32, act float32) bool

Adapt adapts the given gi multiplier factor as function of target and actual average activation, given current params.

func (*ActAvgParams) AvgFmAct 

func (aa *ActAvgParams) AvgFmAct(avg *float32, act float32, dt float32)

AvgFmAct updates the running-average activation given average activity level in layer

func (*ActAvgParams) Defaults 

func (aa *ActAvgParams) Defaults()

func (*ActAvgParams) Update 

func (aa *ActAvgParams) Update()

type ActAvgVals added in v1.2.32 

type ActAvgVals struct {

	// running-average minus-phase activity integrated at Dt.LongAvgTau -- used for adapting inhibition relative to target level
	ActMAvg float32 `` /* 141-byte string literal not displayed */

	// running-average plus-phase activity integrated at Dt.LongAvgTau
	ActPAvg float32 `inactive:"+" desc:"running-average plus-phase activity integrated at Dt.LongAvgTau"`

	// running-average max of minus-phase Ge value across the layer integrated at Dt.LongAvgTau
	AvgMaxGeM float32 `inactive:"+" desc:"running-average max of minus-phase Ge value across the layer integrated at Dt.LongAvgTau"`

	// running-average max of minus-phase Gi value across the layer integrated at Dt.LongAvgTau
	AvgMaxGiM float32 `inactive:"+" desc:"running-average max of minus-phase Gi value across the layer integrated at Dt.LongAvgTau"`

	// multiplier on inhibition -- adapted to maintain target activity level
	GiMult float32 `inactive:"+" desc:"multiplier on inhibition -- adapted to maintain target activity level"`

	// adaptive threshold -- only used for specialized layers, e.g., VSPatch
	AdaptThr float32 `inactive:"+" desc:"adaptive threshold -- only used for specialized layers, e.g., VSPatch"`
	// contains filtered or unexported fields
}

ActAvgVals are long-running-average activation levels stored in the LayerVals, for monitoring and adapting inhibition and possibly scaling parameters. All of these integrate over NData within a network, so are the same across them.

func (*ActAvgVals) Init added in v1.7.9 

func (lv *ActAvgVals) Init()

type ActInitParams 

type ActInitParams struct {

	// [def: 0.3] initial membrane potential -- see Erev.L for the resting potential (typically .3)
	Vm float32 `def:"0.3" desc:"initial membrane potential -- see Erev.L for the resting potential (typically .3)"`

	// [def: 0] initial activation value -- typically 0
	Act float32 `def:"0" desc:"initial activation value -- typically 0"`

	// [def: 0] baseline level of excitatory conductance (net input) -- Ge is initialized to this value, and it is added in as a constant background level of excitatory input -- captures all the other inputs not represented in the model, and intrinsic excitability, etc
	GeBase float32 `` /* 268-byte string literal not displayed */

	// [def: 0] baseline level of inhibitory conductance (net input) -- Gi is initialized to this value, and it is added in as a constant background level of inhibitory input -- captures all the other inputs not represented in the model
	GiBase float32 `` /* 235-byte string literal not displayed */

	// [def: 0] variance (sigma) of gaussian distribution around baseline Ge values, per unit, to establish variability in intrinsic excitability.  value never goes < 0
	GeVar float32 `` /* 167-byte string literal not displayed */

	// [def: 0] variance (sigma) of gaussian distribution around baseline Gi values, per unit, to establish variability in intrinsic excitability.  value never goes < 0
	GiVar float32 `` /* 167-byte string literal not displayed */
	// contains filtered or unexported fields
}

ActInitParams are initial values for key network state variables. Initialized in InitActs called by InitWts, and provides target values for DecayState.

func (*ActInitParams) Defaults 

func (ai *ActInitParams) Defaults()

func (*ActInitParams) GetGeBase added in v1.7.7 

func (ai *ActInitParams) GetGeBase(rnd erand.Rand) float32

GeBase returns the baseline Ge value: Ge + rand(GeVar) > 0

func (*ActInitParams) GetGiBase added in v1.7.7 

func (ai *ActInitParams) GetGiBase(rnd erand.Rand) float32

GiBase returns the baseline Gi value: Gi + rand(GiVar) > 0

func (*ActInitParams) Update 

func (ai *ActInitParams) Update()

type ActParams 

type ActParams struct {

	// [view: inline] Spiking function parameters
	Spikes SpikeParams `view:"inline" desc:"Spiking function parameters"`

	// [view: inline] dendrite-specific parameters
	Dend DendParams `view:"inline" desc:"dendrite-specific parameters"`

	// [view: inline] initial values for key network state variables -- initialized in InitActs called by InitWts, and provides target values for DecayState
	Init ActInitParams `` /* 155-byte string literal not displayed */

	// [view: inline] amount to decay between AlphaCycles, simulating passage of time and effects of saccades etc, especially important for environments with random temporal structure (e.g., most standard neural net training corpora)
	Decay DecayParams `` /* 233-byte string literal not displayed */

	// [view: inline] time and rate constants for temporal derivatives / updating of activation state
	Dt DtParams `view:"inline" desc:"time and rate constants for temporal derivatives / updating of activation state"`

	// [view: inline] [Defaults: 1, .2, 1, 1] maximal conductances levels for channels
	Gbar chans.Chans `view:"inline" desc:"[Defaults: 1, .2, 1, 1] maximal conductances levels for channels"`

	// [view: inline] [Defaults: 1, .3, .25, .1] reversal potentials for each channel
	Erev chans.Chans `view:"inline" desc:"[Defaults: 1, .3, .25, .1] reversal potentials for each channel"`

	// [view: inline] how external inputs drive neural activations
	Clamp ClampParams `view:"inline" desc:"how external inputs drive neural activations"`

	// [view: inline] how, where, when, and how much noise to add
	Noise SpikeNoiseParams `view:"inline" desc:"how, where, when, and how much noise to add"`

	// [view: inline] range for Vm membrane potential -- [0.1, 1.0] -- important to keep just at extreme range of reversal potentials to prevent numerical instability
	VmRange minmax.F32 `` /* 165-byte string literal not displayed */

	// [view: inline] M-type medium time-scale afterhyperpolarization mAHP current -- this is the primary form of adaptation on the time scale of multiple sequences of spikes
	Mahp chans.MahpParams `` /* 173-byte string literal not displayed */

	// [view: inline] slow time-scale afterhyperpolarization sAHP current -- integrates CaSpkD at theta cycle intervals and produces a hard cutoff on sustained activity for any neuron
	Sahp chans.SahpParams `` /* 182-byte string literal not displayed */

	// [view: inline] sodium-gated potassium channel adaptation parameters -- activates a leak-like current as a function of neural activity (firing = Na influx) at two different time-scales (Slick = medium, Slack = slow)
	KNa chans.KNaMedSlow `` /* 220-byte string literal not displayed */

	// [view: inline] NMDA channel parameters used in computing Gnmda conductance for bistability, and postsynaptic calcium flux used in learning.  Note that Learn.Snmda has distinct parameters used in computing sending NMDA parameters used in learning.
	NMDA chans.NMDAParams `` /* 252-byte string literal not displayed */

	// [view: inline] NMDA channel parameters used in computing Gnmda conductance for bistability, and postsynaptic calcium flux used in learning.  Note that Learn.Snmda has distinct parameters used in computing sending NMDA parameters used in learning.
	MaintNMDA chans.NMDAParams `` /* 252-byte string literal not displayed */

	// [view: inline] GABA-B / GIRK channel parameters
	GabaB chans.GABABParams `view:"inline" desc:"GABA-B / GIRK channel parameters"`

	// [view: inline] voltage gated calcium channels -- provide a key additional source of Ca for learning and positive-feedback loop upstate for active neurons
	VGCC chans.VGCCParams `` /* 159-byte string literal not displayed */

	// [view: inline] A-type potassium (K) channel that is particularly important for limiting the runaway excitation from VGCC channels
	AK chans.AKsParams `` /* 135-byte string literal not displayed */

	// [view: inline] small-conductance calcium-activated potassium channel produces the pausing function as a consequence of rapid bursting.
	SKCa chans.SKCaParams `` /* 140-byte string literal not displayed */

	// [view: inline] Attentional modulation parameters: how Attn modulates Ge
	AttnMod AttnParams `view:"inline" desc:"Attentional modulation parameters: how Attn modulates Ge"`

	// [view: inline] provides encoding population codes, used to represent a single continuous (scalar) value, across a population of units / neurons (1 dimensional)
	PopCode PopCodeParams `` /* 165-byte string literal not displayed */
}

axon.ActParams contains all the activation computation params and functions for basic Axon, at the neuron level . This is included in axon.Layer to drive the computation.

func (*ActParams) AddGeNoise added in v1.8.0 

func (ac *ActParams) AddGeNoise(ctx *Context, ni, di uint32)

AddGeNoise updates nrn.GeNoise if active

func (*ActParams) AddGiNoise added in v1.8.0 

func (ac *ActParams) AddGiNoise(ctx *Context, ni, di uint32)

AddGiNoise updates nrn.GiNoise if active

func (*ActParams) DecayAHP added in v1.7.18 

func (ac *ActParams) DecayAHP(ctx *Context, ni, di uint32, decay float32)

DecayAHP decays after-hyperpolarization variables by given factor (typically Decay.AHP)

func (*ActParams) DecayLearnCa added in v1.7.16 

func (ac *ActParams) DecayLearnCa(ctx *Context, ni, di uint32, decay float32)

DecayLearnCa decays neuron-level calcium learning and spiking variables by given factor. Note: this is generally NOT useful, causing variability in these learning factors as a function of the decay parameter that then has impacts on learning rates etc. see Act.Decay.LearnCa param controlling this

func (*ActParams) DecayState 

func (ac *ActParams) DecayState(ctx *Context, ni, di uint32, decay, glong, ahp float32)

DecayState decays the activation state toward initial values in proportion to given decay parameter. Special case values such as Glong and KNa are also decayed with their separately parameterized values. Called with ac.Decay.Act by Layer during NewState

func (*ActParams) Defaults 

func (ac *ActParams) Defaults()

func (*ActParams) GSkCaFmCa added in v1.7.0 

func (ac *ActParams) GSkCaFmCa(ctx *Context, ni, di uint32)

GSkCaFmCa updates the SKCa channel if used

func (*ActParams) GeFmSyn added in v1.5.12 

func (ac *ActParams) GeFmSyn(ctx *Context, ni, di uint32, geSyn, geExt float32)

GeFmSyn integrates Ge excitatory conductance from GeSyn. geExt is extra conductance to add to the final Ge value

func (*ActParams) GiFmSyn added in v1.5.12 

func (ac *ActParams) GiFmSyn(ctx *Context, ni, di uint32, giSyn float32) float32

GiFmSyn integrates GiSyn inhibitory synaptic conductance from GiRaw value (can add other terms to geRaw prior to calling this)

func (*ActParams) GkFmVm added in v1.6.0 

func (ac *ActParams) GkFmVm(ctx *Context, ni, di uint32)

GkFmVm updates all the Gk-based conductances: Mahp, KNa, Gak

func (*ActParams) GvgccFmVm added in v1.3.24 

func (ac *ActParams) GvgccFmVm(ctx *Context, ni, di uint32)

GvgccFmVm updates all the VGCC voltage-gated calcium channel variables from VmDend

func (*ActParams) InetFmG 

func (ac *ActParams) InetFmG(vm, ge, gl, gi, gk float32) float32

InetFmG computes net current from conductances and Vm

func (*ActParams) InitActs 

func (ac *ActParams) InitActs(ctx *Context, ni, di uint32)

InitActs initializes activation state in neuron -- called during InitWts but otherwise not automatically called (DecayState is used instead)

func (*ActParams) InitLongActs added in v1.2.66 

func (ac *ActParams) InitLongActs(ctx *Context, ni, di uint32)

InitLongActs initializes longer time-scale activation states in neuron (SpkPrv, SpkSt*, ActM, ActP) Called from InitActs, which is called from InitWts, but otherwise not automatically called (DecayState is used instead)

func (*ActParams) KNaNewState added in v1.7.18 

func (ac *ActParams) KNaNewState(ctx *Context, ni, di uint32)

KNaNewState does TrialSlow version of KNa during NewState if option is set

func (*ActParams) MaintNMDAFmRaw added in v1.7.19 

func (ac *ActParams) MaintNMDAFmRaw(ctx *Context, ni, di uint32)

MaintNMDAFmRaw updates all the Maint NMDA variables from GModRaw and current Vm, Spiking

func (*ActParams) NMDAFmRaw added in v1.3.1 

func (ac *ActParams) NMDAFmRaw(ctx *Context, ni, di uint32, geTot float32)

NMDAFmRaw updates all the NMDA variables from total Ge (GeRaw + Ext) and current Vm, Spiking

func (*ActParams) SpikeFmVm added in v1.6.12 

func (ac *ActParams) SpikeFmVm(ctx *Context, ni, di uint32)

SpikeFmVm computes Spike from Vm and ISI-based activation

func (*ActParams) SpikeFmVmVars added in v1.8.0 

func (ac *ActParams) SpikeFmVmVars(nrnISI, nrnISIAvg, nrnSpike, nrnSpiked, nrnAct *float32, nrnVm float32)

SpikeFmVmVars computes Spike from Vm and ISI-based activation, using pointers to variables

func (*ActParams) Update 

func (ac *ActParams) Update()

Update must be called after any changes to parameters

func (*ActParams) VmFmG 

func (ac *ActParams) VmFmG(ctx *Context, ni, di uint32)

VmFmG computes membrane potential Vm from conductances Ge, Gi, and Gk.

func (*ActParams) VmFmInet added in v1.2.95 

func (ac *ActParams) VmFmInet(vm, dt, inet float32) float32

VmFmInet computes new Vm value from inet, clamping range

func (*ActParams) VmInteg added in v1.2.96 

func (ac *ActParams) VmInteg(vm, dt, ge, gl, gi, gk float32, nvm, inet *float32)

VmInteg integrates Vm over VmSteps to obtain a more stable value Returns the new Vm and inet values.

type AttnParams added in v1.2.85 

type AttnParams struct {

	// is attentional modulation active?
	On slbool.Bool `desc:"is attentional modulation active?"`

	// [viewif: On] minimum act multiplier if attention is 0
	Min float32 `viewif:"On" desc:"minimum act multiplier if attention is 0"`

	// threshold on CaSpkP for determining the reaction time for the Layer -- starts after MaxCycStart to ensure that prior trial activity has had a chance to dissipate.
	RTThr float32 `` /* 169-byte string literal not displayed */
	// contains filtered or unexported fields
}

AttnParams determine how the Attn modulates Ge

func (*AttnParams) Defaults added in v1.2.85 

func (at *AttnParams) Defaults()

func (*AttnParams) ModVal added in v1.2.85 

func (at *AttnParams) ModVal(val float32, attn float32) float32

ModVal returns the attn-modulated value -- attn must be between 1-0

func (*AttnParams) Update added in v1.2.85 

func (at *AttnParams) Update()

type AvgMaxI32 added in v1.7.9 

type AvgMaxI32 struct {

	// Average, from Calc when last computed as Sum / N
	Avg float32 `inactive:"+" desc:"Average, from Calc when last computed as Sum / N"`

	// Maximum value, copied from CurMax in Calc
	Max float32 `inactive:"+" desc:"Maximum value, copied from CurMax in Calc"`

	// sum for computing average -- incremented in UpdateVal, reset in Calc
	Sum int32 `inactive:"+" desc:"sum for computing average -- incremented in UpdateVal, reset in Calc"`

	// current maximum value, updated via UpdateVal, reset in Calc
	CurMax int32 `inactive:"+" desc:"current maximum value, updated via UpdateVal, reset in Calc"`

	// number of items in the sum -- this must be set in advance to a known value and it is used in computing the float <-> int conversion factor to maximize precision.
	N int32 `` /* 181-byte string literal not displayed */
	// contains filtered or unexported fields
}

AvgMaxI32 holds average and max statistics for float32, and values used for computing them incrementally, using a fixed precision int32 based float representation that can be used with GPU-based atomic add and max functions. This ONLY works for positive values with averages around 1, and the N must be set IN ADVANCE to the correct number of items. Once Calc() is called, the incremental values are reset via Init() so it is always ready for updating without a separate Init() pass.

func (*AvgMaxI32) Calc added in v1.7.9 

func (am *AvgMaxI32) Calc(refIdx int32)

Calc computes the average given the current Sum and copies over CurMax to Max refIdx is a reference index of thing being computed, which will be printed in case there is an overflow, for debugging (can't be a string because this code runs on GPU).

func (*AvgMaxI32) FloatFmIntFactor added in v1.7.9 

func (am *AvgMaxI32) FloatFmIntFactor() float32

FloatFmIntFactor returns the factor used for converting int32 back to float32 -- this is 1 / FloatToIntFactor for faster multiplication instead of dividing.

func (*AvgMaxI32) FloatFromInt added in v1.7.9 

func (am *AvgMaxI32) FloatFromInt(ival, refIdx int32) float32

FloatFromInt converts the given int32 value produced via FloatToInt back into a float32 (divides by factor)

func (*AvgMaxI32) FloatToInt added in v1.7.9 

func (am *AvgMaxI32) FloatToInt(val float32) int32

FloatToInt converts the given floating point value to a large int for max updating.

func (*AvgMaxI32) FloatToIntFactor added in v1.7.9 

func (am *AvgMaxI32) FloatToIntFactor() float32

FloatToIntFactor returns the factor used for converting float32 to int32 for Max updating, assuming that the overall value is in the general order of 0-1 (127 is the max).

func (*AvgMaxI32) FloatToIntSum added in v1.7.9 

func (am *AvgMaxI32) FloatToIntSum(val float32) int32

FloatToIntSum converts the given floating point value to a large int for sum accumulating -- divides by N.

func (*AvgMaxI32) Init added in v1.7.9 

func (am *AvgMaxI32) Init()

Init initializes incremental values used during updating.

func (*AvgMaxI32) String added in v1.7.9 

func (am *AvgMaxI32) String() string

func (*AvgMaxI32) UpdateVal added in v1.7.9 

func (am *AvgMaxI32) UpdateVal(val float32)

UpdateVal updates stats from given value

func (*AvgMaxI32) Zero added in v1.7.9 

func (am *AvgMaxI32) Zero()

Zero resets everything completely -- back to "as new" state.

type AvgMaxPhases added in v1.7.0 

type AvgMaxPhases struct {

	// [view: inline] updated every cycle -- this is the source of all subsequent time scales
	Cycle AvgMaxI32 `view:"inline" desc:"updated every cycle -- this is the source of all subsequent time scales"`

	// [view: inline] at the end of the minus phase
	Minus AvgMaxI32 `view:"inline" desc:"at the end of the minus phase"`

	// [view: inline] at the end of the plus phase
	Plus AvgMaxI32 `view:"inline" desc:"at the end of the plus phase"`

	// [view: inline] at the end of the previous plus phase
	Prev AvgMaxI32 `view:"inline" desc:"at the end of the previous plus phase"`
}

AvgMaxPhases contains the average and maximum values over a Pool of neurons, at different time scales within a standard ThetaCycle of updating. It is much more efficient on the GPU to just grab everything in one pass at the cycle level, and then take snapshots from there. All of the cycle level values are updated at the *start* of the cycle based on values from the prior cycle -- thus are 1 cycle behind in general.

func (*AvgMaxPhases) Calc added in v1.7.9 

func (am *AvgMaxPhases) Calc(refIdx int32)

Calc does Calc on Cycle, which is then ready for aggregation again

func (*AvgMaxPhases) CycleToMinus added in v1.7.0 

func (am *AvgMaxPhases) CycleToMinus()

CycleToMinus grabs current Cycle values into the Minus phase values

func (*AvgMaxPhases) CycleToPlus added in v1.7.0 

func (am *AvgMaxPhases) CycleToPlus()

CycleToPlus grabs current Cycle values into the Plus phase values

func (*AvgMaxPhases) Zero added in v1.7.9 

func (am *AvgMaxPhases) Zero()

Zero does a full reset on everything -- for InitActs

type AxonLayer 

type AxonLayer interface {
	emer.Layer

	// AsAxon returns this layer as a axon.Layer -- so that the AxonLayer
	// interface does not need to include accessors to all the basic stuff
	AsAxon() *Layer

	// PostBuild performs special post-Build() configuration steps for specific algorithms,
	// using configuration data set in BuildConfig during the ConfigNet process.
	PostBuild()
}

AxonLayer defines the essential algorithmic API for Axon, at the layer level. These are the methods that the axon.Network calls on its layers at each step of processing. Other Layer types can selectively re-implement (override) these methods to modify the computation, while inheriting the basic behavior for non-overridden methods.

All of the structural API is in emer.Layer, which this interface also inherits for convenience.

type AxonNetwork 

type AxonNetwork interface {
	emer.Network

	// AsAxon returns this network as a axon.Network -- so that the
	// AxonNetwork interface does not need to include accessors
	// to all the basic stuff
	AsAxon() *Network
}

AxonNetwork defines the essential algorithmic API for Axon, at the network level. These are the methods that the user calls in their Sim code: * NewState * Cycle * NewPhase * DWt * WtFmDwt Because we don't want to have to force the user to use the interface cast in calling these methods, we provide Impl versions here that are the implementations which the user-facing method calls through the interface cast. Specialized algorithms should thus only change the Impl version, which is what is exposed here in this interface.

There is now a strong constraint that all Cycle level computation takes place in one pass at the Layer level, which greatly improves threading efficiency.

All of the structural API is in emer.Network, which this interface also inherits for convenience.

type AxonPrjn 

type AxonPrjn interface {
	emer.Prjn

	// AsAxon returns this prjn as a axon.Prjn -- so that the AxonPrjn
	// interface does not need to include accessors to all the basic stuff.
	AsAxon() *Prjn
}

AxonPrjn defines the essential algorithmic API for Axon, at the projection level. These are the methods that the axon.Layer calls on its prjns at each step of processing. Other Prjn types can selectively re-implement (override) these methods to modify the computation, while inheriting the basic behavior for non-overridden methods.

All of the structural API is in emer.Prjn, which this interface also inherits for convenience.

type AxonPrjns added in v1.6.14 

type AxonPrjns []*Prjn

func (*AxonPrjns) Add added in v1.6.14 

func (ap *AxonPrjns) Add(pj *Prjn)

type BLAPrjnParams added in v1.7.18 

type BLAPrjnParams struct {

	// [def: 0.01,1] use 0.01 for acquisition (don't unlearn) and 1 for extinction -- negative delta learning rate multiplier
	NegDeltaLRate float32 `def:"0.01,1" desc:"use 0.01 for acquisition (don't unlearn) and 1 for extinction -- negative delta learning rate multiplier"`

	// [def: 0.1] threshold on this layer's ACh level for trace learning updates
	AChThr float32 `def:"0.1" desc:"threshold on this layer's ACh level for trace learning updates"`

	// [def: 0,0.5] proportion of US time stimulus activity to use for the trace component of
	USTrace float32 `def:"0,0.5" desc:"proportion of US time stimulus activity to use for the trace component of "`
	// contains filtered or unexported fields
}

BLAPrjnParams has parameters for basolateral amygdala learning. Learning is driven by the Tr trace as function of ACh * Send Act recorded prior to US, and at US, recv unit delta: CaSpkP - SpkPrv times normalized GeIntNorm for recv unit credit assignment. The Learn.Trace.Tau time constant determines trace updating over trials when ACh is above threshold -- this determines strength of second-order conditioning -- default of 1 means none, but can be increased as needed.

func (*BLAPrjnParams) Defaults added in v1.7.18 

func (bp *BLAPrjnParams) Defaults()

func (*BLAPrjnParams) Update added in v1.7.18 

func (bp *BLAPrjnParams) Update()

type BurstParams added in v1.7.0 

type BurstParams struct {

	// [def: 0.1] [max: 1] Relative component of threshold on superficial activation value, below which it does not drive Burst (and above which, Burst = CaSpkP).  This is the distance between the average and maximum activation values within layer (e.g., 0 = average, 1 = max).  Overall effective threshold is MAX of relative and absolute thresholds.
	ThrRel float32 `` /* 348-byte string literal not displayed */

	// [def: 0.1] [min: 0] [max: 1] Absolute component of threshold on superficial activation value, below which it does not drive Burst (and above which, Burst = CaSpkP).  Overall effective threshold is MAX of relative and absolute thresholds.
	ThrAbs float32 `` /* 241-byte string literal not displayed */
	// contains filtered or unexported fields
}

BurstParams determine how the 5IB Burst activation is computed from CaSpkP integrated spiking values in Super layers -- thresholded.

func (*BurstParams) Defaults added in v1.7.0 

func (bp *BurstParams) Defaults()

func (*BurstParams) ThrFmAvgMax added in v1.7.0 

func (bp *BurstParams) ThrFmAvgMax(avg, mx float32) float32

ThrFmAvgMax returns threshold from average and maximum values

func (*BurstParams) Update added in v1.7.0 

func (bp *BurstParams) Update()

type CTParams added in v1.7.0 

type CTParams struct {

	// [def: 0.05,0.1,1,2] gain factor for context excitatory input, which is constant as compared to the spiking input from other projections, so it must be downscaled accordingly.  This can make a difference and may need to be scaled up or down.
	GeGain float32 `` /* 246-byte string literal not displayed */

	// [def: 0,50] decay time constant for context Ge input -- if > 0, decays over time so intrinsic circuit dynamics have to take over.  For single-step copy-based cases, set to 0, while longer-time-scale dynamics should use 50
	DecayTau float32 `` /* 227-byte string literal not displayed */

	// [view: -] 1 / tau
	DecayDt float32 `view:"-" json:"-" xml:"-" desc:"1 / tau"`
	// contains filtered or unexported fields
}

CTParams control the CT corticothalamic neuron special behavior

func (*CTParams) Defaults added in v1.7.0 

func (cp *CTParams) Defaults()

func (*CTParams) Update added in v1.7.0 

func (cp *CTParams) Update()

type CaLrnParams added in v1.5.1 

type CaLrnParams struct {

	// [def: 80] denomenator used for normalizing CaLrn, so the max is roughly 1 - 1.5 or so, which works best in terms of previous standard learning rules, and overall learning performance
	Norm float32 `` /* 188-byte string literal not displayed */

	// [def: true] use spikes to generate VGCC instead of actual VGCC current -- see SpkVGCCa for calcium contribution from each spike
	SpkVGCC slbool.Bool `` /* 133-byte string literal not displayed */

	// [def: 35] multiplier on spike for computing Ca contribution to CaLrn in SpkVGCC mode
	SpkVgccCa float32 `def:"35" desc:"multiplier on spike for computing Ca contribution to CaLrn in SpkVGCC mode"`

	// [def: 10] time constant of decay for VgccCa calcium -- it is highly transient around spikes, so decay and diffusion factors are more important than for long-lasting NMDA factor.  VgccCa is integrated separately int VgccCaInt prior to adding into NMDA Ca in CaLrn
	VgccTau float32 `` /* 268-byte string literal not displayed */

	// [view: inline] time constants for integrating CaLrn across M, P and D cascading levels
	Dt kinase.CaDtParams `view:"inline" desc:"time constants for integrating CaLrn across M, P and D cascading levels"`

	// [def: 0.01,0.02,0.5] Threshold on CaSpkP CaSpkD value for updating synapse-level Ca values (SynCa) -- this is purely a performance optimization that excludes random infrequent spikes -- 0.05 works well on larger networks but not smaller, which require the .01 default.
	UpdtThr float32 `` /* 274-byte string literal not displayed */

	// [view: -] rate = 1 / tau
	VgccDt float32 `view:"-" json:"-" xml:"-" inactive:"+" desc:"rate = 1 / tau"`

	// [view: -] = 1 / Norm
	NormInv float32 `view:"-" json:"-" xml:"-" inactive:"+" desc:"= 1 / Norm"`
	// contains filtered or unexported fields
}

CaLrnParams parameterizes the neuron-level calcium signals driving learning: CaLrn = NMDA + VGCC Ca sources, where VGCC can be simulated from spiking or use the more complex and dynamic VGCC channel directly. CaLrn is then integrated in a cascading manner at multiple time scales: CaM (as in calmodulin), CaP (ltP, CaMKII, plus phase), CaD (ltD, DAPK1, minus phase).

func (*CaLrnParams) CaLrns added in v1.8.0 

func (np *CaLrnParams) CaLrns(ctx *Context, ni, di uint32)

CaLrns updates the CaLrn value and its cascaded values, based on NMDA, VGCC Ca it first calls VgccCa to update the spike-driven version of that variable, and perform its time-integration.

func (*CaLrnParams) Defaults added in v1.5.1 

func (np *CaLrnParams) Defaults()

func (*CaLrnParams) Update added in v1.5.1 

func (np *CaLrnParams) Update()

func (*CaLrnParams) VgccCaFmSpike added in v1.8.0 

func (np *CaLrnParams) VgccCaFmSpike(ctx *Context, ni, di uint32)

VgccCa updates the simulated VGCC calcium from spiking, if that option is selected, and performs time-integration of VgccCa

type CaSpkParams added in v1.5.1 

type CaSpkParams struct {

	// [def: 8,12] gain multiplier on spike for computing CaSpk: increasing this directly affects the magnitude of the trace values, learning rate in Target layers, and other factors that depend on CaSpk values: RLRate, UpdtThr.  Prjn.KinaseCa.SpikeG provides an additional gain factor specific to the synapse-level trace factors, without affecting neuron-level CaSpk values.  Larger networks require higher gain factors at the neuron level -- 12, vs 8 for smaller.
	SpikeG float32 `` /* 464-byte string literal not displayed */

	// [def: 30] [min: 1] time constant for integrating spike-driven calcium trace at sender and recv neurons, CaSyn, which then drives synapse-level integration of the joint pre * post synapse-level activity, in cycles (msec).  Note: if this param is changed, then there will be a change in effective learning rate that can be compensated for by multiplying PrjnParams.Learn.KinaseCa.SpikeG by sqrt(30 / sqrt(SynTau)
	SynTau float32 `` /* 415-byte string literal not displayed */

	// [view: -] rate = 1 / tau
	SynDt float32 `view:"-" json:"-" xml:"-" inactive:"+" desc:"rate = 1 / tau"`

	// [view: inline] time constants for integrating CaSpk across M, P and D cascading levels -- these are typically the same as in CaLrn and Prjn level for synaptic integration, except for the M factor.
	Dt kinase.CaDtParams `` /* 202-byte string literal not displayed */
	// contains filtered or unexported fields
}

CaSpkParams parameterizes the neuron-level spike-driven calcium signals, starting with CaSyn that is integrated at the neuron level and drives synapse-level, pre * post Ca integration, which provides the Tr trace that multiplies error signals, and drives learning directly for Target layers. CaSpk* values are integrated separately at the Neuron level and used for UpdtThr and RLRate as a proxy for the activation (spiking) based learning signal.

func (*CaSpkParams) CaFmSpike added in v1.5.1 

func (np *CaSpkParams) CaFmSpike(ctx *Context, ni, di uint32)

CaFmSpike computes CaSpk* and CaSyn calcium signals based on current spike.

func (*CaSpkParams) Defaults added in v1.5.1 

func (np *CaSpkParams) Defaults()

func (*CaSpkParams) Update added in v1.5.1 

func (np *CaSpkParams) Update()

type ClampParams 

type ClampParams struct {

	// is this a clamped input layer?  set automatically based on layer type at initialization
	IsInput slbool.Bool `inactive:"+" desc:"is this a clamped input layer?  set automatically based on layer type at initialization"`

	// is this a target layer?  set automatically based on layer type at initialization
	IsTarget slbool.Bool `inactive:"+" desc:"is this a target layer?  set automatically based on layer type at initialization"`

	// [def: 0.8,1.5] amount of Ge driven for clamping -- generally use 0.8 for Target layers, 1.5 for Input layers
	Ge float32 `def:"0.8,1.5" desc:"amount of Ge driven for clamping -- generally use 0.8 for Target layers, 1.5 for Input layers"`

	// [def: false] [view: add external conductance on top of any existing -- generally this is not a good idea for target layers (creates a main effect that learning can never match), but may be ok for input layers]
	Add slbool.Bool `` /* 207-byte string literal not displayed */

	// [def: 0.5] threshold on neuron Act activity to count as active for computing error relative to target in PctErr method
	ErrThr float32 `def:"0.5" desc:"threshold on neuron Act activity to count as active for computing error relative to target in PctErr method"`
	// contains filtered or unexported fields
}

ClampParams specify how external inputs drive excitatory conductances (like a current clamp) -- either adds or overwrites existing conductances. Noise is added in either case.

func (*ClampParams) Defaults 

func (cp *ClampParams) Defaults()

func (*ClampParams) Update 

func (cp *ClampParams) Update()

type Context added in v1.7.0 

type Context struct {

	// current evaluation mode, e.g., Train, Test, etc
	Mode etime.Modes `desc:"current evaluation mode, e.g., Train, Test, etc"`

	// if true, the model is being run in a testing mode, so no weight changes or other associated computations are needed.  this flag should only affect learning-related behavior.  Is automatically updated based on Mode != Train
	Testing slbool.Bool `` /* 242-byte string literal not displayed */

	// phase counter: typicaly 0-1 for minus-plus but can be more phases for other algorithms
	Phase int32 `desc:"phase counter: typicaly 0-1 for minus-plus but can be more phases for other algorithms"`

	// true if this is the plus phase, when the outcome / bursting is occurring, driving positive learning -- else minus phase
	PlusPhase slbool.Bool `` /* 126-byte string literal not displayed */

	// cycle within current phase -- minus or plus
	PhaseCycle int32 `desc:"cycle within current phase -- minus or plus"`

	// cycle counter: number of iterations of activation updating (settling) on the current state -- this counts time sequentially until reset with NewState
	Cycle int32 `` /* 156-byte string literal not displayed */

	// [def: 200] length of the theta cycle in terms of 1 msec Cycles -- some network update steps depend on doing something at the end of the theta cycle (e.g., CTCtxtPrjn).
	ThetaCycles int32 `` /* 173-byte string literal not displayed */

	// total cycle count -- increments continuously from whenever it was last reset -- typically this is number of milliseconds in simulation time -- is int32 and not uint32 b/c used with Synapse CaUpT which needs to have a -1 case for expired update time
	CyclesTotal int32 `` /* 255-byte string literal not displayed */

	// accumulated amount of time the network has been running, in simulation-time (not real world time), in seconds
	Time float32 `desc:"accumulated amount of time the network has been running, in simulation-time (not real world time), in seconds"`

	// total trial count -- increments continuously in NewState call *only in Train mode* from whenever it was last reset -- can be used for synchronizing weight updates across nodes
	TrialsTotal int32 `` /* 182-byte string literal not displayed */

	// [def: 0.001] amount of time to increment per cycle
	TimePerCycle float32 `def:"0.001" desc:"amount of time to increment per cycle"`

	// [def: 100] how frequently to perform slow adaptive processes such as synaptic scaling, inhibition adaptation, associated in the brain with sleep, in the SlowAdapt method.  This should be long enough for meaningful changes to accumulate -- 100 is default but could easily be longer in larger models.  Because SlowCtr is incremented by NData, high NData cases (e.g. 16) likely need to increase this value -- e.g., 400 seems to produce overall consistent results in various models.
	SlowInterval int32 `` /* 484-byte string literal not displayed */

	// counter for how long it has been since last SlowAdapt step.  Note that this is incremented by NData to maintain consistency across different values of this parameter.
	SlowCtr int32 `` /* 186-byte string literal not displayed */

	// synaptic calcium counter, which drives the CaUpT synaptic value to optimize updating of this computationally expensive factor. It is incremented by 1 for each cycle, and reset at the SlowInterval, at which point the synaptic calcium values are all reset.
	SynCaCtr float32 `` /* 274-byte string literal not displayed */

	// [view: inline] indexes and sizes of current network
	NetIdxs NetIdxs `view:"inline" desc:"indexes and sizes of current network"`

	// [view: -] stride offsets for accessing neuron variables
	NeuronVars NeuronVarStrides `view:"-" desc:"stride offsets for accessing neuron variables"`

	// [view: -] stride offsets for accessing neuron average variables
	NeuronAvgVars NeuronAvgVarStrides `view:"-" desc:"stride offsets for accessing neuron average variables"`

	// [view: -] stride offsets for accessing neuron indexes
	NeuronIdxs NeuronIdxStrides `view:"-" desc:"stride offsets for accessing neuron indexes"`

	// [view: -] stride offsets for accessing synapse variables
	SynapseVars SynapseVarStrides `view:"-" desc:"stride offsets for accessing synapse variables"`

	// [view: -] stride offsets for accessing synapse Ca variables
	SynapseCaVars SynapseCaStrides `view:"-" desc:"stride offsets for accessing synapse Ca variables"`

	// [view: -] stride offsets for accessing synapse indexes
	SynapseIdxs SynapseIdxStrides `view:"-" desc:"stride offsets for accessing synapse indexes"`

	// random counter -- incremented by maximum number of possible random numbers generated per cycle, regardless of how many are actually used -- this is shared across all layers so must encompass all possible param settings.
	RandCtr slrand.Counter `` /* 226-byte string literal not displayed */
	// contains filtered or unexported fields
}

Context contains all of the global context state info that is shared across every step of the computation. It is passed around to all relevant computational functions, and is updated on the CPU and synced to the GPU after every cycle. It is the *only* mechanism for communication from CPU to GPU. It contains timing, Testing vs. Training mode, random number context, global neuromodulation, etc.

func NewContext added in v1.7.0 

func NewContext() *Context

NewContext returns a new Time struct with default parameters

func (*Context) CopyNetStridesFrom added in v1.8.0 

func (ctx *Context) CopyNetStridesFrom(srcCtx *Context)

CopyNetStridesFrom copies strides and NetIdxs for accessing variables on a Network -- these must be set properly for the Network in question (from its Ctx field) before calling any compute methods with the context. See SetCtxStrides on Network.

func (*Context) CycleInc added in v1.7.0 

func (ctx *Context) CycleInc()

CycleInc increments at the cycle level

func (*Context) Defaults added in v1.7.0 

func (ctx *Context) Defaults()

Defaults sets default values

func (*Context) GlobalIdx added in v1.8.0 

func (ctx *Context) GlobalIdx(di uint32, gvar GlobalVars) uint32

GlobalIdx returns index into main global variables, before GvVtaDA

func (*Context) GlobalUSnegIdx added in v1.8.0 

func (ctx *Context) GlobalUSnegIdx(di uint32, gvar GlobalVars, negIdx uint32) uint32

GlobalUSnegIdx returns index into USneg global variables

func (*Context) GlobalUSposIdx added in v1.8.16 

func (ctx *Context) GlobalUSposIdx(di uint32, gvar GlobalVars, posIdx uint32) uint32

GlobalUSposIdx returns index into USpos, Drive, VSPatch global variables

func (*Context) GlobalVNFloats added in v1.8.0 

func (ctx *Context) GlobalVNFloats() uint32

GlobalVNFloats number of floats to allocate for Globals

func (*Context) NewPhase added in v1.7.0 

func (ctx *Context) NewPhase(plusPhase bool)

NewPhase resets PhaseCycle = 0 and sets the plus phase as specified

func (*Context) NewState added in v1.7.0 

func (ctx *Context) NewState(mode etime.Modes)

NewState resets counters at start of new state (trial) of processing. Pass the evaluation model associated with this new state -- if !Train then testing will be set to true.

func (*Context) Reset added in v1.7.0 

func (ctx *Context) Reset()

Reset resets the counters all back to zero

func (*Context) SetGlobalStrides added in v1.8.0 

func (ctx *Context) SetGlobalStrides()

SetGlobalStrides sets global variable access offsets and strides

func (*Context) SlowInc added in v1.8.0 

func (ctx *Context) SlowInc() bool

SlowInc increments the Slow counter and returns true if time to perform SlowAdapt functions (associated with sleep).

type CorSimStats added in v1.3.35 

type CorSimStats struct {

	// correlation (centered cosine aka normalized dot product) activation difference between ActP and ActM on this alpha-cycle for this layer -- computed by CorSimFmActs called by PlusPhase
	Cor float32 `` /* 203-byte string literal not displayed */

	// running average of correlation similarity between ActP and ActM -- computed with CorSim.Tau time constant in PlusPhase
	Avg float32 `` /* 138-byte string literal not displayed */

	// running variance of correlation similarity between ActP and ActM -- computed with CorSim.Tau time constant in PlusPhase
	Var float32 `` /* 139-byte string literal not displayed */
	// contains filtered or unexported fields
}

CorSimStats holds correlation similarity (centered cosine aka normalized dot product) statistics at the layer level

func (*CorSimStats) Init added in v1.3.35 

func (cd *CorSimStats) Init()

type DAModTypes added in v1.7.0 

type DAModTypes int32

DAModTypes are types of dopamine modulation of neural activity. 

const (
	// NoDAMod means there is no effect of dopamine on neural activity
	NoDAMod DAModTypes = iota

	// D1Mod is for neurons that primarily express dopamine D1 receptors,
	// which are excitatory from DA bursts, inhibitory from dips.
	// Cortical neurons can generally use this type, while subcortical
	// populations are more diverse in having both D1 and D2 subtypes.
	D1Mod

	// D2Mod is for neurons that primarily express dopamine D2 receptors,
	// which are excitatory from DA dips, inhibitory from bursts.
	D2Mod

	// D1AbsMod is like D1Mod, except the absolute value of DA is used
	// instead of the signed value.
	// There are a subset of DA neurons that send increased DA for
	// both negative and positive outcomes, targeting frontal neurons.
	D1AbsMod

	DAModTypesN
)

func (DAModTypes) Desc added in v1.8.14 

func (i DAModTypes) Desc() string

func (*DAModTypes) FromString added in v1.7.0 

func (i *DAModTypes) FromString(s string) error

func (DAModTypes) String added in v1.7.0 

func (i DAModTypes) String() string

type DecayParams added in v1.2.59 

type DecayParams struct {

	// [def: 0,0.2,0.5,1] [min: 0] [max: 1] proportion to decay most activation state variables toward initial values at start of every ThetaCycle (except those controlled separately below) -- if 1 it is effectively equivalent to full clear, resetting other derived values.  ISI is reset every AlphaCycle to get a fresh sample of activations (doesn't affect direct computation -- only readout).
	Act float32 `` /* 391-byte string literal not displayed */

	// [def: 0,0.6] [min: 0] [max: 1] proportion to decay long-lasting conductances, NMDA and GABA, and also the dendritic membrane potential -- when using random stimulus order, it is important to decay this significantly to allow a fresh start -- but set Act to 0 to enable ongoing activity to keep neurons in their sensitive regime.
	Glong float32 `` /* 332-byte string literal not displayed */

	// [def: 0] [min: 0] [max: 1] decay of afterhyperpolarization currents, including mAHP, sAHP, and KNa -- has a separate decay because often useful to have this not decay at all even if decay is on.
	AHP float32 `` /* 198-byte string literal not displayed */

	// [def: 0] [min: 0] [max: 1] decay of Ca variables driven by spiking activity used in learning: CaSpk* and Ca* variables. These are typically not decayed but may need to be in some situations.
	LearnCa float32 `` /* 194-byte string literal not displayed */

	// decay layer at end of ThetaCycle when there is a global reward -- true by default for PTPred, PTMaint and PFC Super layers
	OnRew slbool.Bool `` /* 129-byte string literal not displayed */
	// contains filtered or unexported fields
}

DecayParams control the decay of activation state in the DecayState function called in NewState when a new state is to be processed.

func (*DecayParams) Defaults added in v1.2.59 

func (dp *DecayParams) Defaults()

func (*DecayParams) Update added in v1.2.59 

func (dp *DecayParams) Update()

type DendParams added in v1.2.95 

type DendParams struct {

	// [def: 0.2,0.5] dendrite-specific strength multiplier of the exponential spiking drive on Vm -- e.g., .5 makes it half as strong as at the soma (which uses Gbar.L as a strength multiplier per the AdEx standard model)
	GbarExp float32 `` /* 221-byte string literal not displayed */

	// [def: 3,6] dendrite-specific conductance of Kdr delayed rectifier currents, used to reset membrane potential for dendrite -- applied for Tr msec
	GbarR float32 `` /* 150-byte string literal not displayed */

	// [def: 0,2] SST+ somatostatin positive slow spiking inhibition level specifically affecting dendritic Vm (VmDend) -- this is important for countering a positive feedback loop from NMDA getting stronger over the course of learning -- also typically requires SubMean = 1 for TrgAvgAct and learning to fully counter this feedback loop.
	SSGi float32 `` /* 337-byte string literal not displayed */

	// set automatically based on whether this layer has any recv projections that have a GType conductance type of Modulatory -- if so, then multiply GeSyn etc by GModSyn
	HasMod slbool.Bool `` /* 184-byte string literal not displayed */

	// multiplicative gain factor on the total modulatory input -- this can also be controlled by the PrjnScale.Abs factor on ModulatoryG inputs, but it is convenient to be able to control on the layer as well.
	ModGain float32 `` /* 210-byte string literal not displayed */

	// baseline modulatory level for modulatory effects -- net modulation is ModBase + ModGain * GModSyn
	ModBase float32 `desc:"baseline modulatory level for modulatory effects -- net modulation is ModBase + ModGain * GModSyn"`
	// contains filtered or unexported fields
}

DendParams are the parameters for updating dendrite-specific dynamics

func (*DendParams) Defaults added in v1.2.95 

func (dp *DendParams) Defaults()

func (*DendParams) Update added in v1.2.95 

func (dp *DendParams) Update()

type DriveParams added in v1.8.16 

type DriveParams struct {

	// minimum effective drive value -- this is an automatic baseline ensuring that a positive US results in at least some minimal level of reward.  Unlike Base values, this is not reflected in the activity of the drive values -- applies at the time of reward calculation as a minimum baseline.
	DriveMin float32 `` /* 294-byte string literal not displayed */

	// baseline levels for each drive -- what they naturally trend toward in the absence of any input.  Set inactive drives to 0 baseline, active ones typically elevated baseline (0-1 range).
	Base []float32 `` /* 191-byte string literal not displayed */

	// time constants in ThetaCycle (trial) units for natural update toward Base values -- 0 values means no natural update.
	Tau []float32 `desc:"time constants in ThetaCycle (trial) units for natural update toward Base values -- 0 values means no natural update."`

	// decrement in drive value when US is consumed, thus partially satisfying the drive -- positive values are subtracted from current Drive value.
	Satisfaction []float32 `` /* 148-byte string literal not displayed */

	// [view: -] 1/Tau
	Dt []float32 `view:"-" desc:"1/Tau"`
}

DriveParams manages the drive parameters for computing and updating drive state. Most of the params are for optional case where drives are automatically updated based on US consumption (which satisfies drives) and time passing (which increases drives).

func (*DriveParams) AddTo added in v1.8.16 

func (dp *DriveParams) AddTo(ctx *Context, di uint32, drv uint32, delta float32) float32

AddTo increments drive by given amount, subject to 0-1 range clamping. Returns new val.

func (*DriveParams) Alloc added in v1.8.16 

func (dp *DriveParams) Alloc(nDrives int)

func (*DriveParams) Defaults added in v1.8.16 

func (dp *DriveParams) Defaults()

func (*DriveParams) EffectiveDrive added in v1.8.16 

func (dp *DriveParams) EffectiveDrive(ctx *Context, di uint32, i uint32) float32

EffectiveDrive returns the Max of Drives at given index and DriveMin. note that index 0 is the novelty / curiosity drive, which doesn't use DriveMin.

func (*DriveParams) ExpStep added in v1.8.16 

func (dp *DriveParams) ExpStep(ctx *Context, di uint32, drv uint32, dt, base float32) float32

ExpStep updates drive with an exponential step with given dt value toward given baseline value.

func (*DriveParams) ExpStepAll added in v1.8.16 

func (dp *DriveParams) ExpStepAll(ctx *Context, di uint32)

ExpStepAll updates given drives with an exponential step using dt values toward baseline values.

func (*DriveParams) SoftAdd added in v1.8.16 

func (dp *DriveParams) SoftAdd(ctx *Context, di uint32, drv uint32, delta float32) float32

SoftAdd increments drive by given amount, using soft-bounding to 0-1 extremes. if delta is positive, multiply by 1-val, else val. Returns new val.

func (*DriveParams) ToBaseline added in v1.8.16 

func (dp *DriveParams) ToBaseline(ctx *Context, di uint32)

ToBaseline sets all drives to their baseline levels

func (*DriveParams) ToZero added in v1.8.16 

func (dp *DriveParams) ToZero(ctx *Context, di uint32)

ToZero sets all drives to 0

func (*DriveParams) Update added in v1.8.16 

func (dp *DriveParams) Update()

func (*DriveParams) VarToZero added in v1.8.16 

func (dp *DriveParams) VarToZero(ctx *Context, di uint32, gvar GlobalVars)

VarToZero sets all values of given drive-sized variable to 0

type DtParams 

type DtParams struct {

	// [def: 1,0.5] [min: 0] overall rate constant for numerical integration, for all equations at the unit level -- all time constants are specified in millisecond units, with one cycle = 1 msec -- if you instead want to make one cycle = 2 msec, you can do this globally by setting this integ value to 2 (etc).  However, stability issues will likely arise if you go too high.  For improved numerical stability, you may even need to reduce this value to 0.5 or possibly even lower (typically however this is not necessary).  MUST also coordinate this with network.time_inc variable to ensure that global network.time reflects simulated time accurately
	Integ float32 `` /* 649-byte string literal not displayed */

	// [def: 2.81] [min: 1] membrane potential time constant in cycles, which should be milliseconds typically (tau is roughly how long it takes for value to change significantly -- 1.4x the half-life) -- reflects the capacitance of the neuron in principle -- biological default for AdEx spiking model C = 281 pF = 2.81 normalized
	VmTau float32 `` /* 328-byte string literal not displayed */

	// [def: 5] [min: 1] dendritic membrane potential time constant in cycles, which should be milliseconds typically (tau is roughly how long it takes for value to change significantly -- 1.4x the half-life) -- reflects the capacitance of the neuron in principle -- biological default for AdEx spiking model C = 281 pF = 2.81 normalized
	VmDendTau float32 `` /* 335-byte string literal not displayed */

	// [def: 2] [min: 1] number of integration steps to take in computing new Vm value -- this is the one computation that can be most numerically unstable so taking multiple steps with proportionally smaller dt is beneficial
	VmSteps int32 `` /* 223-byte string literal not displayed */

	// [def: 5] [min: 1] time constant for decay of excitatory AMPA receptor conductance.
	GeTau float32 `def:"5" min:"1" desc:"time constant for decay of excitatory AMPA receptor conductance."`

	// [def: 7] [min: 1] time constant for decay of inhibitory GABAa receptor conductance.
	GiTau float32 `def:"7" min:"1" desc:"time constant for decay of inhibitory GABAa receptor conductance."`

	// [def: 40] [min: 1] time constant for integrating values over timescale of an individual input state (e.g., roughly 200 msec -- theta cycle), used in computing ActInt, GeInt from Ge, and GiInt from GiSyn -- this is used for scoring performance, not for learning, in cycles, which should be milliseconds typically (tau is roughly how long it takes for value to change significantly -- 1.4x the half-life),
	IntTau float32 `` /* 409-byte string literal not displayed */

	// [def: 20] [min: 1] time constant for integrating slower long-time-scale averages, such as nrn.ActAvg, Pool.ActsMAvg, ActsPAvg -- computed in NewState when a new input state is present (i.e., not msec but in units of a theta cycle) (tau is roughly how long it takes for value to change significantly) -- set lower for smaller models
	LongAvgTau float32 `` /* 336-byte string literal not displayed */

	// [def: 10] [min: 0] cycle to start updating the SpkMaxCa, SpkMax values within a theta cycle -- early cycles often reflect prior state
	MaxCycStart int32 `` /* 138-byte string literal not displayed */

	// [view: -] nominal rate = Integ / tau
	VmDt float32 `view:"-" json:"-" xml:"-" desc:"nominal rate = Integ / tau"`

	// [view: -] nominal rate = Integ / tau
	VmDendDt float32 `view:"-" json:"-" xml:"-" desc:"nominal rate = Integ / tau"`

	// [view: -] 1 / VmSteps
	DtStep float32 `view:"-" json:"-" xml:"-" desc:"1 / VmSteps"`

	// [view: -] rate = Integ / tau
	GeDt float32 `view:"-" json:"-" xml:"-" desc:"rate = Integ / tau"`

	// [view: -] rate = Integ / tau
	GiDt float32 `view:"-" json:"-" xml:"-" desc:"rate = Integ / tau"`

	// [view: -] rate = Integ / tau
	IntDt float32 `view:"-" json:"-" xml:"-" desc:"rate = Integ / tau"`

	// [view: -] rate = 1 / tau
	LongAvgDt float32 `view:"-" json:"-" xml:"-" desc:"rate = 1 / tau"`
}

DtParams are time and rate constants for temporal derivatives in Axon (Vm, G)

func (*DtParams) AvgVarUpdt added in v1.2.45 

func (dp *DtParams) AvgVarUpdt(avg, vr *float32, val float32)

AvgVarUpdt updates the average and variance from current value, using LongAvgDt

func (*DtParams) Defaults 

func (dp *DtParams) Defaults()

func (*DtParams) GeSynFmRaw added in v1.2.97 

func (dp *DtParams) GeSynFmRaw(geSyn, geRaw float32) float32

GeSynFmRaw integrates a synaptic conductance from raw spiking using GeTau

func (*DtParams) GeSynFmRawSteady added in v1.5.12 

func (dp *DtParams) GeSynFmRawSteady(geRaw float32) float32

GeSynFmRawSteady returns the steady-state GeSyn that would result from receiving a steady increment of GeRaw every time step = raw * GeTau. dSyn = Raw - dt*Syn; solve for dSyn = 0 to get steady state: dt*Syn = Raw; Syn = Raw / dt = Raw * Tau

func (*DtParams) GiSynFmRaw added in v1.2.97 

func (dp *DtParams) GiSynFmRaw(giSyn, giRaw float32) float32

GiSynFmRaw integrates a synaptic conductance from raw spiking using GiTau

func (*DtParams) GiSynFmRawSteady added in v1.5.12 

func (dp *DtParams) GiSynFmRawSteady(giRaw float32) float32

GiSynFmRawSteady returns the steady-state GiSyn that would result from receiving a steady increment of GiRaw every time step = raw * GiTau. dSyn = Raw - dt*Syn; solve for dSyn = 0 to get steady state: dt*Syn = Raw; Syn = Raw / dt = Raw * Tau

func (*DtParams) Update 

func (dp *DtParams) Update()

type GPLayerTypes added in v1.7.0 

type GPLayerTypes int32

GPLayerTypes is a GPLayer axon-specific layer type enum. 

const (
	// GPeOut is Outer layer of GPe neurons, receiving inhibition from MtxGo
	GPeOut GPLayerTypes = iota

	// GPeIn is Inner layer of GPe neurons, receiving inhibition from GPeOut and MtxNo
	GPeIn

	// GPeTA is arkypallidal layer of GPe neurons, receiving inhibition from GPeIn
	// and projecting inhibition to Mtx
	GPeTA

	// GPi is the inner globus pallidus, functionally equivalent to SNr,
	// receiving from MtxGo and GPeIn, and sending inhibition to VThal
	GPi

	GPLayerTypesN
)

The GPLayer types

func (GPLayerTypes) Desc added in v1.8.14 

func (i GPLayerTypes) Desc() string

func (*GPLayerTypes) FromString added in v1.7.0 

func (i *GPLayerTypes) FromString(s string) error

func (GPLayerTypes) MarshalJSON added in v1.7.0 

func (ev GPLayerTypes) MarshalJSON() ([]byte, error)

func (GPLayerTypes) String added in v1.7.0 

func (i GPLayerTypes) String() string

func (*GPLayerTypes) UnmarshalJSON added in v1.7.0 

func (ev *GPLayerTypes) UnmarshalJSON(b []byte) error

type GPParams added in v1.7.0 

type GPParams struct {

	// [view: inline] [viewif: LayType=GPLayer] type of GP Layer -- must set during config using SetBuildConfig of GPType.
	GPType GPLayerTypes `viewif:"LayType=GPLayer" view:"inline" desc:"type of GP Layer -- must set during config using SetBuildConfig of GPType."`
	// contains filtered or unexported fields
}

GPLayer represents a globus pallidus layer, including: GPeOut, GPeIn, GPeTA (arkypallidal), and GPi (see GPType for type). Typically just a single unit per Pool representing a given stripe.

func (*GPParams) Defaults added in v1.7.0 

func (gp *GPParams) Defaults()

func (*GPParams) Update added in v1.7.0 

func (gp *GPParams) Update()

type GPU added in v1.7.9 

type GPU struct {

	// if true, actually use the GPU
	On          bool `desc:"if true, actually use the GPU"`
	RecFunTimes bool `` /* 284-byte string literal not displayed */

	// if true, process each cycle one at a time.  Otherwise, 10 cycles at a time are processed in one batch.
	CycleByCycle bool `desc:"if true, process each cycle one at a time.  Otherwise, 10 cycles at a time are processed in one batch."`

	// [view: -] the network we operate on -- we live under this net
	Net *Network `view:"-" desc:"the network we operate on -- we live under this net"`

	// [view: -] the context we use
	Ctx *Context `view:"-" desc:"the context we use"`

	// [view: -] the vgpu compute system
	Sys *vgpu.System `view:"-" desc:"the vgpu compute system"`

	// [view: -] VarSet = 0: the uniform LayerParams
	Params *vgpu.VarSet `view:"-" desc:"VarSet = 0: the uniform LayerParams"`

	// [view: -] VarSet = 1: the storage indexes and PrjnParams
	Idxs *vgpu.VarSet `view:"-" desc:"VarSet = 1: the storage indexes and PrjnParams"`

	// [view: -] VarSet = 2: the Storage buffer for RW state structs and neuron floats
	Structs *vgpu.VarSet `view:"-" desc:"VarSet = 2: the Storage buffer for RW state structs and neuron floats"`

	// [view: -] Varset = 3: the Storage buffer for synapses
	Syns *vgpu.VarSet `view:"-" desc:"Varset = 3: the Storage buffer for synapses"`

	// [view: -] Varset = 4: the Storage buffer for SynCa banks
	SynCas *vgpu.VarSet `view:"-" desc:"Varset = 4: the Storage buffer for SynCa banks"`

	// [view: -] for sequencing commands
	Semaphores map[string]vk.Semaphore `view:"-" desc:"for sequencing commands"`

	// [def: 64] [view: -] number of warp threads -- typically 64 -- must update all hlsl files if changed!
	NThreads int `view:"-" inactive:"-" def:"64" desc:"number of warp threads -- typically 64 -- must update all hlsl files if changed!"`

	// [view: -] maximum number of bytes per individual storage buffer element, from GPUProps.Limits.MaxStorageBufferRange
	MaxBufferBytes uint32 `view:"-" desc:"maximum number of bytes per individual storage buffer element, from GPUProps.Limits.MaxStorageBufferRange"`

	// [view: -] bank of floats for GPU access
	SynapseCas0 []float32 `view:"-" desc:"bank of floats for GPU access"`

	// [view: -] bank of floats for GPU access
	SynapseCas1 []float32 `view:"-" desc:"bank of floats for GPU access"`

	// [view: -] bank of floats for GPU access
	SynapseCas2 []float32 `view:"-" desc:"bank of floats for GPU access"`

	// [view: -] bank of floats for GPU access
	SynapseCas3 []float32 `view:"-" desc:"bank of floats for GPU access"`

	// [view: -] bank of floats for GPU access
	SynapseCas4 []float32 `view:"-" desc:"bank of floats for GPU access"`

	// [view: -] bank of floats for GPU access
	SynapseCas5 []float32 `view:"-" desc:"bank of floats for GPU access"`

	// [view: -] bank of floats for GPU access
	SynapseCas6 []float32 `view:"-" desc:"bank of floats for GPU access"`

	// [view: -] bank of floats for GPU access
	SynapseCas7 []float32 `view:"-" desc:"bank of floats for GPU access"`

	// [view: -] tracks var binding
	DidBind map[string]bool `view:"-" desc:"tracks var binding"`
}

GPU manages all of the GPU-based computation for a given Network. Lives within the network.

func (*GPU) Config added in v1.7.9 

func (gp *GPU) Config(ctx *Context, net *Network)

Config configures the network -- must call on an already-built network

func (*GPU) ConfigSynCaBuffs added in v1.8.2 

func (gp *GPU) ConfigSynCaBuffs()

ConfigSynCaBuffs configures special SynapseCas buffers needed for larger memory access

func (*GPU) CopyContextFmStaging added in v1.7.9 

func (gp *GPU) CopyContextFmStaging()

CopyContextFmStaging copies Context from staging to CPU, after Sync back down.

func (*GPU) CopyContextToStaging added in v1.7.9 

func (gp *GPU) CopyContextToStaging()

CopyContextToStaging copies current context to staging from CPU. Must call SyncMemToGPU after this (see SyncContextToGPU). See SetContext if there is a new one.

func (*GPU) CopyExtsToStaging added in v1.7.9 

func (gp *GPU) CopyExtsToStaging()

CopyExtsToStaging copies external inputs to staging from CPU. Typically used in RunApplyExts which also does the Sync.

func (*GPU) CopyGBufToStaging added in v1.8.0 

func (gp *GPU) CopyGBufToStaging()

CopyGBufToStaging copies the GBuf and GSyns memory to staging.

func (*GPU) CopyIdxsToStaging added in v1.7.9 

func (gp *GPU) CopyIdxsToStaging()

CopyIdxsToStaging is only called when the network is built to copy the indexes specifying connectivity etc to staging from CPU.

func (*GPU) CopyLayerStateFmStaging added in v1.7.9 

func (gp *GPU) CopyLayerStateFmStaging()

CopyLayerStateFmStaging copies Context, LayerVals and Pools from staging to CPU, after Sync.

func (*GPU) CopyLayerValsFmStaging added in v1.7.9 

func (gp *GPU) CopyLayerValsFmStaging()

CopyLayerValsFmStaging copies LayerVals from staging to CPU, after Sync back down.

func (*GPU) CopyLayerValsToStaging added in v1.7.9 

func (gp *GPU) CopyLayerValsToStaging()

CopyLayerValsToStaging copies LayerVals to staging from CPU. Must call SyncMemToGPU after this (see SyncLayerValsToGPU).

func (*GPU) CopyNeuronsFmStaging added in v1.7.9 

func (gp *GPU) CopyNeuronsFmStaging()

CopyNeuronsFmStaging copies Neurons from staging to CPU, after Sync back down.

func (*GPU) CopyNeuronsToStaging added in v1.7.9 

func (gp *GPU) CopyNeuronsToStaging()

CopyNeuronsToStaging copies neuron state up to staging from CPU. Must call SyncMemToGPU after this (see SyncNeuronsToGPU).

func (*GPU) CopyParamsToStaging added in v1.7.9 

func (gp *GPU) CopyParamsToStaging()

CopyParamsToStaging copies the LayerParams and PrjnParams to staging from CPU. Must call SyncMemToGPU after this (see SyncParamsToGPU).

func (*GPU) CopyPoolsFmStaging added in v1.7.9 

func (gp *GPU) CopyPoolsFmStaging()

CopyPoolsFmStaging copies Pools from staging to CPU, after Sync back down.

func (*GPU) CopyPoolsToStaging added in v1.7.9 

func (gp *GPU) CopyPoolsToStaging()

CopyPoolsToStaging copies Pools to staging from CPU. Must call SyncMemToGPU after this (see SyncPoolsToGPU).

func (*GPU) CopyStateFmStaging added in v1.7.9 

func (gp *GPU) CopyStateFmStaging()

CopyStateFmStaging copies Context, LayerVals, Pools, and Neurons from staging to CPU, after Sync.

func (*GPU) CopyStateToStaging added in v1.7.9 

func (gp *GPU) CopyStateToStaging()

CopyStateToStaging copies LayerVals, Pools, Neurons state to staging from CPU. this is typically sufficient for most syncing -- only missing the Synapses which must be copied separately. Must call SyncMemToGPU after this (see SyncStateToGPU).

func (*GPU) CopySynCaFmStaging added in v1.8.1 

func (gp *GPU) CopySynCaFmStaging()

CopySynCaFmStaging copies the SynCa variables to GPU, which are per-Di (even larger). This is only used for GUI viewing -- SynCa vars otherwise managed entirely on GPU.

func (*GPU) CopySynCaToStaging added in v1.8.1 

func (gp *GPU) CopySynCaToStaging()

CopySynCaToStaging copies the SynCa variables to GPU, which are per-Di (even larger). This is only used for initialization -- SynCa vars otherwise managed entirely on GPU. Must call SyncMemToGPU after this (see SyncSynCaToGPU).

func (*GPU) CopySynapsesFmStaging added in v1.7.9 

func (gp *GPU) CopySynapsesFmStaging()

CopySynapsesFmStaging copies Synapses from staging to CPU, after Sync back down. Does not copy SynCa synapse state -- see SynCa methods.

func (*GPU) CopySynapsesToStaging added in v1.7.9 

func (gp *GPU) CopySynapsesToStaging()

CopySynapsesToStaging copies the synapse memory to staging (large). Does not copy SynCa synapse state -- see SynCa methods. This is not typically needed except when weights are initialized or for the Slow weight update processes that are not on GPU. Must call SyncMemToGPU after this (see SyncSynapsesToGPU).

func (*GPU) Destroy added in v1.7.9 

func (gp *GPU) Destroy()

Destroy should be called to release all the resources allocated by the network

func (*GPU) RunApplyExts added in v1.7.9 

func (gp *GPU) RunApplyExts()

RunApplyExts copies Exts external input memory to the GPU and then runs the ApplyExts shader that applies those external inputs to the GPU-side neuron state. The caller must check the On flag before running this, to use CPU vs. GPU

func (*GPU) RunApplyExtsCmd added in v1.7.11 

func (gp *GPU) RunApplyExtsCmd() vk.CommandBuffer

RunApplyExtsCmd returns the commands to copy Exts external input memory to the GPU and then runs the ApplyExts shader that applies those external inputs to the GPU-side neuron state. The caller must check the On flag before running this, to use CPU vs. GPU

func (*GPU) RunCycle added in v1.7.9 

func (gp *GPU) RunCycle()

RunCycle is the main cycle-level update loop for updating one msec of neuron state. It copies current Context up to GPU for updated Cycle counter state and random number state, Different versions of the code are run depending on various flags. By default, it will run the entire minus and plus phase in one big chunk. The caller must check the On flag before running this, to use CPU vs. GPU.

func (*GPU) RunCycleOne added in v1.7.9 

func (gp *GPU) RunCycleOne()

RunCycleOne does one cycle of updating in an optimized manner using Events to sequence each of the pipeline calls. It is for CycleByCycle mode and syncs back full state every cycle.

func (*GPU) RunCycleOneCmd added in v1.7.11 

func (gp *GPU) RunCycleOneCmd() vk.CommandBuffer

RunCycleOneCmd returns commands to do one cycle of updating in an optimized manner using Events to sequence each of the pipeline calls. It is for CycleByCycle mode and syncs back full state every cycle.

func (*GPU) RunCycleSeparateFuns added in v1.7.9 

func (gp *GPU) RunCycleSeparateFuns()

RunCycleSeparateFuns does one cycle of updating in a very slow manner that allows timing to be recorded for each function call, for profiling.

func (*GPU) RunCycles added in v1.7.9 

func (gp *GPU) RunCycles()

RunCycles does multiple cycles of updating in one chunk

func (*GPU) RunCyclesCmd added in v1.7.11 

func (gp *GPU) RunCyclesCmd() vk.CommandBuffer

RunCyclesCmd returns the RunCycles commands to do multiple cycles of updating in one chunk

func (*GPU) RunDWt added in v1.7.9 

func (gp *GPU) RunDWt()

RunDWt runs the DWt shader to compute weight changes. The caller must check the On flag before running this, to use CPU vs. GPU

func (*GPU) RunDWtCmd added in v1.8.0 

func (gp *GPU) RunDWtCmd() vk.CommandBuffer

RunDWtCmd returns the commands to run the DWt shader to compute weight changes.

func (*GPU) RunMinusPhase added in v1.7.9 

func (gp *GPU) RunMinusPhase()

RunMinusPhase runs the MinusPhase shader to update snapshot variables at the end of the minus phase. All non-synapse state is copied back down after this, so it is available for action calls The caller must check the On flag before running this, to use CPU vs. GPU

func (*GPU) RunMinusPhaseCmd added in v1.7.11 

func (gp *GPU) RunMinusPhaseCmd() vk.CommandBuffer

RunMinusPhaseCmd returns the commands to run the MinusPhase shader to update snapshot variables at the end of the minus phase. All non-synapse state is copied back down after this, so it is available for action calls

func (*GPU) RunNewState added in v1.7.9 

func (gp *GPU) RunNewState()

RunNewState runs the NewState shader to initialize state at start of new ThetaCycle trial. The caller must check the On flag before running this, to use CPU vs. GPU

func (*GPU) RunNewStateCmd added in v1.8.0 

func (gp *GPU) RunNewStateCmd() vk.CommandBuffer

RunNewStateCmd returns the commands to run the NewState shader to update variables at the start of a new trial.

func (*GPU) RunPipelineMemWait added in v1.8.2 

func (gp *GPU) RunPipelineMemWait(cmd vk.CommandBuffer, name string, n int)

RunPipelineMemWait records command to run given pipeline with a WaitMemWriteRead after it, so subsequent pipeline run will have access to values updated by this command.

func (*GPU) RunPipelineNoWait added in v1.8.2 

func (gp *GPU) RunPipelineNoWait(cmd vk.CommandBuffer, name string, n int)

RunPipelineNoWait records command to run given pipeline without any waiting after it for writes to complete. This should be the last command in the sequence.

func (*GPU) RunPipelineOffset added in v1.8.2 

func (gp *GPU) RunPipelineOffset(cmd vk.CommandBuffer, name string, n, off int)

RunPipelineOffset records command to run given pipeline with a push constant offset for the starting index to compute. This is needed when the total number of dispatch indexes exceeds GPU.MaxComputeWorkGroupCount1D. Does NOT wait for writes, assuming a parallel launch of all.

func (*GPU) RunPipelineWait added in v1.7.9 

func (gp *GPU) RunPipelineWait(name string, n int)

RunPipelineWait runs given pipeline in "single shot" mode, which is maximally inefficient if multiple commands need to be run. This is the only mode in which timer information is available.

func (*GPU) RunPlusPhase added in v1.7.9 

func (gp *GPU) RunPlusPhase()

RunPlusPhase runs the PlusPhase shader to update snapshot variables and do additional stats-level processing at end of the plus phase. All non-synapse state is copied back down after this. The caller must check the On flag before running this, to use CPU vs. GPU

func (*GPU) RunPlusPhaseCmd added in v1.7.11 

func (gp *GPU) RunPlusPhaseCmd() vk.CommandBuffer

RunPlusPhaseCmd returns the commands to run the PlusPhase shader to update snapshot variables and do additional stats-level processing at end of the plus phase. All non-synapse state is copied back down after this.

func (*GPU) RunPlusPhaseStart added in v1.7.10 

func (gp *GPU) RunPlusPhaseStart()

RunPlusPhaseStart runs the PlusPhaseStart shader does updating at the start of the plus phase: applies Target inputs as External inputs.

func (*GPU) RunWtFmDWt added in v1.7.9 

func (gp *GPU) RunWtFmDWt()

RunWtFmDWt runs the WtFmDWt shader to update weights from weigh changes. The caller must check the On flag before running this, to use CPU vs. GPU

func (*GPU) RunWtFmDWtCmd added in v1.7.11 

func (gp *GPU) RunWtFmDWtCmd() vk.CommandBuffer

RunWtFmDWtCmd returns the commands to run the WtFmDWt shader to update weights from weight changes. This also syncs neuron state from CPU -> GPU because TrgAvgFmD has updated that state.

func (*GPU) SetContext added in v1.7.9 

func (gp *GPU) SetContext(ctx *Context)

SetContext sets our context to given context and syncs it to the GPU. Typically a single context is used as it must be synced into the GPU. The GPU never writes to the CPU

func (*GPU) StartRun added in v1.7.9 

func (gp *GPU) StartRun(cmd vk.CommandBuffer)

StartRun resets the given command buffer in preparation for recording commands for a multi-step run. It is much more efficient to record all commands to one buffer, and use Events to synchronize the steps between them, rather than using semaphores. The submit call is by far the most expensive so that should only happen once!

func (*GPU) SynCaBuff added in v1.8.2 

func (gp *GPU) SynCaBuff(idx uint32) []float32

func (*GPU) SynDataNs added in v1.8.2 

func (gp *GPU) SynDataNs() (nCmd, nPer, nLast int)

SynDataNs returns the numbers for processing SynapseCas vars = Synapses * MaxData. Can exceed thread count limit and require multiple command launches with different offsets. The offset is in terms of synapse index, so everything is computed in terms of synapse indexes, with MaxData then multiplied to get final values. nCmd = number of command launches, nPer = number of synapses per cmd, nLast = number of synapses for last command launch.

func (*GPU) SyncAllFmGPU added in v1.7.9 

func (gp *GPU) SyncAllFmGPU()

SyncAllFmCPU copies State except Context plus Synapses from GPU to CPU. This is called before SlowAdapt, which is run CPU-side

func (*GPU) SyncAllToGPU added in v1.7.9 

func (gp *GPU) SyncAllToGPU()

SyncAllToGPU copies LayerVals, Pools, Neurons, Synapses to GPU. Calls SyncMemToGPU -- use when this is the only copy taking place.

func (*GPU) SyncContextFmGPU added in v1.7.9 

func (gp *GPU) SyncContextFmGPU()

SyncContextFmGPU copies Context from GPU to CPU. This is done at the end of each cycle to get state back from GPU for CPU-side computations. Use only when only thing being copied -- more efficient to get all at once. e.g. see SyncStateFmGPU

func (*GPU) SyncContextToGPU added in v1.7.9 

func (gp *GPU) SyncContextToGPU()

SyncContextToGPU copies current context to GPU from CPU. Calls SyncMemToGPU -- use when this is the only copy taking place. See SetContext if there is a new one.

func (*GPU) SyncGBufToGPU added in v1.7.9 

func (gp *GPU) SyncGBufToGPU()

SyncGBufToGPU copies the GBuf and GSyns memory to the GPU.

func (*GPU) SyncLayerStateFmGPU added in v1.7.9 

func (gp *GPU) SyncLayerStateFmGPU()

SyncLayerStateFmCPU copies Context, LayerVals, and Pools from GPU to CPU. This is the main GPU->CPU sync step automatically called after each Cycle.

func (*GPU) SyncLayerValsFmGPU added in v1.7.9 

func (gp *GPU) SyncLayerValsFmGPU()

SyncLayerValsFmGPU copies LayerVals from GPU to CPU. This is done at the end of each cycle to get state back from staging for CPU-side computations. Use only when only thing being copied -- more efficient to get all at once. e.g. see SyncStateFmGPU

func (*GPU) SyncLayerValsToGPU added in v1.7.9 

func (gp *GPU) SyncLayerValsToGPU()

SyncLayerValsToGPU copies LayerVals to GPU from CPU. Calls SyncMemToGPU -- use when this is the only copy taking place.

func (*GPU) SyncMemToGPU added in v1.7.9 

func (gp *GPU) SyncMemToGPU()

SyncMemToGPU synchronizes any staging memory buffers that have been updated with a Copy function, actually sending the updates from the staging -> GPU. The CopyTo commands just copy Network-local data to a staging buffer, and this command then actually moves that onto the GPU. In unified GPU memory architectures, this staging buffer is actually the same one used directly by the GPU -- otherwise it is a separate staging buffer.

func (*GPU) SyncNeuronsFmGPU added in v1.7.9 

func (gp *GPU) SyncNeuronsFmGPU()

SyncNeuronsFmGPU copies Neurons from GPU to CPU. Use only when only thing being copied -- more efficient to get all at once. e.g. see SyncStateFmGPU

func (*GPU) SyncNeuronsToGPU added in v1.7.9 

func (gp *GPU) SyncNeuronsToGPU()

SyncNeuronsToGPU copies neuron state up to GPU from CPU. Calls SyncMemToGPU -- use when this is the only copy taking place.

func (*GPU) SyncParamsToGPU added in v1.7.9 

func (gp *GPU) SyncParamsToGPU()

SyncParamsToGPU copies the LayerParams and PrjnParams to the GPU from CPU. Calls SyncMemToGPU -- use when this is the only copy taking place.

func (*GPU) SyncPoolsFmGPU added in v1.7.9 

func (gp *GPU) SyncPoolsFmGPU()

SyncPoolsFmGPU copies Pools from GPU to CPU. Use only when only thing being copied -- more efficient to get all at once. e.g. see SyncStateFmGPU

func (*GPU) SyncPoolsToGPU added in v1.7.9 

func (gp *GPU) SyncPoolsToGPU()

SyncPoolsToGPU copies Pools to GPU from CPU. Calls SyncMemToGPU -- use when this is the only copy taking place.

func (*GPU) SyncRegionStruct added in v1.7.9 

func (gp *GPU) SyncRegionStruct(vnm string) vgpu.MemReg

SyncRegionStruct returns the SyncRegion with error panic

func (*GPU) SyncRegionSynCas added in v1.8.1 

func (gp *GPU) SyncRegionSynCas(vnm string) vgpu.MemReg

SyncRegionSynCas returns the SyncRegion with error panic

func (*GPU) SyncRegionSyns added in v1.8.0 

func (gp *GPU) SyncRegionSyns(vnm string) vgpu.MemReg

SyncRegionSyns returns the SyncRegion with error panic

func (*GPU) SyncStateFmGPU added in v1.7.9 

func (gp *GPU) SyncStateFmGPU()

SyncStateFmCPU copies Neurons, LayerVals, and Pools from GPU to CPU. This is the main GPU->CPU sync step automatically called in PlusPhase.

func (*GPU) SyncStateGBufToGPU added in v1.8.0 

func (gp *GPU) SyncStateGBufToGPU()

SyncStateGBufToGPU copies LayVals, Pools, Neurons, GBuf state to GPU this is typically sufficient for most syncing -- only missing the Synapses which must be copied separately. Calls SyncMemToGPU -- use when this is the only copy taking place.

func (*GPU) SyncStateToGPU added in v1.7.9 

func (gp *GPU) SyncStateToGPU()

SyncStateToGPU copies LayVals, Pools, Neurons state to GPU this is typically sufficient for most syncing -- only missing the Synapses which must be copied separately. Calls SyncMemToGPU -- use when this is the only copy taking place.

func (*GPU) SyncSynCaFmGPU added in v1.8.1 

func (gp *GPU) SyncSynCaFmGPU()

SyncSynCaFmGPU copies the SynCa variables to GPU, which are per-Di (even larger). This is only used for GUI viewing -- SynCa vars otherwise managed entirely on GPU. Use only when only thing being copied -- more efficient to get all at once.

func (*GPU) SyncSynCaToGPU added in v1.8.1 

func (gp *GPU) SyncSynCaToGPU()

SyncSynCaToGPU copies the SynCa variables to GPU, which are per-Di (even larger). This is only used for initialization -- SynCa vars otherwise managed entirely on GPU. Calls SyncMemToGPU -- use when this is the only copy taking place.

func (*GPU) SyncSynapsesFmGPU added in v1.7.9 

func (gp *GPU) SyncSynapsesFmGPU()

SyncSynapsesFmGPU copies Synapses from GPU to CPU. Does not copy SynCa synapse state -- see SynCa methods. Use only when only thing being copied -- more efficient to get all at once.

func (*GPU) SyncSynapsesToGPU added in v1.7.9 

func (gp *GPU) SyncSynapsesToGPU()

SyncSynapsesToGPU copies the synapse memory to GPU (large). This is not typically needed except when weights are initialized or for the Slow weight update processes that are not on GPU. Calls SyncMemToGPU -- use when this is the only copy taking place.

func (*GPU) TestSynCa added in v1.8.2 

func (gp *GPU) TestSynCa() bool

TestSynCa tests writing to SynCa -- returns true if passed

func (*GPU) TestSynCaCmd added in v1.8.2 

func (gp *GPU) TestSynCaCmd() vk.CommandBuffer

type GScaleVals added in v1.2.37 

type GScaleVals struct {

	// scaling factor for integrating synaptic input conductances (G's), originally computed as a function of sending layer activity and number of connections, and typically adapted from there -- see Prjn.PrjnScale adapt params
	Scale float32 `` /* 240-byte string literal not displayed */

	// normalized relative proportion of total receiving conductance for this projection: PrjnScale.Rel / sum(PrjnScale.Rel across relevant prjns)
	Rel float32 `` /* 159-byte string literal not displayed */
	// contains filtered or unexported fields
}

GScaleVals holds the conductance scaling values. These are computed once at start and remain constant thereafter, and therefore belong on Params and not on PrjnVals.

type GiveUpParams added in v1.8.16 

type GiveUpParams struct {

	// [def: 1] threshold factor that multiplies integrated pvNeg value to establish a threshold for whether the integrated pvPos value is good enough to drive overall net positive reward
	NegThr float32 `` /* 186-byte string literal not displayed */

	// [def: 10] multiplier on pos - neg for logistic probability function -- higher gain values produce more binary give up behavior and lower values produce more graded stochastic behavior around the threshold
	Gain float32 `` /* 210-byte string literal not displayed */

	// minimum estimated PVpos value -- deals with any errors in the estimation process to make sure that erroneous GiveUp doesn't happen.
	MinPVposEst float32 `` /* 138-byte string literal not displayed */
}

GiveUpParams are parameters for computing when to give up

func (*GiveUpParams) Defaults added in v1.8.16 

func (gp *GiveUpParams) Defaults()

func (*GiveUpParams) Prob added in v1.8.16 

func (gp *GiveUpParams) Prob(pvDiff float32, rnd erand.Rand) (float32, bool)

type GlobalVars added in v1.8.0 

type GlobalVars int32

GlobalVars are network-wide variables, such as neuromodulators, reward, drives, etc including the state for the PVLV phasic dopamine model. 

const (

	// Rew is reward value -- this is set here in the Context struct, and the RL Rew layer grabs it from there -- must also set HasRew flag when rew is set -- otherwise is ignored.
	GvRew GlobalVars = iota

	// HasRew must be set to true when a reward is present -- otherwise Rew is ignored.  Also set when PVLV BOA model gives up.  This drives ACh release in the PVLV model.
	GvHasRew

	// RewPred is reward prediction -- computed by a special reward prediction layer
	GvRewPred

	// PrevPred is previous time step reward prediction -- e.g., for TDPredLayer
	GvPrevPred

	// HadRew is HasRew state from the previous trial -- copied from HasRew in NewState -- used for updating Effort, Urgency at start of new trial
	GvHadRew

	// DA is dopamine -- represents reward prediction error, signaled as phasic increases or decreases in activity relative to a tonic baseline, which is represented by a value of 0.  Released by the VTA -- ventral tegmental area, or SNc -- substantia nigra pars compacta.
	GvDA

	// ACh is acetylcholine -- activated by salient events, particularly at the onset of a reward / punishment outcome (US), or onset of a conditioned stimulus (CS).  Driven by BLA -> PPtg that detects changes in BLA activity, via LDTLayer type
	GvACh

	// NE is norepinepherine -- not yet in use
	GvNE

	// Ser is serotonin -- not yet in use
	GvSer

	// AChRaw is raw ACh value used in updating global ACh value by LDTLayer
	GvAChRaw

	// NotMaint is activity of the PTNotMaintLayer -- drives top-down inhibition of LDT layer / ACh activity.
	GvNotMaint

	// VSMatrixJustGated is VSMatrix just gated (to engage goal maintenance in PFC areas), set at end of plus phase -- this excludes any gating happening at time of US
	GvVSMatrixJustGated

	// VSMatrixHasGated is VSMatrix has gated since the last time HasRew was set (US outcome received or expected one failed to be received
	GvVSMatrixHasGated

	// CuriosityPoolGated is true if VSMatrixJustGated and the first pool representing the curiosity / novelty drive gated -- this can change the giving up Effort.Max parameter.
	GvCuriosityPoolGated

	// Time is raw time counter, incrementing upward during goal engaged window.
	// This is also copied directly into NegUS[0] which tracks time, but we maintain
	// a separate effort value to make it clearer.
	GvTime

	// Effort is raw effort counter -- incrementing upward for each effort step
	// during goal engaged window.
	// This is also copied directly into NegUS[1] which tracks effort, but we maintain
	// a separate effort value to make it clearer.
	GvEffort

	// UrgencyRaw is raw effort for urgency -- incrementing upward from effort
	// increments per step when _not_ goal engaged
	GvUrgencyRaw

	// Urgency is the overall urgency activity level (normalized 0-1),
	// computed from logistic function of GvUrgencyRaw
	GvUrgency

	// HasPosUS indicates has positive US on this trial -- drives goal accomplishment logic
	// and gating.
	GvHasPosUS

	// HadPosUS is state from the previous trial (copied from HasPosUS in NewState).
	GvHadPosUS

	// NegUSOutcome indicates that a strong negative US stimulus was experienced,
	// driving phasic ACh, VSMatrix gating to reset current goal engaged plan (if any),
	// and phasic dopamine based on the outcome.
	GvNegUSOutcome

	// HadNegUSOutcome is state from the previous trial (copied from NegUSOutcome in NewState)
	GvHadNegUSOutcome

	// PVposSum is total weighted positive valence primary value = sum of Weight * USpos * Drive
	GvPVposSum

	// PVpos is normalized positive valence primary value = (1 - 1/(1+PVposGain * PVposSum))
	GvPVpos

	// PVnegSum is total weighted negative valence primary value = sum of Weight * USneg
	GvPVnegSum

	// PVpos is normalized negative valence primary value = (1 - 1/(1+PVnegGain * PVnegSum))
	GvPVneg

	// PVposEst is the estimated PVpos value based on OFCposUSPT and VSMatrix gating
	GvPVposEst

	// PVposEstSum is the sum that goes into computing estimated PVpos
	// value based on OFCposUSPT and VSMatrix gating
	GvPVposEstSum

	// PVposEstDisc is the discounted version of PVposEst, subtracting VSPatchPosSum,
	// which represents the accumulated expectation of PVpos to this point.
	GvPVposEstDisc

	// GiveUpDiff is the difference: PVposEstDisc - PVneg representing the
	// expected positive outcome up to this point.  When this turns negative,
	// the chance of giving up goes up proportionally, as a logistic
	// function of this difference.
	GvGiveUpDiff

	// GiveUpProb is the probability from the logistic function of GiveUpDiff
	GvGiveUpProb

	// GiveUp is true if a reset was triggered probabilistically based on GiveUpProb
	GvGiveUp

	// GaveUp is copy of GiveUp from previous trial
	GvGaveUp

	// VSPatchPos is net shunting input from VSPatch (PosD1, named PVi in original PVLV)
	// computed as the Max of US-specific VSPatch saved values.
	// This is also stored as GvRewPred.
	GvVSPatchPos

	// VSPatchPosPrev is the previous-trial version of VSPatchPos -- for adjusting the
	// VSPatchThr threshold
	GvVSPatchPosPrev

	// VSPatchPosSum is the sum of VSPatchPos over goal engaged trials,
	// representing the integrated prediction that the US is going to occur
	GvVSPatchPosSum

	// computed LHb activity level that drives dipping / pausing of DA firing,
	// when VSPatch pos prediction > actual PV reward drive
	// or PVneg > PVpos
	GvLHbDip

	// LHbBurst is computed LHb activity level that drives bursts of DA firing, when actual PV reward drive > VSPatch pos prediction
	GvLHbBurst

	// LHbPVDA is GvLHbBurst - GvLHbDip -- the LHb contribution to DA, reflecting PV and VSPatch (PVi), but not the CS (LV) contributions
	GvLHbPVDA

	// CeMpos is positive valence central nucleus of the amygdala (CeM) LV (learned value) activity, reflecting |BLAPosAcqD1 - BLAPosExtD2|_+ positively rectified.  CeM sets Raw directly.  Note that a positive US onset even with no active Drive will be reflected here, enabling learning about unexpected outcomes
	GvCeMpos

	// CeMneg is negative valence central nucleus of the amygdala (CeM) LV (learned value) activity, reflecting |BLANegAcqD2 - BLANegExtD1|_+ positively rectified.  CeM sets Raw directly
	GvCeMneg

	// VtaDA is overall dopamine value reflecting all of the different inputs
	GvVtaDA

	// USneg are negative valence US outcomes -- normalized version of raw,
	// NNegUSs of them
	GvUSneg

	// USnegRaw are raw, linearly incremented negative valence US outcomes,
	// this value is also integrated together with all US vals for PVneg
	GvUSnegRaw

	// Drives is current drive state -- updated with optional homeostatic exponential return to baseline values
	GvDrives

	// USpos is current positive-valence drive-satisfying input(s) (unconditioned stimuli = US)
	GvUSpos

	// VSPatch is current reward predicting VSPatch (PosD1) values
	GvVSPatch

	// VSPatch is previous reward predicting VSPatch (PosD1) values
	GvVSPatchPrev

	// OFCposUSPTMaint is activity level of given OFCposUSPT maintenance pool
	// used in anticipating potential USpos outcome value
	GvOFCposUSPTMaint

	// VSMatrixPoolGated indicates whether given VSMatrix pool gated
	// this is reset after last goal accomplished -- records gating since then.
	GvVSMatrixPoolGated

	GlobalVarsN
)

func (GlobalVars) Desc added in v1.8.14 

func (i GlobalVars) Desc() string

func (*GlobalVars) FromString added in v1.8.0 

func (i *GlobalVars) FromString(s string) error

func (GlobalVars) MarshalJSON added in v1.8.0 

func (ev GlobalVars) MarshalJSON() ([]byte, error)

func (GlobalVars) String added in v1.8.0 

func (i GlobalVars) String() string

func (*GlobalVars) UnmarshalJSON added in v1.8.0 

func (ev *GlobalVars) UnmarshalJSON(b []byte) error

type HipConfig added in v1.8.6 

type HipConfig struct {

	// size of EC2
	EC2Size evec.Vec2i `nest:"+" desc:"size of EC2"`

	// number of EC3 pools (outer dimension)
	EC3NPool evec.Vec2i `nest:"+" desc:"number of EC3 pools (outer dimension)"`

	// number of neurons in one EC3 pool
	EC3NNrn evec.Vec2i `nest:"+" desc:"number of neurons in one EC3 pool"`

	// number of neurons in one CA1 pool
	CA1NNrn evec.Vec2i `nest:"+" desc:"number of neurons in one CA1 pool"`

	// size of CA3
	CA3Size evec.Vec2i `nest:"+" desc:"size of CA3"`

	// [def: 2.236] size of DG / CA3
	DGRatio float32 `def:"2.236" desc:"size of DG / CA3"`

	// [def: 0.1] percent connectivity from EC3 to EC2
	EC3ToEC2PCon float32 `def:"0.1" desc:"percent connectivity from EC3 to EC2"`

	// [def: 0.25] percent connectivity from EC2 to DG
	EC2ToDGPCon float32 `def:"0.25" desc:"percent connectivity from EC2 to DG"`

	// [def: 0.25] percent connectivity from EC2 to CA3
	EC2ToCA3PCon float32 `def:"0.25" desc:"percent connectivity from EC2 to CA3"`

	// [def: 0.25] percent connectivity from CA3 to CA1
	CA3ToCA1PCon float32 `def:"0.25" desc:"percent connectivity from CA3 to CA1"`

	// [def: 0.02] percent connectivity into CA3 from DG
	DGToCA3PCon float32 `def:"0.02" desc:"percent connectivity into CA3 from DG"`

	// lateral radius of connectivity in EC2
	EC2LatRadius int `desc:"lateral radius of connectivity in EC2"`

	// lateral gaussian sigma in EC2 for how quickly weights fall off with distance
	EC2LatSigma float32 `desc:"lateral gaussian sigma in EC2 for how quickly weights fall off with distance"`

	// [def: 1] proportion of full mossy fiber strength (PrjnScale.Rel) for CA3 EDL in training, applied at the start of a trial to reduce DG -> CA3 strength.  1 = fully reduce strength, .5 = 50% reduction, etc
	MossyDelta float32 `` /* 209-byte string literal not displayed */

	// [def: 0.75] proportion of full mossy fiber strength (PrjnScale.Rel) for CA3 EDL in testing, applied during 2nd-3rd quarters to reduce DG -> CA3 strength.  1 = fully reduce strength, .5 = 50% reduction, etc
	MossyDeltaTest float32 `` /* 211-byte string literal not displayed */

	// [def: 0.9] low theta modulation value for temporal difference EDL -- sets PrjnScale.Rel on CA1 <-> EC prjns consistent with Theta phase model
	ThetaLow float32 `` /* 147-byte string literal not displayed */

	// [def: 1] high theta modulation value for temporal difference EDL -- sets PrjnScale.Rel on CA1 <-> EC prjns consistent with Theta phase model
	ThetaHigh float32 `` /* 146-byte string literal not displayed */

	// [def: true] flag for clamping the EC5 from EC5ClampSrc
	EC5Clamp bool `def:"true" desc:"flag for clamping the EC5 from EC5ClampSrc"`

	// [def: EC3] source layer for EC5 clamping activations in the plus phase -- biologically it is EC3 but can use an Input layer if available
	EC5ClampSrc string `` /* 142-byte string literal not displayed */

	// [def: true] clamp the EC5 from EC5ClampSrc during testing as well as training -- this will overwrite any target values that might be used in stats (e.g., in the basic hip example), so it must be turned off there
	EC5ClampTest bool `` /* 217-byte string literal not displayed */

	// [def: 0.1] threshold for binarizing EC5 clamp values -- any value above this is clamped to 1, else 0 -- helps produce a cleaner learning signal.  Set to 0 to not perform any binarization.
	EC5ClampThr float32 `` /* 193-byte string literal not displayed */
}

HipConfig have the hippocampus size and connectivity parameters

func (*HipConfig) Defaults added in v1.8.6 

func (hip *HipConfig) Defaults()

type HipPrjnParams added in v1.8.5 

type HipPrjnParams struct {

	// [def: 0] Hebbian learning proportion
	Hebb float32 `def:"0" desc:"Hebbian learning proportion"`

	// [def: 1] EDL proportion
	Err float32 `def:"1" desc:"EDL proportion"`

	// [def: 0.4:0.8] [min: 0] [max: 1] proportion of correction to apply to sending average activation for hebbian learning component (0=none, 1=all, .5=half, etc)
	SAvgCor float32 `` /* 161-byte string literal not displayed */

	// [def: 0.01] [min: 0] threshold of sending average activation below which learning does not occur (prevents learning when there is no input)
	SAvgThr float32 `` /* 144-byte string literal not displayed */

	// [def: 0.1] [min: 0] sending layer Nominal (need to manually set it to be the same as the sending layer)
	SNominal float32 `def:"0.1" min:"0" desc:"sending layer Nominal (need to manually set it to be the same as the sending layer)"`
	// contains filtered or unexported fields
}

HipPrjnParams define behavior of hippocampus prjns, which have special learning rules

func (*HipPrjnParams) Defaults added in v1.8.5 

func (hp *HipPrjnParams) Defaults()

func (*HipPrjnParams) Update added in v1.8.5 

func (hp *HipPrjnParams) Update()

type InhibParams 

type InhibParams struct {

	// [view: inline] layer-level and pool-level average activation initial values and updating / adaptation thereof -- initial values help determine initial scaling factors.
	ActAvg ActAvgParams `` /* 173-byte string literal not displayed */

	// [view: inline] inhibition across the entire layer -- inputs generally use Gi = 0.8 or 0.9, 1.3 or higher for sparse layers.  If the layer has sub-pools (4D shape) then this is effectively between-pool inhibition.
	Layer fsfffb.GiParams `` /* 218-byte string literal not displayed */

	// [view: inline] inhibition within sub-pools of units, for layers with 4D shape -- almost always need this if the layer has pools.
	Pool fsfffb.GiParams `` /* 134-byte string literal not displayed */
}

axon.InhibParams contains all the inhibition computation params and functions for basic Axon This is included in axon.Layer to support computation. This also includes other misc layer-level params such as expected average activation in the layer which is used for Ge rescaling and potentially for adapting inhibition over time

func (*InhibParams) Defaults 

func (ip *InhibParams) Defaults()

func (*InhibParams) Update 

func (ip *InhibParams) Update()

type LDTParams added in v1.7.18 

type LDTParams struct {

	// [def: 0.05] threshold per input source, on absolute value (magnitude), to count as a significant reward event, which then drives maximal ACh -- set to 0 to disable this nonlinear behavior
	SrcThr float32 `` /* 193-byte string literal not displayed */

	// [def: true] use the global Context.NeuroMod.HasRew flag -- if there is some kind of external reward being given, then ACh goes to 1, else 0 for this component
	Rew slbool.Bool `` /* 164-byte string literal not displayed */

	// [def: 2] extent to which active maintenance (via Context.NeuroMod.NotMaint PTNotMaintLayer activity) inhibits ACh signals -- when goal engaged, distractability is lower.
	MaintInhib float32 `` /* 175-byte string literal not displayed */

	// [def: 0.4] maximum NeuroMod.NotMaint activity for computing Maint as 1-NotMaint -- when NotMaint is >= NotMaintMax, then Maint = 0.
	NotMaintMax float32 `` /* 137-byte string literal not displayed */

	// idx of Layer to get max activity from -- set during Build from BuildConfig SrcLay1Name if present -- -1 if not used
	SrcLay1Idx int32 `` /* 135-byte string literal not displayed */

	// idx of Layer to get max activity from -- set during Build from BuildConfig SrcLay2Name if present -- -1 if not used
	SrcLay2Idx int32 `` /* 135-byte string literal not displayed */

	// idx of Layer to get max activity from -- set during Build from BuildConfig SrcLay3Name if present -- -1 if not used
	SrcLay3Idx int32 `` /* 135-byte string literal not displayed */

	// idx of Layer to get max activity from -- set during Build from BuildConfig SrcLay4Name if present -- -1 if not used
	SrcLay4Idx int32 `` /* 135-byte string literal not displayed */
}

LDTParams compute reward salience as ACh global neuromodulatory signal as a function of the MAX activation of its inputs.

func (*LDTParams) ACh added in v1.7.19 

func (lp *LDTParams) ACh(ctx *Context, di uint32, srcLay1Act, srcLay2Act, srcLay3Act, srcLay4Act float32) float32

ACh returns the computed ACh salience value based on given source layer activations and key values from the ctx Context.

func (*LDTParams) Defaults added in v1.7.18 

func (lp *LDTParams) Defaults()

func (*LDTParams) MaintFmNotMaint added in v1.7.18 

func (lp *LDTParams) MaintFmNotMaint(notMaint float32) float32

MaintFmNotMaint returns a 0-1 value reflecting strength of active maintenance based on the activity of the PTNotMaintLayer as recorded in NeuroMod.NotMaint.

func (*LDTParams) MaxSrcAct added in v1.7.19 

func (lp *LDTParams) MaxSrcAct(maxSrcAct, srcLayAct float32) float32

MaxSrcAct returns the updated maxSrcAct value from given source layer activity value.

func (*LDTParams) Thr added in v1.7.18 

func (lp *LDTParams) Thr(val float32) float32

Thr applies SrcThr threshold to given value

func (*LDTParams) Update added in v1.7.18 

func (lp *LDTParams) Update()

type LHbParams added in v1.8.16 

type LHbParams struct {

	// [def: 1] threshold factor that multiplies integrated pvNeg value to establish a threshold for whether the integrated pvPos value is good enough to drive overall net positive reward
	NegThr float32 `` /* 186-byte string literal not displayed */

	// [def: 1] gain multiplier on PVpos for purposes of generating bursts (not for  discounting negative dips) -- 4 renormalizes for typical ~.5 values (.5 * .5 = .25)
	BurstGain float32 `` /* 167-byte string literal not displayed */

	// [def: 1] gain multiplier on PVneg for purposes of generating dips (not for  discounting positive bursts) -- 4 renormalizes for typical ~.5 values (.5 * .5 = .25)
	DipGain float32 `` /* 167-byte string literal not displayed */
}

LHbParams has values for computing LHb & RMTg which drives dips / pauses in DA firing. LHb handles all US-related (PV = primary value) processing. Positive net LHb activity drives dips / pauses in VTA DA activity, e.g., when predicted pos > actual or actual neg > predicted. Negative net LHb activity drives bursts in VTA DA activity, e.g., when actual pos > predicted (redundant with LV / Amygdala) or "relief" burst when actual neg < predicted.

func (*LHbParams) DAFmPVs added in v1.8.17 

func (lh *LHbParams) DAFmPVs(pvPos, pvNeg, vsPatchPos float32) (burst, dip, da, rew float32)

DAFmPVs computes the overall PV DA in terms of LHb burst and dip activity from given pvPos, pvNeg, and vsPatchPos values. Also returns the net "reward" value as the discounted PV value, separate from the vsPatchPos prediction error factor.

func (*LHbParams) DAforNoUS added in v1.8.16 

func (lh *LHbParams) DAforNoUS(ctx *Context, di uint32, vsPatchPos float32) float32

DAforNoUS computes the overall LHb Dip = vsPatchPos, and PVDA ~= -Dip, for case when there is _NOT_ a primary positive reward value or a give-up state. In this case, inhibition of VS via ACh is assumed to prevent activity of PVneg (and there is no PVpos), so only vsPatchPos is operative. Returns net dopamine which is -vsPatchPos.

func (*LHbParams) DAforUS added in v1.8.16 

func (lh *LHbParams) DAforUS(ctx *Context, di uint32, pvPos, pvNeg, vsPatchPos float32) float32

DAforUS computes the overall LHb Dip or Burst (one is always 0), and PVDA ~= Burst - Dip, for case when there is a primary positive reward value or a give-up state has triggered. Returns the overall net reward magnitude, prior to VSPatch discounting.

func (*LHbParams) Defaults added in v1.8.16 

func (lh *LHbParams) Defaults()

func (*LHbParams) Reset added in v1.8.16 

func (lh *LHbParams) Reset(ctx *Context, di uint32)

Reset resets all LHb vars back to 0

func (*LHbParams) Update added in v1.8.16 

func (lh *LHbParams) Update()

type LRateMod added in v1.6.13 

type LRateMod struct {

	// toggle use of this modulation factor
	On slbool.Bool `desc:"toggle use of this modulation factor"`

	// [viewif: On] [min: 0] [max: 1] baseline learning rate -- what you get for correct cases
	Base float32 `viewif:"On" min:"0" max:"1" desc:"baseline learning rate -- what you get for correct cases"`

	// [viewif: On] defines the range over which modulation occurs for the modulator factor -- Min and below get the Base level of learning rate modulation, Max and above get a modulation of 1
	Range minmax.F32 `` /* 191-byte string literal not displayed */
	// contains filtered or unexported fields
}

LRateMod implements global learning rate modulation, based on a performance-based factor, for example error. Increasing levels of the factor = higher learning rate. This can be added to a Sim and called prior to DWt() to dynamically change lrate based on overall network performance.

func (*LRateMod) Defaults added in v1.6.13 

func (lr *LRateMod) Defaults()

func (*LRateMod) LRateMod added in v1.6.13 

func (lr *LRateMod) LRateMod(net *Network, fact float32) float32

LRateMod calls LRateMod on given network, using computed Mod factor based on given normalized modulation factor (0 = no error = Base learning rate, 1 = maximum error). returns modulation factor applied.

func (*LRateMod) Mod added in v1.6.13 

func (lr *LRateMod) Mod(fact float32) float32

Mod returns the learning rate modulation factor as a function of any kind of normalized modulation factor, e.g., an error measure. If fact <= Range.Min, returns Base If fact >= Range.Max, returns 1 otherwise, returns proportional value between Base..1

func (*LRateMod) Update added in v1.6.13 

func (lr *LRateMod) Update()

type LRateParams added in v1.6.13 

type LRateParams struct {

	// [def: 0.04,0.1,0.2] base learning rate for this projection -- can be modulated by other factors below -- for larger networks, use slower rates such as 0.04, smaller networks can use faster 0.2.
	Base float32 `` /* 199-byte string literal not displayed */

	// scheduled learning rate multiplier, simulating reduction in plasticity over aging
	Sched float32 `desc:"scheduled learning rate multiplier, simulating reduction in plasticity over aging"`

	// dynamic learning rate modulation due to neuromodulatory or other such factors
	Mod float32 `desc:"dynamic learning rate modulation due to neuromodulatory or other such factors"`

	// effective actual learning rate multiplier used in computing DWt: Eff = eMod * Sched * Base
	Eff float32 `inactive:"+" desc:"effective actual learning rate multiplier used in computing DWt: Eff = eMod * Sched * Base"`
}

LRateParams manages learning rate parameters

func (*LRateParams) Defaults added in v1.6.13 

func (ls *LRateParams) Defaults()

func (*LRateParams) Init added in v1.6.13 

func (ls *LRateParams) Init()

Init initializes modulation values back to 1 and updates Eff

func (*LRateParams) Update added in v1.6.13 

func (ls *LRateParams) Update()

func (*LRateParams) UpdateEff added in v1.7.0 

func (ls *LRateParams) UpdateEff()

type LaySpecialVals added in v1.7.0 

type LaySpecialVals struct {

	// one value
	V1 float32 `inactive:"+" desc:"one value"`

	// one value
	V2 float32 `inactive:"+" desc:"one value"`

	// one value
	V3 float32 `inactive:"+" desc:"one value"`

	// one value
	V4 float32 `inactive:"+" desc:"one value"`
}

LaySpecialVals holds special values used to communicate to other layers based on neural values, used for special algorithms such as RL where some of the computation is done algorithmically.

func (*LaySpecialVals) Init added in v1.7.9 

func (lv *LaySpecialVals) Init()

type Layer 

type Layer struct {
	LayerBase

	// all layer-level parameters -- these must remain constant once configured
	Params *LayerParams `desc:"all layer-level parameters -- these must remain constant once configured"`
}

axon.Layer implements the basic Axon spiking activation function, and manages learning in the projections.

func (*Layer) AdaptInhib added in v1.2.37 

func (ly *Layer) AdaptInhib(ctx *Context)

AdaptInhib adapts inhibition

func (*Layer) AllParams 

func (ly *Layer) AllParams() string

AllParams returns a listing of all parameters in the Layer

func (*Layer) AnyGated added in v1.7.0 

func (ly *Layer) AnyGated(di uint32) bool

AnyGated returns true if the layer-level pool Gated flag is true, which indicates if any of the layers gated.

func (*Layer) ApplyExt 

func (ly *Layer) ApplyExt(ctx *Context, di uint32, ext etensor.Tensor)

ApplyExt applies external input in the form of an etensor.Float32 or 64. Negative values and NaNs are not valid, and will be interpreted as missing inputs. The given data index di is the data parallel index (0 < di < MaxData): must present inputs separately for each separate data parallel set. If dimensionality of tensor matches that of layer, and is 2D or 4D, then each dimension is iterated separately, so any mismatch preserves dimensional structure. Otherwise, the flat 1D view of the tensor is used. If the layer is a Target or Compare layer type, then it goes in Target otherwise it goes in Ext. Also sets the Exts values on layer, which are used for the GPU version, which requires calling the network ApplyExts() method -- is a no-op for CPU.

func (*Layer) ApplyExt1D 

func (ly *Layer) ApplyExt1D(ctx *Context, di uint32, ext []float64)

ApplyExt1D applies external input in the form of a flat 1-dimensional slice of floats If the layer is a Target or Compare layer type, then it goes in Target otherwise it goes in Ext

func (*Layer) ApplyExt1D32 

func (ly *Layer) ApplyExt1D32(ctx *Context, di uint32, ext []float32)

ApplyExt1D32 applies external input in the form of a flat 1-dimensional slice of float32s. If the layer is a Target or Compare layer type, then it goes in Target otherwise it goes in Ext

func (*Layer) ApplyExt1DTsr 

func (ly *Layer) ApplyExt1DTsr(ctx *Context, di uint32, ext etensor.Tensor)

ApplyExt1DTsr applies external input using 1D flat interface into tensor. If the layer is a Target or Compare layer type, then it goes in Target otherwise it goes in Ext

func (*Layer) ApplyExt2D 

func (ly *Layer) ApplyExt2D(ctx *Context, di uint32, ext etensor.Tensor)

ApplyExt2D applies 2D tensor external input

func (*Layer) ApplyExt2Dto4D 

func (ly *Layer) ApplyExt2Dto4D(ctx *Context, di uint32, ext etensor.Tensor)

ApplyExt2Dto4D applies 2D tensor external input to a 4D layer

func (*Layer) ApplyExt4D 

func (ly *Layer) ApplyExt4D(ctx *Context, di uint32, ext etensor.Tensor)

ApplyExt4D applies 4D tensor external input

func (*Layer) ApplyExtFlags 

func (ly *Layer) ApplyExtFlags() (clearMask, setMask NeuronFlags, toTarg bool)

ApplyExtFlags gets the clear mask and set mask for updating neuron flags based on layer type, and whether input should be applied to Target (else Ext)

func (*Layer) ApplyExtVal added in v1.7.9 

func (ly *Layer) ApplyExtVal(ctx *Context, lni, di uint32, val float32, clearMask, setMask NeuronFlags, toTarg bool)

ApplyExtVal applies given external value to given neuron using clearMask, setMask, and toTarg from ApplyExtFlags. Also saves Val in Exts for potential use by GPU.

func (*Layer) AsAxon 

func (ly *Layer) AsAxon() *Layer

AsAxon returns this layer as a axon.Layer -- all derived layers must redefine this to return the base Layer type, so that the AxonLayer interface does not need to include accessors to all the basic stuff

func (*Layer) AvgDifFmTrgAvg added in v1.6.0 

func (ly *Layer) AvgDifFmTrgAvg(ctx *Context)

AvgDifFmTrgAvg updates neuron-level AvgDif values from AvgPct - TrgAvg which is then used for synaptic scaling of LWt values in Prjn SynScale.

func (*Layer) AvgMaxVarByPool added in v1.6.0 

func (ly *Layer) AvgMaxVarByPool(ctx *Context, varNm string, poolIdx, di int) minmax.AvgMax32

AvgMaxVarByPool returns the average and maximum value of given variable for given pool index (0 = entire layer, 1.. are subpools for 4D only). Uses fast index-based variable access.

func (*Layer) BGThalDefaults added in v1.7.18 

func (ly *Layer) BGThalDefaults()

func (*Layer) BLADefaults added in v1.7.11 

func (ly *Layer) BLADefaults()

func (*Layer) BetweenLayerGi added in v1.7.5 

func (ly *Layer) BetweenLayerGi(ctx *Context)

BetweenLayerGi computes inhibition Gi between layers

func (*Layer) BetweenLayerGiMax added in v1.7.5 

func (ly *Layer) BetweenLayerGiMax(net *Network, di uint32, maxGi float32, layIdx int32) float32

BetweenLayerGiMax returns max gi value for input maxGi vs the given layIdx layer

func (*Layer) CTDefParamsFast added in v1.7.18 

func (ly *Layer) CTDefParamsFast()

CTDefParamsFast sets fast time-integration parameters for CTLayer. This is what works best in the deep_move 1 trial history case, vs Medium and Long

func (*Layer) CTDefParamsLong added in v1.7.18 

func (ly *Layer) CTDefParamsLong()

CTDefParamsLong sets long time-integration parameters for CTLayer. This is what works best in the deep_music test case integrating over long time windows, compared to Medium and Fast.

func (*Layer) CTDefParamsMedium added in v1.7.18 

func (ly *Layer) CTDefParamsMedium()

CTDefParamsMedium sets medium time-integration parameters for CTLayer. This is what works best in the FSA test case, compared to Fast (deep_move) and Long (deep_music) time integration.

func (*Layer) CeMDefaults added in v1.7.11 

func (ly *Layer) CeMDefaults()

func (*Layer) ClearTargExt added in v1.2.65 

func (ly *Layer) ClearTargExt(ctx *Context)

ClearTargExt clears external inputs Ext that were set from target values Target. This can be called to simulate alpha cycles within theta cycles, for example.

func (*Layer) CorSimFmActs added in v1.3.35 

func (ly *Layer) CorSimFmActs(ctx *Context)

CorSimFmActs computes the correlation similarity (centered cosine aka normalized dot product) in activation state between minus and plus phases.

func (*Layer) CostEst 

func (ly *Layer) CostEst() (neur, syn, tot int)

CostEst returns the estimated computational cost associated with this layer, separated by neuron-level and synapse-level, in arbitrary units where cost per synapse is 1. Neuron-level computation is more expensive but there are typically many fewer neurons, so in larger networks, synaptic costs tend to dominate. Neuron cost is estimated from TimerReport output for large networks.

func (*Layer) CycleNeuron added in v1.6.0 

func (ly *Layer) CycleNeuron(ctx *Context, ni uint32)

CycleNeuron does one cycle (msec) of updating at the neuron level Called directly by Network, iterates over data.

func (*Layer) CyclePost 

func (ly *Layer) CyclePost(ctx *Context)

CyclePost is called after the standard Cycle update, as a separate network layer loop. This is reserved for any kind of special ad-hoc types that need to do something special after Spiking is finally computed and Sent. Typically used for updating global values in the Context state, such as updating a neuromodulatory signal such as dopamine. Any updates here must also be done in gpu_hlsl/gpu_cyclepost.hlsl

func (*Layer) DTrgSubMean added in v1.6.0 

func (ly *Layer) DTrgSubMean(ctx *Context)

DTrgSubMean subtracts the mean from DTrgAvg values Called by TrgAvgFmD

func (*Layer) DWt 

func (ly *Layer) DWt(ctx *Context, si uint32)

DWt computes the weight change (learning), based on synaptically-integrated spiking, computed at the Theta cycle interval. This is the trace version for hidden units, and uses syn CaP - CaD for targets.

func (*Layer) DWtSubMean added in v1.2.23 

func (ly *Layer) DWtSubMean(ctx *Context, ri uint32)

DWtSubMean computes subtractive normalization of the DWts

func (*Layer) DecayState 

func (ly *Layer) DecayState(ctx *Context, di uint32, decay, glong, ahp float32)

DecayState decays activation state by given proportion (default decay values are ly.Params.Acts.Decay.Act, Glong)

func (*Layer) DecayStateLayer added in v1.7.0 

func (ly *Layer) DecayStateLayer(ctx *Context, di uint32, decay, glong, ahp float32)

DecayStateLayer does layer-level decay, but not neuron level

func (*Layer) DecayStateNeuronsAll added in v1.8.0 

func (ly *Layer) DecayStateNeuronsAll(ctx *Context, decay, glong, ahp float32)

DecayStateNeuronsAll decays neural activation state by given proportion (default decay values are ly.Params.Acts.Decay.Act, Glong, AHP) for all data parallel indexes. Does not decay pool or layer state. This is used for minus phase of Pulvinar layers to clear state in prep for driver plus phase.

func (*Layer) DecayStatePool 

func (ly *Layer) DecayStatePool(ctx *Context, pool int, decay, glong, ahp float32)

DecayStatePool decays activation state by given proportion in given sub-pool index (0 based)

func (*Layer) Defaults 

func (ly *Layer) Defaults()

func (*Layer) GInteg added in v1.5.12 

func (ly *Layer) GInteg(ctx *Context, ni, di uint32, pl *Pool, vals *LayerVals)

GInteg integrates conductances G over time (Ge, NMDA, etc). calls SpecialGFmRawSyn, GiInteg

func (*Layer) GPDefaults added in v1.7.0 

func (ly *Layer) GPDefaults()

func (*Layer) GPPostBuild added in v1.7.0 

func (ly *Layer) GPPostBuild()

func (*Layer) GPiDefaults added in v1.7.0 

func (ly *Layer) GPiDefaults()

func (*Layer) GatedFmSpkMax added in v1.7.0 

func (ly *Layer) GatedFmSpkMax(di uint32, thr float32) (bool, int)

GatedFmSpkMax updates the Gated state in Pools of given layer, based on Avg SpkMax being above given threshold. returns true if any gated, and the pool index if 4D layer (0 = first).

func (*Layer) GatherSpikes added in v1.7.2 

func (ly *Layer) GatherSpikes(ctx *Context, ni uint32)

GatherSpikes integrates G*Raw and G*Syn values for given recv neuron while integrating the Recv Prjn-level GSyn integrated values.

func (*Layer) GiFmSpikes added in v1.5.12 

func (ly *Layer) GiFmSpikes(ctx *Context)

GiFmSpikes gets the Spike, GeRaw and GeExt from neurons in the pools where Spike drives FBsRaw = raw feedback signal, GeRaw drives FFsRaw = aggregate feedforward excitatory spiking input. GeExt represents extra excitatory input from other sources. Then integrates new inhibitory conductances therefrom, at the layer and pool level. Called separately by Network.CycleImpl on all Layers Also updates all AvgMax values at the Cycle level.

func (*Layer) HasPoolInhib added in v1.2.79 

func (ly *Layer) HasPoolInhib() bool

HasPoolInhib returns true if the layer is using pool-level inhibition (implies 4D too). This is the proper check for using pool-level target average activations, for example.

func (*Layer) InitActAvg 

func (ly *Layer) InitActAvg(ctx *Context)

InitActAvg initializes the running-average activation values that drive learning and the longer time averaging values.

func (*Layer) InitActAvgLayer added in v1.8.0 

func (ly *Layer) InitActAvgLayer(ctx *Context)

InitActAvgLayer initializes the running-average activation values that drive learning and the longer time averaging values. version with just overall layer-level inhibition.

func (*Layer) InitActAvgPools added in v1.8.0 

func (ly *Layer) InitActAvgPools(ctx *Context)

InitActAvgPools initializes the running-average activation values that drive learning and the longer time averaging values. version with pooled inhibition.

func (*Layer) InitActs 

func (ly *Layer) InitActs(ctx *Context)

InitActs fully initializes activation state -- only called automatically during InitWts

func (*Layer) InitExt 

func (ly *Layer) InitExt(ctx *Context)

InitExt initializes external input state. Should be called prior to ApplyExt on all layers receiving Ext input.

func (*Layer) InitGScale added in v1.2.37 

func (ly *Layer) InitGScale(ctx *Context)

InitGScale computes the initial scaling factor for synaptic input conductances G, stored in GScale.Scale, based on sending layer initial activation.

func (*Layer) InitPrjnGBuffs added in v1.5.12 

func (ly *Layer) InitPrjnGBuffs(ctx *Context)

InitPrjnGBuffs initializes the projection-level conductance buffers and conductance integration values for receiving projections in this layer.

func (*Layer) InitWtSym 

func (ly *Layer) InitWtSym(ctx *Context)

InitWtsSym initializes the weight symmetry -- higher layers copy weights from lower layers

func (*Layer) InitWts 

func (ly *Layer) InitWts(ctx *Context, nt *Network)

InitWts initializes the weight values in the network, i.e., resetting learning Also calls InitActs

func (*Layer) LDTDefaults added in v1.7.18 

func (ly *Layer) LDTDefaults()

func (*Layer) LDTPostBuild added in v1.7.18 

func (ly *Layer) LDTPostBuild()

func (*Layer) LDTSrcLayAct added in v1.7.19 

func (ly *Layer) LDTSrcLayAct(net *Network, layIdx int32, di uint32) float32

LDTSrcLayAct returns the overall activity level for given source layer for purposes of computing ACh salience value. Typically the input is a superior colliculus (SC) layer that rapidly accommodates after the onset of a stimulus. using lpl.AvgMax.CaSpkP.Cycle.Max for layer activity measure.

func (*Layer) LRateMod added in v1.6.13 

func (ly *Layer) LRateMod(mod float32)

LRateMod sets the LRate modulation parameter for Prjns, which is for dynamic modulation of learning rate (see also LRateSched). Updates the effective learning rate factor accordingly.

func (*Layer) LRateSched added in v1.6.13 

func (ly *Layer) LRateSched(sched float32)

LRateSched sets the schedule-based learning rate multiplier. See also LRateMod. Updates the effective learning rate factor accordingly.

func (*Layer) LesionNeurons 

func (ly *Layer) LesionNeurons(prop float32) int

LesionNeurons lesions (sets the Off flag) for given proportion (0-1) of neurons in layer returns number of neurons lesioned. Emits error if prop > 1 as indication that percent might have been passed

func (*Layer) LocalistErr2D added in v1.5.3 

func (ly *Layer) LocalistErr2D(ctx *Context) (err []bool, minusIdx, plusIdx []int)

LocalistErr2D decodes a 2D layer with Y axis = redundant units, X = localist units returning the indexes of the max activated localist value in the minus and plus phase activities, and whether these are the same or different (err = different) returns one result per data parallel index ([ctx.NetIdxs.NData])

func (*Layer) LocalistErr4D added in v1.5.3 

func (ly *Layer) LocalistErr4D(ctx *Context) (err []bool, minusIdx, plusIdx []int)

LocalistErr4D decodes a 4D layer with each pool representing a localist value. Returns the flat 1D indexes of the max activated localist value in the minus and plus phase activities, and whether these are the same or different (err = different)

func (*Layer) MatrixDefaults added in v1.7.0 

func (ly *Layer) MatrixDefaults()

func (*Layer) MatrixGated added in v1.7.0 

func (ly *Layer) MatrixGated(ctx *Context)

MatrixGated is called after std PlusPhase, on CPU, has Pool info downloaded from GPU, to set Gated flag based on SpkMax activity

func (*Layer) MatrixPostBuild added in v1.7.0 

func (ly *Layer) MatrixPostBuild()

func (*Layer) MinusPhase added in v1.2.63 

func (ly *Layer) MinusPhase(ctx *Context)

MinusPhase does updating at end of the minus phase

func (*Layer) MinusPhasePost added in v1.7.10 

func (ly *Layer) MinusPhasePost(ctx *Context)

MinusPhasePost does special algorithm processing at end of minus

func (*Layer) NewState added in v1.2.63 

func (ly *Layer) NewState(ctx *Context)

NewState handles all initialization at start of new input pattern. Does NOT call InitGScale()

func (*Layer) NewStateNeurons added in v1.8.0 

func (ly *Layer) NewStateNeurons(ctx *Context)

NewStateNeurons only calls the neurons part of new state -- for misbehaving GPU

func (*Layer) Object added in v1.7.0 

func (ly *Layer) Object() any

Object returns the object with parameters to be set by emer.Params

func (*Layer) PTMaintDefaults added in v1.7.2 

func (ly *Layer) PTMaintDefaults()

func (*Layer) PTNotMaintDefaults added in v1.7.11 

func (ly *Layer) PTNotMaintDefaults()

func (*Layer) PVLVPostBuild added in v1.7.11 

func (ly *Layer) PVLVPostBuild()

PVLVPostBuild is used for BLA, VSPatch, and PVLayer types to set NeuroMod params

func (*Layer) PctUnitErr 

func (ly *Layer) PctUnitErr(ctx *Context) []float64

PctUnitErr returns the proportion of units where the thresholded value of Target (Target or Compare types) or ActP does not match that of ActM. If Act > ly.Params.Acts.Clamp.ErrThr, effective activity = 1 else 0 robust to noisy activations. returns one result per data parallel index ([ctx.NetIdxs.NData])

func (*Layer) PlusPhase added in v1.2.63 

func (ly *Layer) PlusPhase(ctx *Context)

PlusPhase does updating at end of the plus phase

func (*Layer) PlusPhaseActAvg added in v1.8.0 

func (ly *Layer) PlusPhaseActAvg(ctx *Context)

PlusPhaseActAvg updates ActAvg and DTrgAvg at the plus phase Note: could be done on GPU but not worth it at this point..

func (*Layer) PlusPhasePost added in v1.7.0 

func (ly *Layer) PlusPhasePost(ctx *Context)

PlusPhasePost does special algorithm processing at end of plus

func (*Layer) PlusPhaseStart added in v1.7.10 

func (ly *Layer) PlusPhaseStart(ctx *Context)

PlusPhaseStart does updating at the start of the plus phase: applies Target inputs as External inputs.

func (*Layer) PoolGiFmSpikes added in v1.5.12 

func (ly *Layer) PoolGiFmSpikes(ctx *Context)

PoolGiFmSpikes computes inhibition Gi from Spikes within sub-pools. and also between different layers based on LayInhib* indexes must happen after LayPoolGiFmSpikes has been called.

func (*Layer) PostBuild added in v1.7.0 

func (ly *Layer) PostBuild()

PostBuild performs special post-Build() configuration steps for specific algorithms, using configuration data set in BuildConfig during the ConfigNet process.

func (*Layer) PostSpike added in v1.7.0 

func (ly *Layer) PostSpike(ctx *Context, ni uint32)

PostSpike does updates at neuron level after spiking has been computed. This is where special layer types add extra code. It also updates the CaSpkPCyc stats. Called directly by Network, iterates over data.

func (*Layer) PulvPostBuild added in v1.7.0 

func (ly *Layer) PulvPostBuild()

PulvPostBuild does post-Build config of Pulvinar based on BuildConfig options

func (*Layer) PulvinarDriver added in v1.7.0 

func (ly *Layer) PulvinarDriver(ctx *Context, lni, di uint32) (drvGe, nonDrivePct float32)

func (*Layer) RWDaPostBuild added in v1.7.0 

func (ly *Layer) RWDaPostBuild()

RWDaPostBuild does post-Build config

func (*Layer) ReadWtsJSON 

func (ly *Layer) ReadWtsJSON(r io.Reader) error

ReadWtsJSON reads the weights from this layer from the receiver-side perspective in a JSON text format. This is for a set of weights that were saved *for one layer only* and is not used for the network-level ReadWtsJSON, which reads into a separate structure -- see SetWts method.

func (*Layer) STNDefaults added in v1.7.0 

func (ly *Layer) STNDefaults()

func (*Layer) SendSpike 

func (ly *Layer) SendSpike(ctx *Context, ni uint32)

SendSpike sends spike to receivers for all neurons that spiked last step in Cycle, integrated the next time around. Called directly by Network, iterates over data.

func (*Layer) SetSubMean added in v1.6.11 

func (ly *Layer) SetSubMean(trgAvg, prjn float32)

SetSubMean sets the SubMean parameters in all the layers in the network trgAvg is for Learn.TrgAvgAct.SubMean prjn is for the prjns Learn.Trace.SubMean in both cases, it is generally best to have both parameters set to 0 at the start of learning

func (*Layer) SetWts 

func (ly *Layer) SetWts(lw *weights.Layer) error

SetWts sets the weights for this layer from weights.Layer decoded values

func (*Layer) SlowAdapt added in v1.2.37 

func (ly *Layer) SlowAdapt(ctx *Context)

SlowAdapt is the layer-level slow adaptation functions. Calls AdaptInhib and AvgDifFmTrgAvg for Synaptic Scaling. Does NOT call projection-level methods.

func (*Layer) SpikeFmG added in v1.6.0 

func (ly *Layer) SpikeFmG(ctx *Context, ni, di uint32, lpl *Pool)

SpikeFmG computes Vm from Ge, Gi, Gl conductances and then Spike from that

func (*Layer) SpkSt1 added in v1.5.10 

func (ly *Layer) SpkSt1(ctx *Context)

SpkSt1 saves current activation state in SpkSt1 variables (using CaP)

func (*Layer) SpkSt2 added in v1.5.10 

func (ly *Layer) SpkSt2(ctx *Context)

SpkSt2 saves current activation state in SpkSt2 variables (using CaP)

func (*Layer) SynCa added in v1.3.1 

func (ly *Layer) SynCa(ctx *Context, ni uint32)

SynCa updates synaptic calcium based on spiking, for SynSpkTheta mode. Optimized version only updates at point of spiking, threaded over neurons. Called directly by Network, iterates over data.

func (*Layer) SynFail added in v1.2.92 

func (ly *Layer) SynFail(ctx *Context)

SynFail updates synaptic weight failure only -- normally done as part of DWt and WtFmDWt, but this call can be used during testing to update failing synapses.

func (*Layer) TDDaPostBuild added in v1.7.0 

func (ly *Layer) TDDaPostBuild()

TDDaPostBuild does post-Build config

func (*Layer) TDIntegPostBuild added in v1.7.0 

func (ly *Layer) TDIntegPostBuild()

TDIntegPostBuild does post-Build config

func (*Layer) TargToExt added in v1.2.65 

func (ly *Layer) TargToExt(ctx *Context)

TargToExt sets external input Ext from target values Target This is done at end of MinusPhase to allow targets to drive activity in plus phase. This can be called separately to simulate alpha cycles within theta cycles, for example.

func (*Layer) TestVals added in v1.8.0 

func (ly *Layer) TestVals(ctrKey string, vals map[string]float32)

TestVals returns a map of key vals for testing ctrKey is a key of counters to contextualize values.

func (*Layer) TrgAvgFmD added in v1.2.32 

func (ly *Layer) TrgAvgFmD(ctx *Context)

TrgAvgFmD updates TrgAvg from DTrgAvg -- called in PlusPhasePost

func (*Layer) UnLesionNeurons 

func (ly *Layer) UnLesionNeurons()

UnLesionNeurons unlesions (clears the Off flag) for all neurons in the layer

func (*Layer) Update added in v1.7.0 

func (ly *Layer) Update()

Update is an interface for generically updating after edits this should be used only for the values on the struct itself. UpdateParams is used to update all parameters, including Prjn.

func (*Layer) UpdateExtFlags 

func (ly *Layer) UpdateExtFlags(ctx *Context)

UpdateExtFlags updates the neuron flags for external input based on current layer Type field -- call this if the Type has changed since the last ApplyExt* method call.

func (*Layer) UpdateParams 

func (ly *Layer) UpdateParams()

UpdateParams updates all params given any changes that might have been made to individual values including those in the receiving projections of this layer. This is not called Update because it is not just about the local values in the struct.

func (*Layer) VSPatchAdaptThr added in v1.8.1 

func (ly *Layer) VSPatchAdaptThr(ctx *Context)

VSPatchAdaptThr adapts the learning threshold

func (*Layer) WriteWtsJSON 

func (ly *Layer) WriteWtsJSON(w io.Writer, depth int)

WriteWtsJSON writes the weights from this layer from the receiver-side perspective in a JSON text format. We build in the indentation logic to make it much faster and more efficient.

func (*Layer) WtFmDWt 

func (ly *Layer) WtFmDWt(ctx *Context, si uint32)

WtFmDWt updates weight values from delta weight changes

func (*Layer) WtFmDWtLayer added in v1.6.0 

func (ly *Layer) WtFmDWtLayer(ctx *Context)

WtFmDWtLayer does weight update at the layer level. does NOT call main projection-level WtFmDWt method. in base, only calls TrgAvgFmD

type LayerBase added in v1.4.5 

type LayerBase struct {

	// [view: -] we need a pointer to ourselves as an AxonLayer (which subsumes emer.Layer), which can always be used to extract the true underlying type of object when layer is embedded in other structs -- function receivers do not have this ability so this is necessary.
	AxonLay AxonLayer `` /* 297-byte string literal not displayed */

	// [view: -] our parent network, in case we need to use it to find other layers etc -- set when added by network
	Network *Network `` /* 141-byte string literal not displayed */

	// Name of the layer -- this must be unique within the network, which has a map for quick lookup and layers are typically accessed directly by name
	Nm string `` /* 151-byte string literal not displayed */

	// Class is for applying parameter styles, can be space separated multple tags
	Cls string `desc:"Class is for applying parameter styles, can be space separated multple tags"`

	// inactivate this layer -- allows for easy experimentation
	Off bool `desc:"inactivate this layer -- allows for easy experimentation"`

	// shape of the layer -- can be 2D for basic layers and 4D for layers with sub-groups (hypercolumns) -- order is outer-to-inner (row major) so Y then X for 2D and for 4D: Y-X unit pools then Y-X neurons within pools
	Shp etensor.Shape `` /* 219-byte string literal not displayed */

	// type of layer -- Hidden, Input, Target, Compare, or extended type in specialized algorithms -- matches against .Class parameter styles (e.g., .Hidden etc)
	Typ LayerTypes `` /* 161-byte string literal not displayed */

	// [view: inline] [tableview: -] Spatial relationship to other layer, determines positioning
	Rel relpos.Rel `tableview:"-" view:"inline" desc:"Spatial relationship to other layer, determines positioning"`

	// [tableview: -] position of lower-left-hand corner of layer in 3D space, computed from Rel.  Layers are in X-Y width - height planes, stacked vertically in Z axis.
	Ps mat32.Vec3 `` /* 168-byte string literal not displayed */

	// [view: -] a 0..n-1 index of the position of the layer within list of layers in the network. For Axon networks, it only has significance in determining who gets which weights for enforcing initial weight symmetry -- higher layers get weights from lower layers.
	Idx int `` /* 278-byte string literal not displayed */

	// [view: -] number of neurons in the layer
	NNeurons uint32 `view:"-" desc:"number of neurons in the layer"`

	// [view: -] starting index of neurons for this layer within the global Network list
	NeurStIdx uint32 `view:"-" inactive:"-" desc:"starting index of neurons for this layer within the global Network list"`

	// [view: -] number of pools based on layer shape -- at least 1 for layer pool + 4D subpools
	NPools uint32 `view:"-" desc:"number of pools based on layer shape -- at least 1 for layer pool + 4D subpools"`

	// [view: -] maximum amount of input data that can be processed in parallel in one pass of the network. Neuron, Pool, Vals storage is allocated to hold this amount.
	MaxData uint32 `` /* 167-byte string literal not displayed */

	// [view: -] indexes of representative units in the layer, for computationally expensive stats or displays -- also set RepShp
	RepIxs []int `` /* 128-byte string literal not displayed */

	// [view: -] shape of representative units in the layer -- if RepIxs is empty or .Shp is nil, use overall layer shape
	RepShp etensor.Shape `view:"-" desc:"shape of representative units in the layer -- if RepIxs is empty or .Shp is nil, use overall layer shape"`

	// list of receiving projections into this layer from other layers
	RcvPrjns AxonPrjns `desc:"list of receiving projections into this layer from other layers"`

	// list of sending projections from this layer to other layers
	SndPrjns AxonPrjns `desc:"list of sending projections from this layer to other layers"`

	// layer-level state values that are updated during computation -- one for each data parallel -- is a sub-slice of network full set
	Vals []LayerVals `` /* 135-byte string literal not displayed */

	// computes FS-FFFB inhibition and other pooled, aggregate state variables -- has at least 1 for entire layer (lpl = layer pool), and one for each sub-pool if shape supports that (4D) * 1 per data parallel (inner loop).  This is a sub-slice from overall Network Pools slice.  You must iterate over index and use pointer to modify values.
	Pools []Pool `` /* 341-byte string literal not displayed */

	// [view: -] [Neurons][Data] external input values for this layer, allocated from network global Exts slice
	Exts []float32 `view:"-" desc:"[Neurons][Data] external input values for this layer, allocated from network global Exts slice"`

	// [tableview: -] configuration data set when the network is configured, that is used during the network Build() process via PostBuild method, after all the structure of the network has been fully constructed.  In particular, the Params is nil until Build, so setting anything specific in there (e.g., an index to another layer) must be done as a second pass.  Note that Params are all applied after Build and can set user-modifiable params, so this is for more special algorithm structural parameters set during ConfigNet() methods.,
	BuildConfig map[string]string `` /* 537-byte string literal not displayed */

	// [tableview: -] default parameters that are applied prior to user-set parameters -- these are useful for specific layer functionality in specialized brain areas (e.g., PVLV, BG etc) not associated with a layer type, which otherwise is used to hard-code initial default parameters -- typically just set to a literal map.
	DefParams params.Params `` /* 324-byte string literal not displayed */

	// [tableview: -] provides a history of parameters applied to the layer
	ParamsHistory params.HistoryImpl `tableview:"-" desc:"provides a history of parameters applied to the layer"`
}

LayerBase manages the structural elements of the layer, which are common to any Layer type. The Base does not have algorithm-specific methods and parameters, so it can be easily reused for different algorithms, and cleanly separates the algorithm-specific code. Any dependency on the algorithm-level Layer can be captured in the AxonLayer interface, accessed via the AxonLay field.

func (*LayerBase) AddClass added in v1.8.4 

func (ly *LayerBase) AddClass(cls string)

func (*LayerBase) ApplyDefParams added in v1.7.18 

func (ly *LayerBase) ApplyDefParams()

ApplyDefParams applies DefParams default parameters if set Called by Layer.Defaults()

func (*LayerBase) ApplyParams added in v1.4.5 

func (ly *LayerBase) ApplyParams(pars *params.Sheet, setMsg bool) (bool, error)

ApplyParams applies given parameter style Sheet to this layer and its recv projections. Calls UpdateParams on anything set to ensure derived parameters are all updated. If setMsg is true, then a message is printed to confirm each parameter that is set. it always prints a message if a parameter fails to be set. returns true if any params were set, and error if there were any errors.

func (*LayerBase) Build added in v1.7.0 

func (ly *LayerBase) Build() error

Build constructs the layer state, including calling Build on the projections

func (*LayerBase) BuildConfigByName added in v1.7.0 

func (ly *LayerBase) BuildConfigByName(nm string) (string, error)

BuildConfigByName looks for given BuildConfig option by name, and reports & returns an error if not found.

func (*LayerBase) BuildConfigFindLayer added in v1.7.0 

func (ly *LayerBase) BuildConfigFindLayer(nm string, mustName bool) int32

BuildConfigFindLayer looks for BuildConfig of given name and if found, looks for layer with corresponding name. if mustName is true, then an error is logged if the BuildConfig name does not exist. An error is always logged if the layer name is not found. -1 is returned in any case of not found.

func (*LayerBase) BuildPools added in v1.7.0 

func (ly *LayerBase) BuildPools(ctx *Context, nn uint32) error

BuildPools builds the inhibitory pools structures -- nu = number of units in layer

func (*LayerBase) BuildPrjns added in v1.7.0 

func (ly *LayerBase) BuildPrjns(ctx *Context) error

BuildPrjns builds the projections, send-side

func (*LayerBase) BuildSubPools added in v1.7.0 

func (ly *LayerBase) BuildSubPools(ctx *Context)

BuildSubPools initializes neuron start / end indexes for sub-pools

func (*LayerBase) Class added in v1.4.5 

func (ly *LayerBase) Class() string

func (*LayerBase) Config added in v1.4.5 

func (ly *LayerBase) Config(shape []int, typ emer.LayerType)

Config configures the basic properties of the layer

func (*LayerBase) Idx4DFrom2D added in v1.4.5 

func (ly *LayerBase) Idx4DFrom2D(x, y int) ([]int, bool)

func (*LayerBase) Index added in v1.4.5 

func (ly *LayerBase) Index() int

func (*LayerBase) InitName added in v1.4.5 

func (ly *LayerBase) InitName(lay emer.Layer, name string, net emer.Network)

InitName MUST be called to initialize the layer's pointer to itself as an emer.Layer which enables the proper interface methods to be called. Also sets the name, and the parent network that this layer belongs to (which layers may want to retain).

func (*LayerBase) Is2D added in v1.4.5 

func (ly *LayerBase) Is2D() bool

func (*LayerBase) Is4D added in v1.4.5 

func (ly *LayerBase) Is4D() bool

func (*LayerBase) IsOff added in v1.4.5 

func (ly *LayerBase) IsOff() bool

func (*LayerBase) Label added in v1.4.5 

func (ly *LayerBase) Label() string

func (*LayerBase) LayerType added in v1.7.0 

func (ly *LayerBase) LayerType() LayerTypes

func (*LayerBase) LayerVals added in v1.8.0 

func (ly *LayerBase) LayerVals(di uint32) *LayerVals

LayerVals returns LayerVals at given data index

func (*LayerBase) NRecvPrjns added in v1.4.5 

func (ly *LayerBase) NRecvPrjns() int

func (*LayerBase) NSendPrjns added in v1.4.5 

func (ly *LayerBase) NSendPrjns() int

func (*LayerBase) NSubPools added in v1.7.9 

func (ly *LayerBase) NSubPools() int

NSubPools returns the number of sub-pools of neurons according to the shape parameters. 2D shapes have 0 sub pools. For a 4D shape, the pools are the first set of 2 Y,X dims and then the neurons within the pools are the 2nd set of 2 Y,X dims.

func (*LayerBase) Name added in v1.4.5 

func (ly *LayerBase) Name() string

func (*LayerBase) NeurStartIdx added in v1.6.0 

func (ly *LayerBase) NeurStartIdx() int

func (*LayerBase) NonDefaultParams added in v1.4.5 

func (ly *LayerBase) NonDefaultParams() string

NonDefaultParams returns a listing of all parameters in the Layer that are not at their default values -- useful for setting param styles etc.

func (*LayerBase) ParamsApplied added in v1.7.11 

func (ly *LayerBase) ParamsApplied(sel *params.Sel)

ParamsApplied is just to satisfy History interface so reset can be applied

func (*LayerBase) ParamsHistoryReset added in v1.7.11 

func (ly *LayerBase) ParamsHistoryReset()

ParamsHistoryReset resets parameter application history

func (*LayerBase) PlaceAbove added in v1.7.11 

func (ly *LayerBase) PlaceAbove(other *Layer)

PlaceAbove positions the layer above the other layer, using default XAlign = Left, YAlign = Front alignment

func (*LayerBase) PlaceBehind added in v1.7.11 

func (ly *LayerBase) PlaceBehind(other *Layer, space float32)

PlaceBehind positions the layer behind the other layer, with given spacing, using default XAlign = Left alignment

func (*LayerBase) PlaceRightOf added in v1.7.11 

func (ly *LayerBase) PlaceRightOf(other *Layer, space float32)

PlaceRightOf positions the layer to the right of the other layer, with given spacing, using default YAlign = Front alignment

func (*LayerBase) Pool added in v1.7.0 

func (ly *LayerBase) Pool(pi, di uint32) *Pool

Pool returns pool at given pool x data index

func (*LayerBase) Pos added in v1.4.5 

func (ly *LayerBase) Pos() mat32.Vec3

func (*LayerBase) RecipToRecvPrjn added in v1.7.2 

func (ly *LayerBase) RecipToRecvPrjn(rpj *Prjn) (*Prjn, bool)

RecipToRecvPrjn finds the reciprocal projection to the given recv projection within the ly layer. i.e., where ly is instead the *sending* layer to same other layer B that is the sender of the rpj projection we're receiving from. 

ly = A, other layer = B:

rpj: R=A <- S=B spj: S=A -> R=B

returns false if not found.

func (*LayerBase) RecipToSendPrjn added in v1.4.5 

func (ly *LayerBase) RecipToSendPrjn(spj *Prjn) (*Prjn, bool)

RecipToSendPrjn finds the reciprocal projection to the given sending projection within the ly layer. i.e., where ly is instead the *receiving* layer from same other layer B that is the receiver of the spj projection we're sending to. 

ly = A,  other layer = B:

spj: S=A -> R=B rpj: R=A <- S=B

returns false if not found.

func (*LayerBase) RecvName added in v1.8.9 

func (ly *LayerBase) RecvName(receiver string) *Prjn

func (*LayerBase) RecvNameTry added in v1.7.0 

func (ly *LayerBase) RecvNameTry(receiver string) (emer.Prjn, error)

func (*LayerBase) RecvNameType added in v1.8.9 

func (ly *LayerBase) RecvNameType(receiver, typ string) *Prjn

func (*LayerBase) RecvNameTypeTry added in v1.7.0 

func (ly *LayerBase) RecvNameTypeTry(receiver, typ string) (emer.Prjn, error)

func (*LayerBase) RecvPrjn added in v1.4.5 

func (ly *LayerBase) RecvPrjn(idx int) emer.Prjn

func (*LayerBase) RecvPrjnVals added in v1.8.0 

func (ly *LayerBase) RecvPrjnVals(vals *[]float32, varNm string, sendLay emer.Layer, sendIdx1D int, prjnType string) error

RecvPrjnVals fills in values of given synapse variable name, for projection into given sending layer and neuron 1D index, for all receiving neurons in this layer, into given float32 slice (only resized if not big enough). prjnType is the string representation of the prjn type -- used if non-empty, useful when there are multiple projections between two layers. Returns error on invalid var name. If the receiving neuron is not connected to the given sending layer or neuron then the value is set to mat32.NaN(). Returns error on invalid var name or lack of recv prjn (vals always set to nan on prjn err).

func (*LayerBase) RecvPrjns added in v1.4.5 

func (ly *LayerBase) RecvPrjns() *AxonPrjns

func (*LayerBase) RelPos added in v1.4.5 

func (ly *LayerBase) RelPos() relpos.Rel

func (*LayerBase) RepIdxs added in v1.4.5 

func (ly *LayerBase) RepIdxs() []int

func (*LayerBase) RepShape added in v1.4.8 

func (ly *LayerBase) RepShape() *etensor.Shape

RepShape returns the shape to use for representative units

func (*LayerBase) SendName added in v1.8.9 

func (ly *LayerBase) SendName(sender string) *Prjn

func (*LayerBase) SendNameTry added in v1.7.0 

func (ly *LayerBase) SendNameTry(sender string) (emer.Prjn, error)

func (*LayerBase) SendNameType added in v1.8.9 

func (ly *LayerBase) SendNameType(sender, typ string) *Prjn

func (*LayerBase) SendNameTypeTry added in v1.7.0 

func (ly *LayerBase) SendNameTypeTry(sender, typ string) (emer.Prjn, error)

func (*LayerBase) SendPrjn added in v1.4.5 

func (ly *LayerBase) SendPrjn(idx int) emer.Prjn

func (*LayerBase) SendPrjnVals added in v1.8.0 

func (ly *LayerBase) SendPrjnVals(vals *[]float32, varNm string, recvLay emer.Layer, recvIdx1D int, prjnType string) error

SendPrjnVals fills in values of given synapse variable name, for projection into given receiving layer and neuron 1D index, for all sending neurons in this layer, into given float32 slice (only resized if not big enough). prjnType is the string representation of the prjn type -- used if non-empty, useful when there are multiple projections between two layers. Returns error on invalid var name. If the sending neuron is not connected to the given receiving layer or neuron then the value is set to mat32.NaN(). Returns error on invalid var name or lack of recv prjn (vals always set to nan on prjn err).

func (*LayerBase) SendPrjns added in v1.4.5 

func (ly *LayerBase) SendPrjns() *AxonPrjns

func (*LayerBase) SetBuildConfig added in v1.7.0 

func (ly *LayerBase) SetBuildConfig(param, val string)

SetBuildConfig sets named configuration parameter to given string value to be used in the PostBuild stage -- mainly for layer names that need to be looked up and turned into indexes, after entire network is built.

func (*LayerBase) SetClass added in v1.4.5 

func (ly *LayerBase) SetClass(cls string)

func (*LayerBase) SetIndex added in v1.4.5 

func (ly *LayerBase) SetIndex(idx int)

func (*LayerBase) SetName added in v1.4.5 

func (ly *LayerBase) SetName(nm string)

func (*LayerBase) SetOff added in v1.4.5 

func (ly *LayerBase) SetOff(off bool)

func (*LayerBase) SetPos added in v1.4.5 

func (ly *LayerBase) SetPos(pos mat32.Vec3)

func (*LayerBase) SetRelPos added in v1.4.5 

func (ly *LayerBase) SetRelPos(rel relpos.Rel)

func (*LayerBase) SetRepIdxsShape added in v1.4.8 

func (ly *LayerBase) SetRepIdxsShape(idxs, shape []int)

SetRepIdxsShape sets the RepIdxs, and RepShape and as list of dimension sizes

func (*LayerBase) SetShape added in v1.4.5 

func (ly *LayerBase) SetShape(shape []int)

SetShape sets the layer shape and also uses default dim names

func (*LayerBase) SetThread added in v1.4.5 

func (ly *LayerBase) SetThread(thr int)

func (*LayerBase) SetType added in v1.4.5 

func (ly *LayerBase) SetType(typ emer.LayerType)

func (*LayerBase) Shape added in v1.4.5 

func (ly *LayerBase) Shape() *etensor.Shape

func (*LayerBase) Size added in v1.4.5 

func (ly *LayerBase) Size() mat32.Vec2

func (*LayerBase) SubPool added in v1.8.0 

func (ly *LayerBase) SubPool(ctx *Context, ni, di uint32) *Pool

SubPool returns subpool for given neuron, at data index

func (*LayerBase) Thread added in v1.4.5 

func (ly *LayerBase) Thread() int

todo: remove from emer.Layer api

func (*LayerBase) Type added in v1.4.5 

func (ly *LayerBase) Type() emer.LayerType

func (*LayerBase) TypeName added in v1.4.5 

func (ly *LayerBase) TypeName() string

func (*LayerBase) UnitVal added in v1.8.0 

func (ly *LayerBase) UnitVal(varNm string, idx []int, di int) float32

UnitVal returns value of given variable name on given unit, using shape-based dimensional index

func (*LayerBase) UnitVal1D added in v1.8.0 

func (ly *LayerBase) UnitVal1D(varIdx int, idx, di int) float32

UnitVal1D returns value of given variable index on given unit, using 1-dimensional index. returns NaN on invalid index. This is the core unit var access method used by other methods, so it is the only one that needs to be updated for derived layer types.

func (*LayerBase) UnitVals added in v1.8.0 

func (ly *LayerBase) UnitVals(vals *[]float32, varNm string, di int) error

UnitVals fills in values of given variable name on unit, for each unit in the layer, into given float32 slice (only resized if not big enough). Returns error on invalid var name.

func (*LayerBase) UnitValsRepTensor added in v1.8.0 

func (ly *LayerBase) UnitValsRepTensor(tsr etensor.Tensor, varNm string, di int) error

UnitValsRepTensor fills in values of given variable name on unit for a smaller subset of representative units in the layer, into given tensor. This is used for computationally intensive stats or displays that work much better with a smaller number of units. The set of representative units are defined by SetRepIdxs -- all units are used if no such subset has been defined. If tensor is not already big enough to hold the values, it is set to RepShape to hold all the values if subset is defined, otherwise it calls UnitValsTensor and is identical to that. Returns error on invalid var name.

func (*LayerBase) UnitValsTensor added in v1.8.0 

func (ly *LayerBase) UnitValsTensor(tsr etensor.Tensor, varNm string, di int) error

UnitValsTensor returns values of given variable name on unit for each unit in the layer, as a float32 tensor in same shape as layer units.

func (*LayerBase) UnitVarIdx added in v1.8.0 

func (ly *LayerBase) UnitVarIdx(varNm string) (int, error)

UnitVarIdx returns the index of given variable within the Neuron, according to *this layer's* UnitVarNames() list (using a map to lookup index), or -1 and error message if not found.

func (*LayerBase) UnitVarNames added in v1.8.0 

func (ly *LayerBase) UnitVarNames() []string

UnitVarNames returns a list of variable names available on the units in this layer

func (*LayerBase) UnitVarNum added in v1.8.0 

func (ly *LayerBase) UnitVarNum() int

UnitVarNum returns the number of Neuron-level variables for this layer. This is needed for extending indexes in derived types.

func (*LayerBase) UnitVarProps added in v1.8.0 

func (ly *LayerBase) UnitVarProps() map[string]string

UnitVarProps returns properties for variables

func (*LayerBase) VarRange added in v1.7.0 

func (ly *LayerBase) VarRange(varNm string) (min, max float32, err error)

VarRange returns the min / max values for given variable todo: support r. s. projection values

type LayerIdxs added in v1.7.0 

type LayerIdxs struct {

	// layer index
	LayIdx uint32 `inactive:"+" desc:"layer index"`

	// maximum number of data parallel elements
	MaxData uint32 `inactive:"+" desc:"maximum number of data parallel elements"`

	// start of pools for this layer -- first one is always the layer-wide pool
	PoolSt uint32 `inactive:"+" desc:"start of pools for this layer -- first one is always the layer-wide pool"`

	// start of neurons for this layer in global array (same as Layer.NeurStIdx)
	NeurSt uint32 `inactive:"+" desc:"start of neurons for this layer in global array (same as Layer.NeurStIdx)"`

	// number of neurons in layer
	NeurN uint32 `inactive:"+" desc:"number of neurons in layer"`

	// start index into RecvPrjns global array
	RecvSt uint32 `inactive:"+" desc:"start index into RecvPrjns global array"`

	// number of recv projections
	RecvN uint32 `inactive:"+" desc:"number of recv projections"`

	// start index into RecvPrjns global array
	SendSt uint32 `inactive:"+" desc:"start index into RecvPrjns global array"`

	// number of recv projections
	SendN uint32 `inactive:"+" desc:"number of recv projections"`

	// starting index in network global Exts list of external input for this layer -- only for Input / Target / Compare layer types
	ExtsSt uint32 `` /* 144-byte string literal not displayed */

	// layer shape Pools Y dimension -- 1 for 2D
	ShpPlY int32 `inactive:"+" desc:"layer shape Pools Y dimension -- 1 for 2D"`

	// layer shape Pools X dimension -- 1 for 2D
	ShpPlX int32 `inactive:"+" desc:"layer shape Pools X dimension -- 1 for 2D"`

	// layer shape Units Y dimension
	ShpUnY int32 `inactive:"+" desc:"layer shape Units Y dimension"`

	// layer shape Units X dimension
	ShpUnX int32 `inactive:"+" desc:"layer shape Units X dimension"`
	// contains filtered or unexported fields
}

LayerIdxs contains index access into network global arrays for GPU.

func (*LayerIdxs) ExtIdx added in v1.8.0 

func (lx *LayerIdxs) ExtIdx(ni, di uint32) uint32

ExtIdx returns the index for accessing Exts values: [Neuron][Data] Neuron is *layer-relative* lni index -- add the ExtsSt for network level access.

func (*LayerIdxs) PoolIdx added in v1.8.0 

func (lx *LayerIdxs) PoolIdx(pi, di uint32) uint32

PoolIdx returns the global network index for pool with given pool (0 = layer pool, 1+ = subpools) and data parallel indexes

func (*LayerIdxs) ValsIdx added in v1.8.0 

func (lx *LayerIdxs) ValsIdx(di uint32) uint32

ValsIdx returns the global network index for LayerVals with given data parallel index.

type LayerInhibIdxs added in v1.7.10 

type LayerInhibIdxs struct {

	// idx of Layer to get layer-level inhibition from -- set during Build from BuildConfig LayInhib1Name if present -- -1 if not used
	Idx1 int32 `` /* 147-byte string literal not displayed */

	// idx of Layer to get layer-level inhibition from -- set during Build from BuildConfig LayInhib2Name if present -- -1 if not used
	Idx2 int32 `` /* 147-byte string literal not displayed */

	// idx of Layer to get layer-level inhibition from -- set during Build from BuildConfig LayInhib3Name if present -- -1 if not used
	Idx3 int32 `` /* 147-byte string literal not displayed */

	// idx of Layer to geta layer-level inhibition from -- set during Build from BuildConfig LayInhib4Name if present -- -1 if not used
	Idx4 int32 `` /* 148-byte string literal not displayed */
}

LayerInhibIdxs contains indexes of layers for between-layer inhibition

type LayerParams added in v1.7.0 

type LayerParams struct {

	// functional type of layer -- determines functional code path for specialized layer types, and is synchronized with the Layer.Typ value
	LayType LayerTypes `` /* 140-byte string literal not displayed */

	// [view: add-fields] Activation parameters and methods for computing activations
	Acts ActParams `view:"add-fields" desc:"Activation parameters and methods for computing activations"`

	// [view: add-fields] Inhibition parameters and methods for computing layer-level inhibition
	Inhib InhibParams `view:"add-fields" desc:"Inhibition parameters and methods for computing layer-level inhibition"`

	// [view: inline] indexes of layers that contribute between-layer inhibition to this layer -- set these indexes via BuildConfig LayInhibXName (X = 1, 2...)
	LayInhib LayerInhibIdxs `` /* 158-byte string literal not displayed */

	// [view: add-fields] Learning parameters and methods that operate at the neuron level
	Learn LearnNeurParams `view:"add-fields" desc:"Learning parameters and methods that operate at the neuron level"`

	// [view: inline] [viewif: LayType=SuperLayer] BurstParams determine how the 5IB Burst activation is computed from CaSpkP integrated spiking values in Super layers -- thresholded.
	Bursts BurstParams `` /* 181-byte string literal not displayed */

	// [view: inline] [viewif: LayType=[CTLayer,PTPredLayer,PTNotMaintLayer,BLALayer]] params for the CT corticothalamic layer and PTPred layer that generates predictions over the Pulvinar using context -- uses the CtxtGe excitatory input plus stronger NMDA channels to maintain context trace
	CT CTParams `` /* 290-byte string literal not displayed */

	// [view: inline] [viewif: LayType=PulvinarLayer] provides parameters for how the plus-phase (outcome) state of Pulvinar thalamic relay cell neurons is computed from the corresponding driver neuron Burst activation (or CaSpkP if not Super)
	Pulv PulvParams `` /* 241-byte string literal not displayed */

	// [view: inline] [viewif: LayType=MatrixLayer] parameters for BG Striatum Matrix MSN layers, which are the main Go / NoGo gating units in BG.
	Matrix MatrixParams `` /* 144-byte string literal not displayed */

	// [view: inline] [viewif: LayType=GPLayer] type of GP Layer.
	GP GPParams `viewif:"LayType=GPLayer" view:"inline" desc:"type of GP Layer."`

	// [view: inline] [viewif: LayType=VSPatchLayer] parameters for VSPatch learning
	VSPatch VSPatchParams `viewif:"LayType=VSPatchLayer" view:"inline" desc:"parameters for VSPatch learning"`

	// [view: inline] [viewif: LayType=LDTLayer] parameterizes laterodorsal tegmentum ACh salience neuromodulatory signal, driven by superior colliculus stimulus novelty, US input / absence, and OFC / ACC inhibition
	LDT LDTParams `` /* 213-byte string literal not displayed */

	// [view: inline] [viewif: LayType=VTALayer] parameterizes computing overall VTA DA based on LHb PVDA (primary value -- at US time, computed at start of each trial and stored in LHbPVDA global value) and Amygdala (CeM) CS / learned value (LV) activations, which update every cycle.
	VTA VTAParams `` /* 283-byte string literal not displayed */

	// [view: inline] [viewif: LayType=RWPredLayer] parameterizes reward prediction for a simple Rescorla-Wagner learning dynamic (i.e., PV learning in the PVLV framework).
	RWPred RWPredParams `` /* 170-byte string literal not displayed */

	// [view: inline] [viewif: LayType=RWDaLayer] parameterizes reward prediction dopamine for a simple Rescorla-Wagner learning dynamic (i.e., PV learning in the PVLV framework).
	RWDa RWDaParams `` /* 177-byte string literal not displayed */

	// [view: inline] [viewif: LayType=TDIntegLayer] parameterizes TD reward integration layer
	TDInteg TDIntegParams `viewif:"LayType=TDIntegLayer" view:"inline" desc:"parameterizes TD reward integration layer"`

	// [view: inline] [viewif: LayType=TDDaLayer] parameterizes dopamine (DA) signal as the temporal difference (TD) between the TDIntegLayer activations in the minus and plus phase.
	TDDa TDDaParams `` /* 180-byte string literal not displayed */

	// recv and send projection array access info
	Idxs LayerIdxs `desc:"recv and send projection array access info"`
	// contains filtered or unexported fields
}

LayerParams contains all of the layer parameters. These values must remain constant over the course of computation. On the GPU, they are loaded into a uniform.

func (*LayerParams) AllParams added in v1.7.0 

func (ly *LayerParams) AllParams() string

AllParams returns a listing of all parameters in the Layer

func (*LayerParams) ApplyExtFlags added in v1.7.9 

func (ly *LayerParams) ApplyExtFlags(clearMask, setMask *NeuronFlags, toTarg *bool)

ApplyExtFlags gets the clear mask and set mask for updating neuron flags based on layer type, and whether input should be applied to Target (else Ext)

func (*LayerParams) ApplyExtVal added in v1.7.9 

func (ly *LayerParams) ApplyExtVal(ctx *Context, ni, di uint32, val float32)

ApplyExtVal applies given external value to given neuron, setting flags based on type of layer. Should only be called on Input, Target, Compare layers. Negative values are not valid, and will be interpreted as missing inputs.

func (*LayerParams) AvgGeM added in v1.7.9 

func (ly *LayerParams) AvgGeM(ctx *Context, vals *LayerVals, geIntMinusMax, giIntMinusMax float32)

AvgGeM computes the average and max GeInt, GiInt in minus phase (AvgMaxGeM, AvgMaxGiM) stats, updated in MinusPhase, using values that already max across NData.

func (*LayerParams) CTDefaults added in v1.7.0 

func (ly *LayerParams) CTDefaults()

func (*LayerParams) CyclePostCeMLayer added in v1.7.18 

func (ly *LayerParams) CyclePostCeMLayer(ctx *Context, di uint32, lpl *Pool)

func (*LayerParams) CyclePostLDTLayer added in v1.7.18 

func (ly *LayerParams) CyclePostLDTLayer(ctx *Context, di uint32, vals *LayerVals, srcLay1Act, srcLay2Act, srcLay3Act, srcLay4Act float32)

func (*LayerParams) CyclePostLayer added in v1.8.0 

func (ly *LayerParams) CyclePostLayer(ctx *Context, di uint32, lpl *Pool, vals *LayerVals)

CyclePostLayer is called for all layer types

func (*LayerParams) CyclePostPTNotMaintLayer added in v1.7.18 

func (ly *LayerParams) CyclePostPTNotMaintLayer(ctx *Context, di uint32, lpl *Pool)

func (*LayerParams) CyclePostRWDaLayer added in v1.7.9 

func (ly *LayerParams) CyclePostRWDaLayer(ctx *Context, di uint32, vals *LayerVals, pvals *LayerVals)

func (*LayerParams) CyclePostTDDaLayer added in v1.7.9 

func (ly *LayerParams) CyclePostTDDaLayer(ctx *Context, di uint32, vals *LayerVals, ivals *LayerVals)

func (*LayerParams) CyclePostTDIntegLayer added in v1.7.9 

func (ly *LayerParams) CyclePostTDIntegLayer(ctx *Context, di uint32, vals *LayerVals, pvals *LayerVals)

func (*LayerParams) CyclePostTDPredLayer added in v1.7.9 

func (ly *LayerParams) CyclePostTDPredLayer(ctx *Context, di uint32, vals *LayerVals)

func (*LayerParams) CyclePostVSPatchLayer added in v1.7.12 

func (ly *LayerParams) CyclePostVSPatchLayer(ctx *Context, di uint32, pi int32, pl *Pool, vals *LayerVals)

note: needs to iterate over sub-pools in layer!

func (*LayerParams) CyclePostVTALayer added in v1.7.12 

func (ly *LayerParams) CyclePostVTALayer(ctx *Context, di uint32)

func (*LayerParams) Defaults added in v1.7.0 

func (ly *LayerParams) Defaults()

func (*LayerParams) DrivesDefaults added in v1.7.11 

func (ly *LayerParams) DrivesDefaults()

func (*LayerParams) GFmRawSyn added in v1.7.0 

func (ly *LayerParams) GFmRawSyn(ctx *Context, ni, di uint32)

GFmRawSyn computes overall Ge and GiSyn conductances for neuron from GeRaw and GeSyn values, including NMDA, VGCC, AMPA, and GABA-A channels. drvAct is for Pulvinar layers, activation of driving neuron

func (*LayerParams) GNeuroMod added in v1.7.0 

func (ly *LayerParams) GNeuroMod(ctx *Context, ni, di uint32, vals *LayerVals)

GNeuroMod does neuromodulation of conductances

func (*LayerParams) GatherSpikesInit added in v1.7.2 

func (ly *LayerParams) GatherSpikesInit(ctx *Context, ni, di uint32)

GatherSpikesInit initializes G*Raw and G*Syn values for given neuron prior to integration

func (*LayerParams) GiInteg added in v1.7.0 

func (ly *LayerParams) GiInteg(ctx *Context, ni, di uint32, pl *Pool, vals *LayerVals)

GiInteg adds Gi values from all sources including SubPool computed inhib and updates GABAB as well

func (*LayerParams) InitExt added in v1.7.9 

func (ly *LayerParams) InitExt(ctx *Context, ni, di uint32)

InitExt initializes external input state for given neuron

func (*LayerParams) IsInput added in v1.7.9 

func (ly *LayerParams) IsInput() bool

IsInput returns true if this layer is an Input layer. By default, returns true for layers of Type == axon.InputLayer Used to prevent adapting of inhibition or TrgAvg values.

func (*LayerParams) IsInputOrTarget added in v1.7.9 

func (ly *LayerParams) IsInputOrTarget() bool

IsInputOrTarget returns true if this layer is either an Input or a Target layer.

func (*LayerParams) IsLearnTrgAvg added in v1.7.9 

func (ly *LayerParams) IsLearnTrgAvg() bool

IsLearnTrgAvg returns true if this layer has Learn.TrgAvgAct.On set for learning adjustments based on target average activity levels, and the layer is not an input or target layer.

func (*LayerParams) IsTarget added in v1.7.9 

func (ly *LayerParams) IsTarget() bool

IsTarget returns true if this layer is a Target layer. By default, returns true for layers of Type == TargetLayer Other Target layers include the TRCLayer in deep predictive learning. It is used in SynScale to not apply it to target layers. In both cases, Target layers are purely error-driven.

func (*LayerParams) LayPoolGiFmSpikes added in v1.7.0 

func (ly *LayerParams) LayPoolGiFmSpikes(ctx *Context, lpl *Pool, vals *LayerVals)

LayPoolGiFmSpikes computes inhibition Gi from Spikes for layer-level pool. Also grabs updated Context NeuroMod values into LayerVals

func (*LayerParams) LearnTrgAvgErrLRate added in v1.7.9 

func (ly *LayerParams) LearnTrgAvgErrLRate() float32

LearnTrgAvgErrLRate returns the effective error-driven learning rate for adjusting target average activity levels. This is 0 if !IsLearnTrgAvg() and otherwise is Learn.TrgAvgAct.ErrLRate

func (*LayerParams) MinusPhaseNeuron added in v1.7.9 

func (ly *LayerParams) MinusPhaseNeuron(ctx *Context, ni, di uint32, pl *Pool, lpl *Pool, vals *LayerVals)

MinusPhaseNeuron does neuron level minus-phase updating

func (*LayerParams) MinusPhasePool added in v1.7.9 

func (ly *LayerParams) MinusPhasePool(ctx *Context, pl *Pool)

func (*LayerParams) NewStateLayer added in v1.7.9 

func (ly *LayerParams) NewStateLayer(ctx *Context, lpl *Pool, vals *LayerVals)

func (*LayerParams) NewStateLayerActAvg added in v1.8.0 

func (ly *LayerParams) NewStateLayerActAvg(ctx *Context, vals *LayerVals, actMinusAvg, actPlusAvg float32)

NewStateLayerActAvg updates ActAvg.ActMAvg and ActPAvg based on current values that have been averaged across NData already.

func (*LayerParams) NewStateNeuron added in v1.7.9 

func (ly *LayerParams) NewStateNeuron(ctx *Context, ni, di uint32, vals *LayerVals, pl *Pool)

NewStateNeuron handles all initialization at start of new input pattern. Should already have presented the external input to the network at this point.

func (*LayerParams) NewStatePool added in v1.7.9 

func (ly *LayerParams) NewStatePool(ctx *Context, pl *Pool)

func (*LayerParams) PTPredDefaults added in v1.7.11 

func (ly *LayerParams) PTPredDefaults()

func (*LayerParams) PVDefaults added in v1.7.11 

func (ly *LayerParams) PVDefaults()

func (*LayerParams) PlusPhaseNeuron added in v1.7.9 

func (ly *LayerParams) PlusPhaseNeuron(ctx *Context, ni, di uint32, pl *Pool, lpl *Pool, vals *LayerVals)

PlusPhaseNeuron does neuron level plus-phase updating

func (*LayerParams) PlusPhaseNeuronSpecial added in v1.7.11 

func (ly *LayerParams) PlusPhaseNeuronSpecial(ctx *Context, ni, di uint32, pl *Pool, lpl *Pool, vals *LayerVals)

PlusPhaseNeuronSpecial does special layer type neuron level plus-phase updating

func (*LayerParams) PlusPhasePool added in v1.7.9 

func (ly *LayerParams) PlusPhasePool(ctx *Context, pl *Pool)

func (*LayerParams) PlusPhaseStartNeuron added in v1.7.10 

func (ly *LayerParams) PlusPhaseStartNeuron(ctx *Context, ni, di uint32, pl *Pool, lpl *Pool, vals *LayerVals)

PlusPhaseStartNeuron does neuron level plus-phase start: applies Target inputs as External inputs.

func (*LayerParams) PostSpike added in v1.7.0 

func (ly *LayerParams) PostSpike(ctx *Context, ni, di uint32, pl *Pool, vals *LayerVals)

PostSpike does updates at neuron level after spiking has been computed. it is called *after* PostSpikeSpecial. It also updates the CaSpkPCyc stats.

func (*LayerParams) PostSpikeSpecial added in v1.7.0 

func (ly *LayerParams) PostSpikeSpecial(ctx *Context, ni, di uint32, pl *Pool, lpl *Pool, vals *LayerVals)

PostSpikeSpecial does updates at neuron level after spiking has been computed. This is where special layer types add extra code. warning: if more than 1 layer writes to vals, gpu will fail!

func (*LayerParams) PulvDefaults added in v1.7.0 

func (ly *LayerParams) PulvDefaults()

called in Defaults for Pulvinar layer type

func (*LayerParams) RWDefaults added in v1.7.0 

func (ly *LayerParams) RWDefaults()

func (*LayerParams) RWPredDefaults added in v1.7.0 

func (ly *LayerParams) RWPredDefaults()

func (*LayerParams) SpecialPostGs added in v1.7.0 

func (ly *LayerParams) SpecialPostGs(ctx *Context, ni, di uint32, saveVal float32)

SpecialPostGs is used for special layer types to do things after the standard updates in GFmRawSyn. It is passed the saveVal from SpecialPreGs

func (*LayerParams) SpecialPreGs added in v1.7.0 

func (ly *LayerParams) SpecialPreGs(ctx *Context, ni, di uint32, pl *Pool, vals *LayerVals, drvGe float32, nonDrivePct float32) float32

SpecialPreGs is used for special layer types to do things to the conductance values prior to doing the standard updates in GFmRawSyn drvAct is for Pulvinar layers, activation of driving neuron

func (*LayerParams) SpikeFmG added in v1.7.0 

func (ly *LayerParams) SpikeFmG(ctx *Context, ni, di uint32, lpl *Pool)

SpikeFmG computes Vm from Ge, Gi, Gl conductances and then Spike from that

func (*LayerParams) SubPoolGiFmSpikes added in v1.7.0 

func (ly *LayerParams) SubPoolGiFmSpikes(ctx *Context, di uint32, pl *Pool, lpl *Pool, lyInhib bool, giMult float32)

SubPoolGiFmSpikes computes inhibition Gi from Spikes within a sub-pool pl is guaranteed not to be the overall layer pool

func (*LayerParams) TDDefaults added in v1.7.0 

func (ly *LayerParams) TDDefaults()

func (*LayerParams) TDPredDefaults added in v1.7.0 

func (ly *LayerParams) TDPredDefaults()

func (*LayerParams) USDefaults added in v1.7.11 

func (ly *LayerParams) USDefaults()

func (*LayerParams) Update added in v1.7.0 

func (ly *LayerParams) Update()

func (*LayerParams) UrgencyDefaults added in v1.7.18 

func (ly *LayerParams) UrgencyDefaults()

func (*LayerParams) VSGatedDefaults added in v1.7.11 

func (ly *LayerParams) VSGatedDefaults()

func (*LayerParams) VSPatchDefaults added in v1.7.11 

func (ly *LayerParams) VSPatchDefaults()

type LayerTypes added in v1.7.0 

type LayerTypes int32

LayerTypes is an axon-specific layer type enum, that encompasses all the different algorithm types supported. Class parameter styles automatically key off of these types. The first entries must be kept synchronized with the emer.LayerType, although we replace Hidden -> Super. 

const (
	// Super is a superficial cortical layer (lamina 2-3-4)
	// which does not receive direct input or targets.
	// In more generic models, it should be used as a Hidden layer,
	// and maps onto the Hidden type in emer.LayerType.
	SuperLayer LayerTypes = iota

	// Input is a layer that receives direct external input
	// in its Ext inputs.  Biologically, it can be a primary
	// sensory layer, or a thalamic layer.
	InputLayer

	// Target is a layer that receives direct external target inputs
	// used for driving plus-phase learning.
	// Simple target layers are generally not used in more biological
	// models, which instead use predictive learning via Pulvinar
	// or related mechanisms.
	TargetLayer

	// Compare is a layer that receives external comparison inputs,
	// which drive statistics but do NOT drive activation
	// or learning directly.  It is rarely used in axon.
	CompareLayer

	// CT are layer 6 corticothalamic projecting neurons,
	// which drive "top down" predictions in Pulvinar layers.
	// They maintain information over time via stronger NMDA
	// channels and use maintained prior state information to
	// generate predictions about current states forming on Super
	// layers that then drive PT (5IB) bursting activity, which
	// are the plus-phase drivers of Pulvinar activity.
	CTLayer

	// Pulvinar are thalamic relay cell neurons in the higher-order
	// Pulvinar nucleus of the thalamus, and functionally isomorphic
	// neurons in the MD thalamus, and potentially other areas.
	// These cells alternately reflect predictions driven by CT projections,
	// and actual outcomes driven by 5IB Burst activity from corresponding
	// PT or Super layer neurons that provide strong driving inputs.
	PulvinarLayer

	// TRNLayer is thalamic reticular nucleus layer for inhibitory competition
	// within the thalamus.
	TRNLayer

	// PTMaintLayer implements the subset of pyramidal tract (PT)
	// layer 5 intrinsic bursting (5IB) deep neurons that exhibit
	// robust, stable maintenance of activity over the duration of a
	// goal engaged window, modulated by basal ganglia (BG) disinhibitory
	// gating, supported by strong MaintNMDA channels and recurrent excitation.
	// The lateral PTSelfMaint projection uses MaintG to drive GMaintRaw input
	// that feeds into the stronger, longer MaintNMDA channels,
	// and the ThalToPT ModulatoryG projection from BGThalamus multiplicatively
	// modulates the strength of other inputs, such that only at the time of
	// BG gating are these strong enough to drive sustained active maintenance.
	// Use Act.Dend.ModGain to parameterize.
	PTMaintLayer

	// PTPredLayer implements the subset of pyramidal tract (PT)
	// layer 5 intrinsic bursting (5IB) deep neurons that combine
	// modulatory input from PTMaintLayer sustained maintenance and
	// CTLayer dynamic predictive learning that helps to predict
	// state changes during the period of active goal maintenance.
	// This layer provides the primary input to VSPatch US-timing
	// prediction layers, and other layers that require predictive dynamic
	PTPredLayer

	// PTNotMaintLayer implements a tonically active layer that is inhibited
	// by the PTMaintLayer, thereby providing an active representation of
	// the *absence* of maintained PT activity, which is useful for driving
	// appropriate actions (e.g., exploration) when not in goal-engaged mode.
	PTNotMaintLayer

	// MatrixLayer represents the matrisome medium spiny neurons (MSNs)
	// that are the main Go / NoGo gating units in BG.
	// These are strongly modulated by phasic dopamine: D1 = Go, D2 = NoGo.
	MatrixLayer

	// STNLayer represents subthalamic nucleus neurons, with two subtypes:
	// STNp are more strongly driven and get over bursting threshold, driving strong,
	// rapid activation of the KCa channels, causing a long pause in firing, which
	// creates a window during which GPe dynamics resolve Go vs. No balance.
	// STNs are more weakly driven and thus more slowly activate KCa, resulting in
	// a longer period of activation, during which the GPi is inhibited to prevent
	// premature gating based only MtxGo inhibition -- gating only occurs when
	// GPeIn signal has had a chance to integrate its MtxNo inputs.
	STNLayer

	// GPLayer represents a globus pallidus layer in the BG, including:
	// GPeOut, GPeIn, GPeTA (arkypallidal), and GPi.
	// Typically just a single unit per Pool representing a given stripe.
	GPLayer

	// BGThalLayer represents a BG gated thalamic layer,
	// which receives BG gating in the form of an
	// inhibitory projection from GPi.  Located
	// mainly in the Ventral thalamus: VA / VM / VL,
	// and also parts of MD mediodorsal thalamus.
	BGThalLayer

	// VSGated represents explicit coding of VS gating status:
	// JustGated and HasGated (since last US or failed predicted US),
	// For visualization and / or motor action signaling.
	VSGatedLayer

	// BLALayer represents a basolateral amygdala layer
	// which learns to associate arbitrary stimuli (CSs)
	// with behaviorally salient outcomes (USs)
	BLALayer

	// CeMLayer represents a central nucleus of the amygdala layer.
	CeMLayer

	// VSPatchLayer represents a ventral striatum patch layer,
	// which learns to represent the expected amount of dopamine reward
	// and projects both directly with shunting inhibition to the VTA
	// and indirectly via the LHb / RMTg to cancel phasic dopamine firing
	// to expected rewards (i.e., reward prediction error).
	VSPatchLayer

	// LHbLayer represents the lateral habenula, which drives dipping
	// in the VTA.  It tracks the Global LHb values for
	// visualization purposes -- updated by VTALayer.
	LHbLayer

	// DrivesLayer represents the Drives in PVLV framework.
	// It tracks the Global Drives values for
	// visualization and predictive learning purposes.
	DrivesLayer

	// UrgencyLayer represents the Urgency factor in PVLV framework.
	// It tracks the Global Urgency.Urge value for
	// visualization and predictive learning purposes.
	UrgencyLayer

	// USLayer represents a US unconditioned stimulus layer (USpos or USneg).
	// It tracks the Global USpos or USneg, for visualization
	// and predictive learning purposes. Actual US inputs are set in PVLV.
	USLayer

	// PVLayer represents a PV primary value layer (PVpos or PVneg) representing
	// the total primary value as a function of US inputs, drives, and effort.
	// It tracks the Global VTA.PVpos, PVneg values for
	// visualization and predictive learning purposes.
	PVLayer

	// LDTLayer represents the laterodorsal tegmentum layer, which
	// is the primary limbic ACh (acetylcholine) driver to other ACh:
	// BG cholinergic interneurons (CIN) and nucleus basalis ACh areas.
	// The phasic ACh release signals reward salient inputs from CS, US
	// and US omssion, and it drives widespread disinhibition of BG gating
	// and VTA DA firing.
	// It receives excitation from superior colliculus which computes
	// a temporal derivative (stimulus specific adaptation, SSA)
	// of sensory inputs, and inhibitory input from OFC, ACC driving
	// suppression of distracting inputs during goal-engaged states.
	LDTLayer

	// VTALayer represents the ventral tegmental area, which releases
	// dopamine.  It computes final DA value from PVLV-computed
	// LHb PVDA (primary value DA), updated at start of each trial from
	// updated US, Effort, etc state, and cycle-by-cycle LV learned value
	// state reflecting CS inputs, in the Amygdala (CeM).
	// Its activity reflects this DA level, which is effectively broadcast
	// vial Global state values to all layers.
	VTALayer

	// RewLayer represents positive or negative reward values across 2 units,
	// showing spiking rates for each, and Act always represents signed value.
	RewLayer

	// RWPredLayer computes reward prediction for a simple Rescorla-Wagner
	// learning dynamic (i.e., PV learning in the PVLV framework).
	// Activity is computed as linear function of excitatory conductance
	// (which can be negative -- there are no constraints).
	// Use with RWPrjn which does simple delta-rule learning on minus-plus.
	RWPredLayer

	// RWDaLayer computes a dopamine (DA) signal based on a simple Rescorla-Wagner
	// learning dynamic (i.e., PV learning in the PVLV framework).
	// It computes difference between r(t) and RWPred values.
	// r(t) is accessed directly from a Rew layer -- if no external input then no
	// DA is computed -- critical for effective use of RW only for PV cases.
	// RWPred prediction is also accessed directly from Rew layer to avoid any issues.
	RWDaLayer

	// TDPredLayer is the temporal differences reward prediction layer.
	// It represents estimated value V(t) in the minus phase, and computes
	// estimated V(t+1) based on its learned weights in plus phase,
	// using the TDPredPrjn projection type for DA modulated learning.
	TDPredLayer

	// TDIntegLayer is the temporal differences reward integration layer.
	// It represents estimated value V(t) from prior time step in the minus phase,
	// and estimated discount * V(t+1) + r(t) in the plus phase.
	// It gets Rew, PrevPred from Context.NeuroMod, and Special
	// LayerVals from TDPredLayer.
	TDIntegLayer

	// TDDaLayer computes a dopamine (DA) signal as the temporal difference (TD)
	// between the TDIntegLayer activations in the minus and plus phase.
	// These are retrieved from Special LayerVals.
	TDDaLayer

	LayerTypesN
)

The layer types

func (LayerTypes) Desc added in v1.8.14 

func (i LayerTypes) Desc() string

func (*LayerTypes) FromString added in v1.7.0 

func (i *LayerTypes) FromString(s string) error

func (LayerTypes) IsExt added in v1.7.9 

func (lt LayerTypes) IsExt() bool

IsExt returns true if the layer type deals with external input: Input, Target, Compare

func (LayerTypes) MarshalJSON added in v1.7.0 

func (ev LayerTypes) MarshalJSON() ([]byte, error)

func (LayerTypes) String added in v1.7.0 

func (i LayerTypes) String() string

func (*LayerTypes) UnmarshalJSON added in v1.7.0 

func (ev *LayerTypes) UnmarshalJSON(b []byte) error

type LayerVals added in v1.7.0 

type LayerVals struct {

	// [view: -] layer index for these vals
	LayIdx uint32 `view:"-" desc:"layer index for these vals"`

	// [view: -] data index for these vals
	DataIdx uint32 `view:"-" desc:"data index for these vals"`

	// reaction time for this layer in cycles, which is -1 until the Max CaSpkP level (after MaxCycStart) exceeds the Act.Attn.RTThr threshold
	RT float32 `` /* 155-byte string literal not displayed */

	// [view: inline] running-average activation levels used for adaptive inhibition, and other adapting values
	ActAvg ActAvgVals `view:"inline" desc:"running-average activation levels used for adaptive inhibition, and other adapting values"`

	// correlation (centered cosine aka normalized dot product) similarity between ActM, ActP states
	CorSim CorSimStats `desc:"correlation (centered cosine aka normalized dot product) similarity between ActM, ActP states"`

	// [view: inline] special values used to communicate to other layers based on neural values computed on the GPU -- special cross-layer computations happen CPU-side and are sent back into the network via Context on the next cycle -- used for special algorithms such as RL / DA etc
	Special LaySpecialVals `` /* 282-byte string literal not displayed */
	// contains filtered or unexported fields
}

LayerVals holds extra layer state that is updated per layer. It is sync'd down from the GPU to the CPU after every Cycle.

func (*LayerVals) Init added in v1.7.9 

func (lv *LayerVals) Init()

type LearnNeurParams 

type LearnNeurParams struct {

	// [view: inline] parameterizes the neuron-level calcium signals driving learning: CaLrn = NMDA + VGCC Ca sources, where VGCC can be simulated from spiking or use the more complex and dynamic VGCC channel directly.  CaLrn is then integrated in a cascading manner at multiple time scales: CaM (as in calmodulin), CaP (ltP, CaMKII, plus phase), CaD (ltD, DAPK1, minus phase).
	CaLearn CaLrnParams `` /* 376-byte string literal not displayed */

	// [view: inline] parameterizes the neuron-level spike-driven calcium signals, starting with CaSyn that is integrated at the neuron level, and drives synapse-level, pre * post Ca integration, which provides the Tr trace that multiplies error signals, and drives learning directly for Target layers. CaSpk* values are integrated separately at the Neuron level and used for UpdtThr and RLRate as a proxy for the activation (spiking) based learning signal.
	CaSpk CaSpkParams `` /* 456-byte string literal not displayed */

	// [view: inline] NMDA channel parameters used for learning, vs. the ones driving activation -- allows exploration of learning parameters independent of their effects on active maintenance contributions of NMDA, and may be supported by different receptor subtypes
	LrnNMDA chans.NMDAParams `` /* 266-byte string literal not displayed */

	// [view: inline] synaptic scaling parameters for regulating overall average activity compared to neuron's own target level
	TrgAvgAct TrgAvgActParams `` /* 126-byte string literal not displayed */

	// [view: inline] recv neuron learning rate modulation params -- an additional error-based modulation of learning for receiver side: RLRate = |CaSpkP - CaSpkD| / Max(CaSpkP, CaSpkD)
	RLRate RLRateParams `` /* 184-byte string literal not displayed */

	// [view: inline] neuromodulation effects on learning rate and activity, as a function of layer-level DA and ACh values, which are updated from global Context values, and computed from reinforcement learning algorithms
	NeuroMod NeuroModParams `` /* 221-byte string literal not displayed */
}

axon.LearnNeurParams manages learning-related parameters at the neuron-level. This is mainly the running average activations that drive learning

func (*LearnNeurParams) CaFmSpike added in v1.3.5 

func (ln *LearnNeurParams) CaFmSpike(ctx *Context, ni, di uint32)

CaFmSpike updates all spike-driven calcium variables, including CaLrn and CaSpk. Computed after new activation for current cycle is updated.

func (*LearnNeurParams) Defaults 

func (ln *LearnNeurParams) Defaults()

func (*LearnNeurParams) InitNeurCa added in v1.3.9 

func (ln *LearnNeurParams) InitNeurCa(ctx *Context, ni, di uint32)

InitCaLrnSpk initializes the neuron-level calcium learning and spking variables. Called by InitWts (at start of learning).

func (*LearnNeurParams) LrnNMDAFmRaw added in v1.3.11 

func (ln *LearnNeurParams) LrnNMDAFmRaw(ctx *Context, ni, di uint32, geTot float32)

LrnNMDAFmRaw updates the separate NMDA conductance and calcium values based on GeTot = GeRaw + external ge conductance. These are the variables that drive learning -- can be the same as activation but also can be different for testing learning Ca effects independent of activation effects.

func (*LearnNeurParams) Update 

func (ln *LearnNeurParams) Update()

type LearnSynParams 

type LearnSynParams struct {

	// enable learning for this projection
	Learn slbool.Bool `desc:"enable learning for this projection"`

	// [viewif: Learn] learning rate parameters, supporting two levels of modulation on top of base learning rate.
	LRate LRateParams `viewif:"Learn" desc:"learning rate parameters, supporting two levels of modulation on top of base learning rate."`

	// [viewif: Learn] trace-based learning parameters
	Trace TraceParams `viewif:"Learn" desc:"trace-based learning parameters"`

	// [view: inline] [viewif: Learn] kinase calcium Ca integration parameters
	KinaseCa kinase.CaParams `viewif:"Learn" view:"inline" desc:"kinase calcium Ca integration parameters"`
	// contains filtered or unexported fields
}

LearnSynParams manages learning-related parameters at the synapse-level.

func (*LearnSynParams) CHLdWt 

func (ls *LearnSynParams) CHLdWt(suCaP, suCaD, ruCaP, ruCaD float32) float32

CHLdWt returns the error-driven weight change component for a CHL contrastive hebbian learning rule, optionally using the checkmark temporally eXtended Contrastive Attractor Learning (XCAL) function

func (*LearnSynParams) Defaults 

func (ls *LearnSynParams) Defaults()

func (*LearnSynParams) DeltaDWt added in v1.5.1 

func (ls *LearnSynParams) DeltaDWt(plus, minus float32) float32

DeltaDWt returns the error-driven weight change component for a simple delta between a minus and plus phase factor, optionally using the checkmark temporally eXtended Contrastive Attractor Learning (XCAL) function

func (*LearnSynParams) Update 

func (ls *LearnSynParams) Update()

type MatrixParams added in v1.7.0 

type MatrixParams struct {

	// [def: 0.05] threshold on layer Avg SpkMax for Matrix Go and VThal layers to count as having gated
	GateThr float32 `def:"0.05" desc:"threshold on layer Avg SpkMax for Matrix Go and VThal layers to count as having gated"`

	// is this a ventral striatum (VS) matrix layer?  if true, the gating status of this layer is recorded in the Global state, and used for updating effort and other factors.
	IsVS slbool.Bool `` /* 175-byte string literal not displayed */

	// index of other matrix (Go if we are NoGo and vice-versa).    Set during Build from BuildConfig OtherMatrixName
	OtherMatrixIdx int32 `` /* 130-byte string literal not displayed */

	// index of thalamus layer that we gate.  needed to get gating information.  Set during Build from BuildConfig ThalLay1Name if present -- -1 if not used
	ThalLay1Idx int32 `` /* 169-byte string literal not displayed */

	// index of thalamus layer that we gate.  needed to get gating information.  Set during Build from BuildConfig ThalLay2Name if present -- -1 if not used
	ThalLay2Idx int32 `` /* 169-byte string literal not displayed */

	// index of thalamus layer that we gate.  needed to get gating information.  Set during Build from BuildConfig ThalLay3Name if present -- -1 if not used
	ThalLay3Idx int32 `` /* 169-byte string literal not displayed */

	// index of thalamus layer that we gate.  needed to get gating information.  Set during Build from BuildConfig ThalLay4Name if present -- -1 if not used
	ThalLay4Idx int32 `` /* 169-byte string literal not displayed */

	// index of thalamus layer that we gate.  needed to get gating information.  Set during Build from BuildConfig ThalLay5Name if present -- -1 if not used
	ThalLay5Idx int32 `` /* 169-byte string literal not displayed */

	// index of thalamus layer that we gate.  needed to get gating information.  Set during Build from BuildConfig ThalLay6Name if present -- -1 if not used
	ThalLay6Idx int32 `` /* 169-byte string literal not displayed */
	// contains filtered or unexported fields
}

MatrixParams has parameters for BG Striatum Matrix MSN layers These are the main Go / NoGo gating units in BG. DA, ACh learning rate modulation is pre-computed on the recv neuron RLRate variable via NeuroMod. Also uses Pool.Gated for InvertNoGate, updated in PlusPhase prior to DWt call. Must set Learn.NeuroMod.DAMod = D1Mod or D2Mod via SetBuildConfig("DAMod").

func (*MatrixParams) Defaults added in v1.7.0 

func (mp *MatrixParams) Defaults()

func (*MatrixParams) Update added in v1.7.0 

func (mp *MatrixParams) Update()

type MatrixPrjnParams added in v1.7.0 

type MatrixPrjnParams struct {

	// [def: 1] learning rate for when ACh was elevated but no gating took place, in proportion to the level of ACh that indicates the salience of the event.  A low level of this learning prevents the highly maladaptive situation where the BG is not gating and thus no learning can occur.
	NoGateLRate float32 `` /* 287-byte string literal not displayed */
	// contains filtered or unexported fields
}

MatrixPrjnParams for trace-based learning in the MatrixPrjn. A trace of synaptic co-activity is formed, and then modulated by dopamine whenever it occurs. This bridges the temporal gap between gating activity and subsequent activity, and is based biologically on synaptic tags. Trace is applied to DWt and reset at the time of reward.

func (*MatrixPrjnParams) Defaults added in v1.7.0 

func (tp *MatrixPrjnParams) Defaults()

func (*MatrixPrjnParams) Update added in v1.7.0 

func (tp *MatrixPrjnParams) Update()

type NetIdxs added in v1.8.0 

type NetIdxs struct {

	// [min: 1] number of data parallel items to process currently
	NData uint32 `min:"1" desc:"number of data parallel items to process currently"`

	// network index in global Networks list of networks -- needed for GPU shader kernel compatible network variable access functions (e.g., NrnV, SynV etc) in CPU mode
	NetIdx uint32 `` /* 181-byte string literal not displayed */

	// maximum amount of data parallel
	MaxData uint32 `inactive:"+" desc:"maximum amount of data parallel"`

	// number of layers in the network
	NLayers uint32 `inactive:"+" desc:"number of layers in the network"`

	// total number of neurons
	NNeurons uint32 `inactive:"+" desc:"total number of neurons"`

	// total number of pools excluding * MaxData factor
	NPools uint32 `inactive:"+" desc:"total number of pools excluding * MaxData factor"`

	// total number of synapses
	NSyns uint32 `inactive:"+" desc:"total number of synapses"`

	// maximum size in float32 (4 bytes) of a GPU buffer -- needed for GPU access
	GPUMaxBuffFloats uint32 `inactive:"+" desc:"maximum size in float32 (4 bytes) of a GPU buffer -- needed for GPU access"`

	// total number of SynCa banks of GPUMaxBufferBytes arrays in GPU
	GPUSynCaBanks uint32 `inactive:"+" desc:"total number of SynCa banks of GPUMaxBufferBytes arrays in GPU"`

	// total number of PVLV Drives / positive USs
	PVLVNPosUSs uint32 `inactive:"+" desc:"total number of PVLV Drives / positive USs"`

	// total number of PVLV Negative USs
	PVLVNNegUSs uint32 `inactive:"+" desc:"total number of PVLV Negative USs"`

	// offset into GlobalVars for USneg values
	GvUSnegOff uint32 `inactive:"+" desc:"offset into GlobalVars for USneg values"`

	// stride into GlobalVars for USneg values
	GvUSnegStride uint32 `inactive:"+" desc:"stride into GlobalVars for USneg values"`

	// offset into GlobalVars for USpos, Drive, VSPatch values values
	GvUSposOff uint32 `inactive:"+" desc:"offset into GlobalVars for USpos, Drive, VSPatch values values"`

	// stride into GlobalVars for USpos, Drive, VSPatch values
	GvUSposStride uint32 `inactive:"+" desc:"stride into GlobalVars for USpos, Drive, VSPatch values"`
	// contains filtered or unexported fields
}

NetIdxs are indexes and sizes for processing network

func (*NetIdxs) DataIdx added in v1.8.0 

func (ctx *NetIdxs) DataIdx(idx uint32) uint32

DataIdx returns the data index from an overall index over N * MaxData

func (*NetIdxs) DataIdxIsValid added in v1.8.0 

func (ctx *NetIdxs) DataIdxIsValid(li uint32) bool

DataIdxIsValid returns true if the data index is valid (< NData)

func (*NetIdxs) ItemIdx added in v1.8.0 

func (ctx *NetIdxs) ItemIdx(idx uint32) uint32

ItemIdx returns the main item index from an overall index over NItems * MaxData (items = layers, neurons, synapeses)

func (*NetIdxs) LayerIdxIsValid added in v1.8.0 

func (ctx *NetIdxs) LayerIdxIsValid(li uint32) bool

LayerIdxIsValid returns true if the layer index is valid (< NLayers)

func (*NetIdxs) NeurIdxIsValid added in v1.8.0 

func (ctx *NetIdxs) NeurIdxIsValid(ni uint32) bool

NeurIdxIsValid returns true if the neuron index is valid (< NNeurons)

func (*NetIdxs) PoolDataIdxIsValid added in v1.8.0 

func (ctx *NetIdxs) PoolDataIdxIsValid(pi uint32) bool

PoolDataIdxIsValid returns true if the pool*data index is valid (< NPools*MaxData)

func (*NetIdxs) PoolIdxIsValid added in v1.8.0 

func (ctx *NetIdxs) PoolIdxIsValid(pi uint32) bool

PoolIdxIsValid returns true if the pool index is valid (< NPools)

func (*NetIdxs) SynIdxIsValid added in v1.8.0 

func (ctx *NetIdxs) SynIdxIsValid(si uint32) bool

SynIdxIsValid returns true if the synapse index is valid (< NSyns)

func (*NetIdxs) ValsIdx added in v1.8.0 

func (ctx *NetIdxs) ValsIdx(li, di uint32) uint32

ValsIdx returns the global network index for LayerVals with given layer index and data parallel index.

type Network 

type Network struct {
	NetworkBase
}

axon.Network implements the Axon spiking model, building on the algorithm-independent NetworkBase that manages all the infrastructure. 

var (
	// TheNetwork is the one current network in use, needed for GPU shader kernel
	// compatible variable access in CPU mode, for !multinet build tags case.
	// Typically there is just one and it is faster to access directly.
	// This is set in Network.InitName.
	TheNetwork *Network

	// Networks is a global list of networks, needed for GPU shader kernel
	// compatible variable access in CPU mode, for multinet build tags case.
	// This is updated in Network.InitName, which sets NetIdx.
	Networks []*Network
)

func GlobalNetwork added in v1.8.1 

func GlobalNetwork(ctx *Context) *Network

func NewNetwork added in v1.2.94 

func NewNetwork(name string) *Network

NewNetwork returns a new axon Network

func (*Network) AddAmygdala added in v1.7.0 

func (net *Network) AddAmygdala(prefix string, neg bool, nNeurY, nNeurX int, space float32) (blaPosAcq, blaPosExt, blaNegAcq, blaNegExt, cemPos, cemNeg, blaNov *Layer)

AddAmygdala adds a full amygdala complex including BLA, CeM, and LDT. Inclusion of negative valence is optional with neg arg -- neg* layers are nil if not included. Uses the network PVLV.NPosUSs and NNegUSs for number of pools -- must be configured prior to calling this.

func (*Network) AddBG added in v1.7.0 

func (net *Network) AddBG(prefix string, nPoolsY, nPoolsX, nNeurY, nNeurX, gpNeurY, gpNeurX int, space float32) (mtxGo, mtxNo, gpeTA, stnp, stns, gpi *Layer)

AddBG adds MtxGo, MtxNo, GPeOut, GPeIn, GPeTA, STNp, STNs, GPi layers, with given optional prefix. Doesn't return GPeOut, GpeIn which are purely internal. Only the Matrix has pool-based 4D shape by default -- use pool for "role" like elements where matches need to be detected. All GP / STN layers have gpNeur neurons. Appropriate connections are made between layers, using standard styles. space is the spacing between layers (2 typical).

func (*Network) AddBG4D added in v1.7.0 

func (net *Network) AddBG4D(prefix string, nPoolsY, nPoolsX, nNeurY, nNeurX, gpNeurY, gpNeurX int, space float32) (mtxGo, mtxNo, gpeTA, stnp, stns, gpi *Layer)

AddBG4D adds MtxGo, MtxNo, GPeOut, GPeIn, GPeTA, STNp, STNs, GPi layers, with given optional prefix. Doesn't return GPeOut, GpeIn which are purely internal. This version makes 4D pools throughout the GP layers, with Pools representing separable gating domains. All GP / STN layers have gpNeur neurons. Appropriate PoolOneToOne connections are made between layers, using standard styles. space is the spacing between layers (2 typical) A CIN or more widely used RSalienceLayer should be added and project ACh to the MtxGo, No layers.

func (*Network) AddBGThalLayer2D added in v1.7.18 

func (net *Network) AddBGThalLayer2D(name string, nNeurY, nNeurX int) *Layer

AddBGThalLayer2D adds a BG gated thalamus (e.g., VA/VL/VM, MD) Layer of given size, with given name. This version has a 2D structure

func (*Network) AddBGThalLayer4D added in v1.7.18 

func (net *Network) AddBGThalLayer4D(name string, nPoolsY, nPoolsX, nNeurY, nNeurX int) *Layer

AddBGThalLayer4D adds a BG gated thalamus (e.g., VA/VL/VM, MD) Layer of given size, with given name. This version has a 4D structure, with Pools representing separable gating domains.

func (*Network) AddBLALayers added in v1.7.0 

func (net *Network) AddBLALayers(prefix string, pos bool, nUs, nNeurY, nNeurX int, rel relpos.Relations, space float32) (acq, ext *Layer)

AddBLALayers adds two BLA layers, acquisition / extinction / D1 / D2, for positive or negative valence

func (*Network) AddBOA added in v1.7.18 

func (net *Network) AddBOA(ctx *Context, nYneur, popY, popX, bgY, bgX, pfcY, pfcX int, space float32) (vSgpi, urgency, pvPos, blaPosAcq, blaPosExt, blaNegAcq, blaNegExt, blaNov, ofcPosUS, ofcPosUSCT, ofcPosUSPTp, ofcPosVal, ofcPosValCT, ofcPosValPTp, ofcNegUS, ofcNegUSCT, ofcNegUSPTp, accNegVal, accNegValCT, accNegValPTp, accUtil, sc, notMaint *Layer)

AddBOA builds a complete BOA (BG, OFC, ACC) for goal-driven decision making. Uses the network PVLV.NPosUSs and NNegUSs for number of pools -- must be configured prior to calling this. Calls: * AddPVLVOFCus -- PVLV, and OFC us coding Makes all appropriate interconnections and sets default parameters. Needs CS -> BLA, OFC connections to be made. Returns layers most likely to be used for remaining connections and positions.

func (*Network) AddCTLayer2D added in v1.7.0 

func (net *Network) AddCTLayer2D(name string, nNeurY, nNeurX int) *Layer

AddCTLayer2D adds a CT Layer of given size, with given name.

func (*Network) AddCTLayer4D added in v1.7.0 

func (net *Network) AddCTLayer4D(name string, nPoolsY, nPoolsX, nNeurY, nNeurX int) *Layer

AddCTLayer4D adds a CT Layer of given size, with given name.

func (*Network) AddClampDaLayer added in v1.7.0 

func (nt *Network) AddClampDaLayer(name string) *Layer

AddClampDaLayer adds a ClampDaLayer of given name

func (*Network) AddDrivesLayer added in v1.7.11 

func (net *Network) AddDrivesLayer(ctx *Context, nNeurY, nNeurX int) *Layer

AddDrivesLayer adds PVLV layer representing current drive activity, from Global Drive.Drives. Uses a PopCode representation based on LayerParams.Act.PopCode, distributed over given numbers of neurons in the X and Y dimensions, per drive pool.

func (*Network) AddDrivesPulvLayer added in v1.7.11 

func (net *Network) AddDrivesPulvLayer(ctx *Context, nNeurY, nNeurX int, space float32) (drv, drvP *Layer)

AddDrivesPulvLayer adds PVLV layer representing current drive activity, from Global Drive.Drives. Uses a PopCode representation based on LayerParams.Act.PopCode, distributed over given numbers of neurons in the X and Y dimensions, per drive pool. Adds Pulvinar predictive layers for Drives.

func (*Network) AddGPeLayer2D added in v1.7.0 

func (net *Network) AddGPeLayer2D(name string, nNeurY, nNeurX int) *Layer

AddGPLayer2D adds a GPLayer of given size, with given name. Must set the GPType BuildConfig setting to appropriate GPLayerType

func (*Network) AddGPeLayer4D added in v1.7.0 

func (net *Network) AddGPeLayer4D(name string, nPoolsY, nPoolsX, nNeurY, nNeurX int) *Layer

AddGPLayer4D adds a GPLayer of given size, with given name. Makes a 4D structure with Pools representing separable gating domains.

func (*Network) AddGPiLayer2D added in v1.7.0 

func (net *Network) AddGPiLayer2D(name string, nNeurY, nNeurX int) *Layer

AddGPiLayer2D adds a GPiLayer of given size, with given name.

func (*Network) AddGPiLayer4D added in v1.7.0 

func (net *Network) AddGPiLayer4D(name string, nPoolsY, nPoolsX, nNeurY, nNeurX int) *Layer

AddGPiLayer4D adds a GPiLayer of given size, with given name. Makes a 4D structure with Pools representing separable gating domains.

func (*Network) AddHip added in v1.8.6 

func (net *Network) AddHip(ctx *Context, hip *HipConfig, space float32) (ec2, ec3, dg, ca3, ca1, ec5 *Layer)

AddHip adds a new Hippocampal network for episodic memory. Returns layers most likely to be used for remaining connections and positions.

func (*Network) AddInputPulv2D added in v1.7.0 

func (net *Network) AddInputPulv2D(name string, nNeurY, nNeurX int, space float32) (*Layer, *Layer)

AddInputPulv2D adds an Input and Layer of given size, with given name. The Input layer is set as the Driver of the Layer. Both layers have SetClass(name) called to allow shared params.

func (*Network) AddInputPulv4D added in v1.7.0 

func (net *Network) AddInputPulv4D(name string, nPoolsY, nPoolsX, nNeurY, nNeurX int, space float32) (*Layer, *Layer)

AddInputPulv4D adds an Input and Layer of given size, with given name. The Input layer is set as the Driver of the Layer. Both layers have SetClass(name) called to allow shared params.

func (*Network) AddLDTLayer added in v1.7.18 

func (net *Network) AddLDTLayer(prefix string) *Layer

AddLDTLayer adds a LDTLayer

func (*Network) AddMatrixLayer added in v1.7.0 

func (net *Network) AddMatrixLayer(name string, nPoolsY, nPoolsX, nNeurY, nNeurX int, da DAModTypes) *Layer

AddMatrixLayer adds a MatrixLayer of given size, with given name. Assumes that a 4D structure will be used, with Pools representing separable gating domains. da gives the DaReceptor type (D1R = Go, D2R = NoGo)

func (*Network) AddOFCnegUS added in v1.8.16 

func (net *Network) AddOFCnegUS(ctx *Context, nUSs, ofcY, ofcX int, space float32) (ofc, ofcCT, ofcPT, ofcPTp, ofcMD *Layer)

AddOFCnegUS adds orbital frontal cortex negative US-coding layers, for given number of neg US pools (first is effort pool), with given number of units per pool.

func (*Network) AddOFCposUS added in v1.8.16 

func (net *Network) AddOFCposUS(ctx *Context, nUSs, nY, ofcY, ofcX int, space float32) (ofc, ofcCT, ofcPT, ofcPTp, ofcMD, notMaint *Layer)

AddOFCposUS adds orbital frontal cortex positive US-coding layers, for given number of pos US pools (first is novelty / curiosity pool), with given number of units per pool. Also adds a PTNotMaintLayer called NotMaint with nY units.

func (*Network) AddPFC2D added in v1.7.18 

func (net *Network) AddPFC2D(name, thalSuffix string, nNeurY, nNeurX int, decayOnRew bool, space float32) (pfc, pfcCT, pfcPT, pfcPTp, pfcThal *Layer)

AddPFC2D adds a "full stack" of 2D PFC layers: * AddSuperCT2D (Super and CT) * AddPTMaintThal (PTMaint, BGThal) * AddPTPredLayer (PTPred) with given name prefix, which is also set as the Class for all layers, and suffix for the BGThal layer (e.g., "MD" or "VM" etc for different thalamic nuclei). Sets PFCLayer as additional class for all cortical layers. OneToOne, full connectivity is used between layers. decayOnRew determines the Act.Decay.OnRew setting (true of OFC, ACC type for sure). CT layer uses the Medium timescale params.

func (*Network) AddPFC4D added in v1.7.18 

func (net *Network) AddPFC4D(name, thalSuffix string, nPoolsY, nPoolsX, nNeurY, nNeurX int, decayOnRew bool, space float32) (pfc, pfcCT, pfcPT, pfcPTp, pfcThal *Layer)

AddPFC4D adds a "full stack" of 4D PFC layers: * AddSuperCT4D (Super and CT) * AddPTMaintThal (PTMaint, BGThal) * AddPTPredLayer (PTPred) with given name prefix, which is also set as the Class for all layers, and suffix for the BGThal layer (e.g., "MD" or "VM" etc for different thalamic nuclei). Sets PFCLayer as additional class for all cortical layers. OneToOne and PoolOneToOne connectivity is used between layers. decayOnRew determines the Act.Decay.OnRew setting (true of OFC, ACC type for sure). CT layer uses the Medium timescale params. use, e.g., pfcCT.DefParams["Layer.Inhib.Layer.Gi"] = "2.8" to change default params.

func (*Network) AddPTMaintLayer2D added in v1.7.2 

func (net *Network) AddPTMaintLayer2D(name string, nNeurY, nNeurX int) *Layer

AddPTMaintLayer2D adds a PTMaintLayer of given size, with given name.

func (*Network) AddPTMaintLayer4D added in v1.7.2 

func (net *Network) AddPTMaintLayer4D(name string, nPoolsY, nPoolsX, nNeurY, nNeurX int) *Layer

AddPTMaintLayer4D adds a PTMaintLayer of given size, with given name.

func (*Network) AddPTMaintThalForSuper added in v1.7.2 

func (net *Network) AddPTMaintThalForSuper(super, ct *Layer, thalSuffix, prjnClass string, superToPT, ptSelf prjn.Pattern, space float32) (pt, thal *Layer)

AddPTMaintThalForSuper adds a PTMaint pyramidal tract active maintenance layer and a BG gated Thalamus layer for given superficial layer (SuperLayer) and associated CT, with given thal suffix (e.g., MD, VM). PT and Thal have SetClass(super.Name()) called to allow shared params. Projections are made with given classes: SuperToPT, PTSelfMaint, PTtoThal, ThalToPT, with optional extra class The PT and BGThal layers are positioned behind the CT layer.

func (*Network) AddPTNotMaintLayer added in v1.7.11 

func (net *Network) AddPTNotMaintLayer(ptMaint *Layer, nNeurY, nNeurX int, space float32) *Layer

AddPTNotMaintLayer adds a PTNotMaintLayer of given size, for given PTMaintLayer -- places it to the right of this layer, and calls ConnectPTNotMaint to connect the two, using full connectivity.

func (*Network) AddPTPredLayer added in v1.7.11 

func (net *Network) AddPTPredLayer(ptMaint, ct *Layer, ptToPredPrjn, ctToPredPrjn prjn.Pattern, prjnClass string, space float32) (ptPred *Layer)

AddPTPredLayer adds a PTPred pyramidal tract prediction layer for given PTMaint layer and associated CT. Sets SetClass(super.Name()) to allow shared params. Projections are made with given classes: PTtoPred, CTtoPred The PTPred layer is positioned behind the PT layer.

func (*Network) AddPTPredLayer2D added in v1.7.11 

func (net *Network) AddPTPredLayer2D(name string, nNeurY, nNeurX int) *Layer

AddPTPredLayer2D adds a PTPredLayer of given size, with given name.

func (*Network) AddPTPredLayer4D added in v1.7.11 

func (net *Network) AddPTPredLayer4D(name string, nPoolsY, nPoolsX, nNeurY, nNeurX int) *Layer

AddPTPredLayer4D adds a PTPredLayer of given size, with given name.

func (*Network) AddPVLVOFCus added in v1.7.18 

func (net *Network) AddPVLVOFCus(ctx *Context, nYneur, popY, popX, bgY, bgX, ofcY, ofcX int, space float32) (vSgpi, vSmtxGo, vSmtxNo, vSpatch, urgency, usPos, pvPos, usNeg, usNegP, pvNeg, pvNegP, blaPosAcq, blaPosExt, blaNegAcq, blaNegExt, blaNov, ofcPosUS, ofcPosUSCT, ofcPosUSPTp, ofcPosVal, ofcPosValCT, ofcPosValPTp, ofcPosValMD, ofcNegUS, ofcNegUSCT, ofcNegUSPTp, accNegVal, accNegValCT, accNegValPTp, accNegValMD, sc, notMaint *Layer)

AddPVLVOFCus builds a complete PVLV network with OFCposUS (orbital frontal cortex) US-coding layers, OFCposVal (overall value), OFCnegUS for negative value inputs, and ACCnegVal layers. Uses the network PVLV.NPosUSs and NNegUSs for number of pools -- must be configured prior to calling this. Calls: * AddVTALHbLDTLayers * AddPVLVPulvLayers * AddVS * AddAmygdala * AddOFCposUS * AddOFCnegUS Makes all appropriate interconnections and sets default parameters. Needs CS -> BLA, OFC connections to be made. Returns layers most likely to be used for remaining connections and positions.

func (*Network) AddPVLVPulvLayers added in v1.7.18 

func (net *Network) AddPVLVPulvLayers(ctx *Context, nYneur, popY, popX int, space float32) (drives, drivesP, urgency, usPos, usNeg, usPosP, usNegP, pvPos, pvNeg, pvPosP, pvNegP *Layer)

AddPVLVPulvLayers adds PVLV layers for PV-related information visualizing the internal states of the Global state, with Pulvinar prediction layers for training PFC layers. Uses the network PVLV.NPosUSs and NNegUSs for number of pools -- must be configured prior to calling this. * drives = popcode representation of drive strength (no activity for 0) number of active drives comes from Context; popY, popX neurons per pool. * urgency = popcode representation of urgency Go bias factor, popY, popX neurons. * us = popcode per US, positive & negative * pv = popcode representation of final primary value on positive and negative valences -- this is what the dopamine value ends up conding (pos - neg). Layers are organized in depth per type: USs in one column, PVs in the next, with Drives in the back; urgency behind that.

func (*Network) AddPVLayers added in v1.7.11 

func (net *Network) AddPVLayers(nNeurY, nNeurX int, rel relpos.Relations, space float32) (pvPos, pvNeg *Layer)

AddPVLayers adds PVpos and PVneg layers for positive or negative valence primary value representations, representing the total drive and effort weighted USpos outcome, or total USneg outcome. Uses a PopCode representation based on LayerParams.Act.PopCode, distributed over given numbers of neurons in the X and Y dimensions.

func (*Network) AddPVPulvLayers added in v1.7.11 

func (net *Network) AddPVPulvLayers(nNeurY, nNeurX int, rel relpos.Relations, space float32) (pvPos, pvNeg, pvPosP, pvNegP *Layer)

AddPVLayers adds PVpos and PVneg layers for positive or negative valence primary value representations, representing the total drive and effort weighted USpos outcomes, or total USneg outcomes. Uses a PopCode representation based on LayerParams.Act.PopCode, distributed over given numbers of neurons in the X and Y dimensions. Adds Pulvinar predictive layers for each.

func (*Network) AddPulvForLayer added in v1.7.11 

func (net *Network) AddPulvForLayer(lay *Layer, space float32) *Layer

AddPulvForLayer adds a Pulvinar for given Layer (typically an Input type layer) with a P suffix. The Pulv.Driver is set to given Layer. The Pulv layer needs other CT connections from higher up to predict this layer. Pulvinar is positioned behind the given Layer.

func (*Network) AddPulvForSuper added in v1.7.0 

func (net *Network) AddPulvForSuper(super *Layer, space float32) *Layer

AddPulvForSuper adds a Pulvinar for given superficial layer (SuperLayer) with a P suffix. The Pulv.Driver is set to Super, as is the Class on Pulv. The Pulv layer needs other CT connections from higher up to predict this layer. Pulvinar is positioned behind the CT layer.

func (*Network) AddPulvLayer2D added in v1.7.0 

func (net *Network) AddPulvLayer2D(name string, nNeurY, nNeurX int) *Layer

AddPulvLayer2D adds a Pulvinar Layer of given size, with given name.

func (*Network) AddPulvLayer4D added in v1.7.0 

func (net *Network) AddPulvLayer4D(name string, nPoolsY, nPoolsX, nNeurY, nNeurX int) *Layer

AddPulvLayer4D adds a Pulvinar Layer of given size, with given name.

func (*Network) AddRWLayers added in v1.7.0 

func (nt *Network) AddRWLayers(prefix string, rel relpos.Relations, space float32) (rew, rp, da *Layer)

AddRWLayers adds simple Rescorla-Wagner (PV only) dopamine system, with a primary Reward layer, a RWPred prediction layer, and a dopamine layer that computes diff. Only generates DA when Rew layer has external input -- otherwise zero.

func (*Network) AddRewLayer added in v1.7.0 

func (nt *Network) AddRewLayer(name string) *Layer

AddRewLayer adds a RewLayer of given name

func (*Network) AddSCLayer2D added in v1.7.18 

func (net *Network) AddSCLayer2D(prefix string, nNeurY, nNeurX int) *Layer

AddSCLayer2D adds superior colliculcus 2D layer which computes stimulus onset via trial-delayed inhibition (Inhib.FFPrv) -- connect with fixed random input from sensory input layers. Sets base name and class name to SC. Must set Inhib.FFPrv > 0 and Act.Decay.* = 0

func (*Network) AddSCLayer4D added in v1.7.18 

func (net *Network) AddSCLayer4D(prefix string, nPoolsY, nPoolsX, nNeurY, nNeurX int) *Layer

AddSCLayer4D adds superior colliculcus 4D layer which computes stimulus onset via trial-delayed inhibition (Inhib.FFPrv) -- connect with fixed random input from sensory input layers. Sets base name and class name to SC. Must set Inhib.FFPrv > 0 and Act.Decay.* = 0

func (*Network) AddSTNLayer2D added in v1.7.0 

func (net *Network) AddSTNLayer2D(name string, nNeurY, nNeurX int) *Layer

AddSTNLayer2D adds a subthalamic nucleus Layer of given size, with given name.

func (*Network) AddSTNLayer4D added in v1.7.0 

func (net *Network) AddSTNLayer4D(name string, nPoolsY, nPoolsX, nNeurY, nNeurX int) *Layer

AddSTNLayer4D adds a subthalamic nucleus Layer of given size, with given name. Makes a 4D structure with Pools representing separable gating domains.

func (*Network) AddSuperCT2D added in v1.7.0 

func (net *Network) AddSuperCT2D(name, prjnClass string, shapeY, shapeX int, space float32, pat prjn.Pattern) (super, ct *Layer)

AddSuperCT2D adds a superficial (SuperLayer) and corresponding CT (CT suffix) layer with CTCtxtPrjn projection from Super to CT using given projection pattern, and NO Pulv Pulvinar. CT is placed Behind Super.

func (*Network) AddSuperCT4D added in v1.7.0 

func (net *Network) AddSuperCT4D(name, prjnClass string, nPoolsY, nPoolsX, nNeurY, nNeurX int, space float32, pat prjn.Pattern) (super, ct *Layer)

AddSuperCT4D adds a superficial (SuperLayer) and corresponding CT (CT suffix) layer with CTCtxtPrjn projection from Super to CT using given projection pattern, and NO Pulv Pulvinar. CT is placed Behind Super.

func (*Network) AddSuperLayer2D added in v1.7.0 

func (net *Network) AddSuperLayer2D(name string, nNeurY, nNeurX int) *Layer

AddSuperLayer2D adds a Super Layer of given size, with given name.

func (*Network) AddSuperLayer4D added in v1.7.0 

func (net *Network) AddSuperLayer4D(name string, nPoolsY, nPoolsX, nNeurY, nNeurX int) *Layer

AddSuperLayer4D adds a Super Layer of given size, with given name.

func (*Network) AddTDLayers added in v1.7.0 

func (nt *Network) AddTDLayers(prefix string, rel relpos.Relations, space float32) (rew, rp, ri, td *Layer)

AddTDLayers adds the standard TD temporal differences layers, generating a DA signal. Projection from Rew to RewInteg is given class TDRewToInteg -- should have no learning and 1 weight.

func (*Network) AddUSLayers added in v1.7.11 

func (net *Network) AddUSLayers(popY, popX int, rel relpos.Relations, space float32) (usPos, usNeg *Layer)

AddUSLayers adds USpos and USneg layers for positive or negative valence unconditioned stimuli (USs), using a pop-code representation of US magnitude. These track the Global USpos or USneg, for visualization and predictive learning. Actual US inputs are set in PVLV. Uses the network PVLV.NPosUSs and NNegUSs for number of pools -- must be configured prior to calling this.

func (*Network) AddUSPulvLayers added in v1.7.11 

func (net *Network) AddUSPulvLayers(popY, popX int, rel relpos.Relations, space float32) (usPos, usNeg, usPosP, usNegP *Layer)

AddUSPulvLayers adds USpos and USneg layers for positive or negative valence unconditioned stimuli (USs), using a pop-code representation of US magnitude. These track the Global USpos or USneg, for visualization and predictive learning. Actual US inputs are set in PVLV. Adds Pulvinar predictive layers for each.

func (*Network) AddUrgencyLayer added in v1.7.18 

func (net *Network) AddUrgencyLayer(nNeurY, nNeurX int) *Layer

AddUrgencyLayer adds PVLV layer representing current urgency factor, from Global Urgency.Urge Uses a PopCode representation based on LayerParams.Act.PopCode, distributed over given numbers of neurons in the X and Y dimensions.

func (*Network) AddVS added in v1.7.18 

func (net *Network) AddVS(nUSs, nNeurY, nNeurX, nY int, space float32) (vSmtxGo, vSmtxNo, vSstnp, vSstns, vSgpi, vSpatch, vSgated *Layer)

AddVS adds a Ventral Striatum (VS, mostly Nucleus Accumbens = NAcc) set of layers including extensive Ventral Pallidum (VP) using the pcore BG framework, via the AddBG method. Also adds VSPatch and VSGated layers. vSmtxGo and No have VSMatrixLayer class set and default params appropriate for multi-pool etc

func (*Network) AddVSGatedLayer added in v1.7.11 

func (net *Network) AddVSGatedLayer(prefix string, nYunits int) *Layer

AddVSGatedLayer adds a VSGatedLayer with given number of Y units and 2 pools, first one represents JustGated, second is HasGated.

func (*Network) AddVSPatchLayer added in v1.7.11 

func (net *Network) AddVSPatchLayer(prefix string, nUs, nNeurY, nNeurX int) *Layer

AddVSPatchLayer adds VSPatch (Pos, D1)

func (*Network) AddVTALHbLDTLayers added in v1.7.18 

func (net *Network) AddVTALHbLDTLayers(rel relpos.Relations, space float32) (vta, lhb, ldt *Layer)

AddVTALHbLDTLayers adds VTA dopamine, LHb DA dipping, and LDT ACh layers which are driven by corresponding values in Global

func (*Network) ApplyExts added in v1.7.9 

func (nt *Network) ApplyExts(ctx *Context)

ApplyExts applies external inputs to layers, based on values that were set in prior layer-specific ApplyExt calls. This does nothing on the CPU, but is critical for the GPU, and should be added to all sims where GPU will be used.

func (*Network) AsAxon 

func (nt *Network) AsAxon() *Network

func (*Network) ClearTargExt added in v1.2.65 

func (nt *Network) ClearTargExt(ctx *Context)

ClearTargExt clears external inputs Ext that were set from target values Target. This can be called to simulate alpha cycles within theta cycles, for example.

func (*Network) CollectDWts 

func (nt *Network) CollectDWts(ctx *Context, dwts *[]float32) bool

CollectDWts writes all of the synaptic DWt values to given dwts slice which is pre-allocated to given nwts size if dwts is nil, in which case the method returns true so that the actual length of dwts can be passed next time around. Used for MPI sharing of weight changes across processors. This calls SyncSynapsesFmGPU() (nop if not GPU) first.

func (*Network) ConfigGPUnoGUI added in v1.7.9 

func (nt *Network) ConfigGPUnoGUI(ctx *Context)

ConfigGPUnoGUI turns on GPU mode in case where no GUI is being used. This directly accesses the GPU hardware. It does not work well when GUI also being used. Configures the GPU -- call after Network is Built, initialized, params are set, and everything is ready to run.

func (*Network) ConfigGPUwithGUI added in v1.7.9 

func (nt *Network) ConfigGPUwithGUI(ctx *Context)

ConfigGPUwithGUI turns on GPU mode in context of an active GUI where Vulkan has been initialized etc. Configures the GPU -- call after Network is Built, initialized, params are set, and everything is ready to run.

func (*Network) ConfigLoopsHip added in v1.8.7 

func (net *Network) ConfigLoopsHip(ctx *Context, man *looper.Manager, hip *HipConfig, pretrain *bool)

ConfigLoopsHip configures the hippocampal looper and should be included in ConfigLoops in model to make sure hip loops is configured correctly. see hip.go for an instance of implementation of this function. ec5ClampFrom specifies the layer to clamp EC5 plus phase values from: EC3 is the biological source, but can use Input layer for simple testing net.

func (*Network) ConnectCSToBLAPos added in v1.7.18 

func (net *Network) ConnectCSToBLAPos(cs, blaAcq, blaNov *Layer) (toAcq, toNov *Prjn)

ConnectCSToBLAPos connects the CS input to BLAPosAcqD1, BLANovelCS layers using fixed, higher-variance weights, full projection. Sets classes to: CSToBLAPos, CSToBLANovel with default params

func (*Network) ConnectCTSelf added in v1.7.0 

func (net *Network) ConnectCTSelf(ly *Layer, pat prjn.Pattern, prjnClass string) (ctxt, maint *Prjn)

ConnectCTSelf adds a Self (Lateral) CTCtxtPrjn projection within a CT layer, in addition to a regular lateral projection, which supports active maintenance. The CTCtxtPrjn has a Class label of CTSelfCtxt, and the regular one is CTSelfMaint with optional class added.

func (*Network) ConnectCtxtToCT added in v1.7.0 

func (net *Network) ConnectCtxtToCT(send, recv *Layer, pat prjn.Pattern) *Prjn

ConnectCtxtToCT adds a CTCtxtPrjn from given sending layer to a CT layer

func (*Network) ConnectPTMaintSelf added in v1.7.2 

func (net *Network) ConnectPTMaintSelf(ly *Layer, pat prjn.Pattern, prjnClass string) *Prjn

ConnectPTMaintSelf adds a Self (Lateral) projection within a PTMaintLayer, which supports active maintenance, with a class of PTSelfMaint

func (*Network) ConnectPTNotMaint added in v1.7.11 

func (net *Network) ConnectPTNotMaint(ptMaint, ptNotMaint *Layer, pat prjn.Pattern) *Prjn

ConnectPTNotMaint adds a projection from PTMaintLayer to PTNotMaintLayer, as fixed inhibitory connections, with class ToPTNotMaintInhib

func (*Network) ConnectPTPredSelf added in v1.7.11 

func (net *Network) ConnectPTPredSelf(ly *Layer, pat prjn.Pattern) *Prjn

ConnectPTPredSelf adds a Self (Lateral) projection within a PTPredLayer, which supports active maintenance, with a class of PTSelfMaint

func (*Network) ConnectPTPredToPulv added in v1.7.11 

func (net *Network) ConnectPTPredToPulv(ptPred, pulv *Layer, toPulvPat, fmPulvPat prjn.Pattern, prjnClass string) (toPulv, toPTPred *Prjn)

ConnectPTPredToPulv connects PTPred with given Pulv: PTPred -> Pulv is class PTPredToPulv, From Pulv = type = Back, class = FmPulv toPulvPat is the prjn.Pattern PTPred -> Pulv and fmPulvPat is Pulv -> PTPred Typically Pulv is a different shape than PTPred, so use Full or appropriate topological pattern. adds optional class name to projection.

func (*Network) ConnectSuperToCT added in v1.7.0 

func (net *Network) ConnectSuperToCT(send, recv *Layer, pat prjn.Pattern, prjnClass string) *Prjn

ConnectSuperToCT adds a CTCtxtPrjn from given sending Super layer to a CT layer This automatically sets the FmSuper flag to engage proper defaults, Uses given projection pattern -- e.g., Full, OneToOne, or PoolOneToOne

func (*Network) ConnectToBLAAcq added in v1.7.11 

func (net *Network) ConnectToBLAAcq(send, recv *Layer, pat prjn.Pattern) *Prjn

ConnectToBLAAcq adds a BLAPrjn from given sending layer to a BLA layer, and configures it for acquisition parameters. Sets class to BLAAcqPrjn.

func (*Network) ConnectToBLAExt added in v1.7.11 

func (net *Network) ConnectToBLAExt(send, recv *Layer, pat prjn.Pattern) *Prjn

ConnectToBLAExt adds a BLAPrjn from given sending layer to a BLA layer, and configures it for extinctrion parameters. Sets class to BLAExtPrjn.

func (*Network) ConnectToMatrix added in v1.7.0 

func (net *Network) ConnectToMatrix(send, recv *Layer, pat prjn.Pattern) *Prjn

ConnectToMatrix adds a MatrixPrjn from given sending layer to a matrix layer

func (*Network) ConnectToPFC added in v1.7.18 

func (net *Network) ConnectToPFC(lay, layP, pfc, pfcCT, pfcPTp *Layer, pat prjn.Pattern)

ConnectToPFC connects given predictively learned input to all relevant PFC layers: lay -> pfc (skipped if lay == nil) layP -> pfc, layP <-> pfcCT pfcPTp <-> layP sets PFCPrjn class name for projections

func (*Network) ConnectToPFCBack added in v1.7.18 

func (net *Network) ConnectToPFCBack(lay, layP, pfc, pfcCT, pfcPTp *Layer, pat prjn.Pattern)

ConnectToPFCBack connects given predictively learned input to all relevant PFC layers: lay -> pfc using a BackPrjn -- weaker layP -> pfc, layP <-> pfcCT pfcPTp <-> layP

func (*Network) ConnectToPFCBidir added in v1.7.18 

func (net *Network) ConnectToPFCBidir(lay, layP, pfc, pfcCT, pfcPTp *Layer, pat prjn.Pattern) (ff, fb *Prjn)

ConnectToPFCBidir connects given predictively learned input to all relevant PFC layers, using bidirectional connections to super layers. lay <-> pfc bidirectional layP -> pfc, layP <-> pfcCT pfcPTp <-> layP

func (*Network) ConnectToPulv added in v1.7.0 

func (net *Network) ConnectToPulv(super, ct, pulv *Layer, toPulvPat, fmPulvPat prjn.Pattern, prjnClass string) (toPulv, toSuper, toCT *Prjn)

ConnectToPulv adds the following projections: layers | class | prjn type | prjn pat ------------+------------+-------------+---------- ct ->pulv | "CTToPulv" | ForwardPrjn | toPulvPat pulv->super | "FmPulv" | BackPrjn | fmPulvPat pulv->ct | "FmPulv" | BackPrjn | fmPulvPat

Typically pulv is a different shape than super and ct, so use Full or appropriate topological pattern. Adds optional prjnClass name as a suffix.

func (*Network) ConnectToRWPrjn added in v1.7.0 

func (nt *Network) ConnectToRWPrjn(send, recv *Layer, pat prjn.Pattern) *Prjn

ConnectToRWPred adds a RWPrjn from given sending layer to a RWPred layer

func (*Network) ConnectToSC added in v1.7.18 

func (net *Network) ConnectToSC(send, recv *Layer, pat prjn.Pattern) *Prjn

ConnectToSC adds a ForwardPrjn from given sending layer to a SC layer, setting class as ToSC -- should set params as fixed random with more variance than usual.

func (*Network) ConnectToSC1to1 added in v1.7.18 

func (net *Network) ConnectToSC1to1(send, recv *Layer) *Prjn

ConnectToSC1to1 adds a 1to1 ForwardPrjn from given sending layer to a SC layer, copying the geometry of the sending layer, setting class as ToSC. The conection weights are set to uniform.

func (*Network) ConnectToVSPatch added in v1.7.11 

func (net *Network) ConnectToVSPatch(send, recv *Layer, pat prjn.Pattern) *Prjn

ConnectToVSPatch adds a VSPatchPrjn from given sending layer to a VSPatch layer

func (*Network) ConnectUSToBLA added in v1.8.14 

func (net *Network) ConnectUSToBLA(us, blaAcq, blaExt *Layer) (toAcq, toExt *Prjn)

ConnectUSToBLA connects the US input to BLAPos(Neg)AcqD1(D2) and BLAPos(Neg)ExtD2(D1) layers, using fixed, higher-variance weights, full projection. Sets classes to: USToBLAAcq and USToBLAExt

func (*Network) Cycle 

func (nt *Network) Cycle(ctx *Context)

Cycle runs one cycle of activation updating using threading methods.

func (*Network) DWt 

func (nt *Network) DWt(ctx *Context)

DWt computes the weight change (learning) based on current running-average activation values

func (*Network) DecayState 

func (nt *Network) DecayState(ctx *Context, decay, glong, ahp float32)

DecayState decays activation state by given proportion e.g., 1 = decay completely, and 0 = decay not at all. glong = separate decay factor for long-timescale conductances (g) This is called automatically in NewState, but is avail here for ad-hoc decay cases.

func (*Network) DecayStateByClass added in v1.5.10 

func (nt *Network) DecayStateByClass(ctx *Context, decay, glong, ahp float32, classes ...string)

DecayStateByClass decays activation state for given class name(s) by given proportion e.g., 1 = decay completely, and 0 = decay not at all. glong = separate decay factor for long-timescale conductances (g)

func (*Network) DecayStateByType added in v1.7.1 

func (nt *Network) DecayStateByType(ctx *Context, decay, glong, ahp float32, types ...LayerTypes)

DecayStateByType decays activation state for given layer types by given proportion e.g., 1 = decay completely, and 0 = decay not at all. glong = separate decay factor for long-timescale conductances (g)

func (*Network) DecayStateLayers added in v1.7.10 

func (nt *Network) DecayStateLayers(ctx *Context, decay, glong, ahp float32, layers ...string)

DecayStateLayers decays activation state for given layers by given proportion e.g., 1 = decay completely, and 0 = decay not at all. glong = separate decay factor for long-timescale conductances (g). If this is not being called at the start, around NewState call, then you should also call: nt.GPU.SyncGBufToGPU() to zero the GBuf values which otherwise will persist spikes in flight.

func (*Network) Defaults 

func (nt *Network) Defaults()

Defaults sets all the default parameters for all layers and projections

func (*Network) InitActs 

func (nt *Network) InitActs(ctx *Context)

InitActs fully initializes activation state -- not automatically called

func (*Network) InitExt 

func (nt *Network) InitExt(ctx *Context)

InitExt initializes external input state. Call prior to applying external inputs to layers.

func (*Network) InitGScale added in v1.2.92 

func (nt *Network) InitGScale(ctx *Context)

InitGScale computes the initial scaling factor for synaptic input conductances G, stored in GScale.Scale, based on sending layer initial activation.

func (*Network) InitName added in v1.8.0 

func (nt *Network) InitName(net emer.Network, name string)

InitName MUST be called to initialize the network's pointer to itself as an emer.Network which enables the proper interface methods to be called. Also sets the name, and initializes NetIdx in global list of Network

func (*Network) InitTopoSWts added in v1.2.75 

func (nt *Network) InitTopoSWts()

InitTopoSWts initializes SWt structural weight parameters from prjn types that support topographic weight patterns, having flags set to support it, includes: prjn.PoolTile prjn.Circle. call before InitWts if using Topo wts

func (*Network) InitWts 

func (nt *Network) InitWts(ctx *Context)

InitWts initializes synaptic weights and all other associated long-term state variables including running-average state values (e.g., layer running average activations etc)

func (*Network) LRateMod added in v1.6.13 

func (nt *Network) LRateMod(mod float32)

LRateMod sets the LRate modulation parameter for Prjns, which is for dynamic modulation of learning rate (see also LRateSched). Updates the effective learning rate factor accordingly.

func (*Network) LRateSched added in v1.6.13 

func (nt *Network) LRateSched(sched float32)

LRateSched sets the schedule-based learning rate multiplier. See also LRateMod. Updates the effective learning rate factor accordingly.

func (*Network) LayersSetOff 

func (nt *Network) LayersSetOff(off bool)

LayersSetOff sets the Off flag for all layers to given setting

func (*Network) MinusPhase added in v1.2.63 

func (nt *Network) MinusPhase(ctx *Context)

MinusPhase does updating after end of minus phase

func (*Network) NeuronsSlice added in v1.8.0 

func (nt *Network) NeuronsSlice(vals *[]float32, nrnVar string, di int)

NeuronsSlice returns a slice of neuron values using given neuron variable, resizing as needed.

func (*Network) NewLayer 

func (nt *Network) NewLayer() emer.Layer

NewLayer returns new layer of proper type

func (*Network) NewPrjn 

func (nt *Network) NewPrjn() emer.Prjn

NewPrjn returns new prjn of proper type

func (*Network) NewState added in v1.2.63 

func (nt *Network) NewState(ctx *Context)

NewState handles all initialization at start of new input pattern. This is called *before* applying external input data and operates across all data parallel values. The current Context.NData should be set properly prior to calling this and subsequent Cycle methods.

func (*Network) PlusPhase added in v1.2.63 

func (nt *Network) PlusPhase(ctx *Context)

PlusPhase does updating after end of plus phase

func (*Network) PlusPhaseStart added in v1.7.10 

func (nt *Network) PlusPhaseStart(ctx *Context)

PlusPhaseStart does updating at the start of the plus phase: applies Target inputs as External inputs.

func (*Network) SetDWts 

func (nt *Network) SetDWts(ctx *Context, dwts []float32, navg int)

SetDWts sets the DWt weight changes from given array of floats, which must be correct size navg is the number of processors aggregated in these dwts -- some variables need to be averaged instead of summed (e.g., ActAvg) This calls SyncSynapsesToGPU() (nop if not GPU) after.

func (*Network) SetSubMean added in v1.6.11 

func (nt *Network) SetSubMean(trgAvg, prjn float32)

SetSubMean sets the SubMean parameters in all the layers in the network trgAvg is for Learn.TrgAvgAct.SubMean prjn is for the prjns Learn.Trace.SubMean in both cases, it is generally best to have both parameters set to 0 at the start of learning

func (*Network) SizeReport 

func (nt *Network) SizeReport(detail bool) string

SizeReport returns a string reporting the size of each layer and projection in the network, and total memory footprint. If detail flag is true, details per layer, projection is included.

func (*Network) SlowAdapt added in v1.2.37 

func (nt *Network) SlowAdapt(ctx *Context)

SlowAdapt is the layer-level slow adaptation functions: Synaptic scaling, and adapting inhibition

func (*Network) SpkSt1 added in v1.5.10 

func (nt *Network) SpkSt1(ctx *Context)

SpkSt1 saves current acts into SpkSt1 (using CaSpkP)

func (*Network) SpkSt2 added in v1.5.10 

func (nt *Network) SpkSt2(ctx *Context)

SpkSt2 saves current acts into SpkSt2 (using CaSpkP)

func (*Network) SynFail added in v1.2.92 

func (nt *Network) SynFail(ctx *Context)

SynFail updates synaptic failure

func (*Network) SynsSlice added in v1.8.0 

func (nt *Network) SynsSlice(vals *[]float32, synvar SynapseVars)

SynsSlice returns a slice of synaptic values, in natural sending order, using given synaptic variable, resizing as needed.

func (*Network) TargToExt added in v1.2.65 

func (nt *Network) TargToExt(ctx *Context)

TargToExt sets external input Ext from target values Target This is done at end of MinusPhase to allow targets to drive activity in plus phase. This can be called separately to simulate alpha cycles within theta cycles, for example.

func (*Network) UnLesionNeurons 

func (nt *Network) UnLesionNeurons(ctx *Context)

UnLesionNeurons unlesions neurons in all layers in the network. Provides a clean starting point for subsequent lesion experiments.

func (*Network) UpdateExtFlags 

func (nt *Network) UpdateExtFlags(ctx *Context)

UpdateExtFlags updates the neuron flags for external input based on current layer Type field -- call this if the Type has changed since the last ApplyExt* method call.

func (*Network) UpdateParams 

func (nt *Network) UpdateParams()

UpdateParams updates all the derived parameters if any have changed, for all layers and projections

func (*Network) WtFmDWt 

func (nt *Network) WtFmDWt(ctx *Context)

WtFmDWt updates the weights from delta-weight changes. Also does ctx.SlowInc() and calls SlowAdapt at SlowInterval

func (*Network) WtsHash added in v1.7.14 

func (nt *Network) WtsHash() string

WtsHash returns a hash code of all weight values

type NetworkBase added in v1.4.5 

type NetworkBase struct {

	// [view: -] we need a pointer to ourselves as an emer.Network, which can always be used to extract the true underlying type of object when network is embedded in other structs -- function receivers do not have this ability so this is necessary.
	EmerNet emer.Network `` /* 274-byte string literal not displayed */

	// overall name of network -- helps discriminate if there are multiple
	Nm string `desc:"overall name of network -- helps discriminate if there are multiple"`

	// filename of last weights file loaded or saved
	WtsFile string `desc:"filename of last weights file loaded or saved"`

	// PVLV system for phasic dopamine signaling, including internal drives, US outcomes.  Core LHb (lateral habenula) and VTA (ventral tegmental area) dopamine are computed in equations using inputs from specialized network layers (LDTLayer driven by BLA, CeM layers, VSPatchLayer).  Renders USLayer, PVLayer, DrivesLayer representations based on state updated here.
	PVLV PVLV `` /* 367-byte string literal not displayed */

	// [view: -] map of name to layers -- layer names must be unique
	LayMap map[string]*Layer `view:"-" desc:"map of name to layers -- layer names must be unique"`

	// [view: -] map of layer classes -- made during Build
	LayClassMap map[string][]string `view:"-" desc:"map of layer classes -- made during Build"`

	// [view: -] minimum display position in network
	MinPos mat32.Vec3 `view:"-" desc:"minimum display position in network"`

	// [view: -] maximum display position in network
	MaxPos mat32.Vec3 `view:"-" desc:"maximum display position in network"`

	// optional metadata that is saved in network weights files -- e.g., can indicate number of epochs that were trained, or any other information about this network that would be useful to save
	MetaData map[string]string `` /* 194-byte string literal not displayed */

	// if true, the neuron and synapse variables will be organized into a gpu-optimized memory order, otherwise cpu-optimized. This must be set before network Build() is called.
	UseGPUOrder bool `` /* 190-byte string literal not displayed */

	// [view: -] network index in global Networks list of networks -- needed for GPU shader kernel compatible network variable access functions (e.g., NrnV, SynV etc) in CPU mode
	NetIdx uint32 `` /* 177-byte string literal not displayed */

	// [view: -] maximum synaptic delay across any projection in the network -- used for sizing the GBuf accumulation buffer.
	MaxDelay uint32 `` /* 137-byte string literal not displayed */

	// maximum number of data inputs that can be processed in parallel in one pass of the network. Neuron storage is allocated to hold this amount during Build process, and this value reflects that.
	MaxData uint32 `` /* 211-byte string literal not displayed */

	// total number of neurons
	NNeurons uint32 `inactive:"+" desc:"total number of neurons"`

	// total number of synapses
	NSyns uint32 `inactive:"+" desc:"total number of synapses"`

	// [view: -] storage for global vars
	Globals []float32 `view:"-" desc:"storage for global vars"`

	// array of layers
	Layers []*Layer `desc:"array of layers"`

	// [view: -] [Layers] array of layer parameters, in 1-to-1 correspondence with Layers
	LayParams []LayerParams `view:"-" desc:"[Layers] array of layer parameters, in 1-to-1 correspondence with Layers"`

	// [view: -] [Layers][MaxData] array of layer values, with extra per data
	LayVals []LayerVals `view:"-" desc:"[Layers][MaxData] array of layer values, with extra per data"`

	// [view: -] [Layers][Pools][MaxData] array of inhibitory pools for all layers.
	Pools []Pool `view:"-" desc:"[Layers][Pools][MaxData] array of inhibitory pools for all layers."`

	// [view: -] [Layers][Neurons][MaxData] entire network's allocation of neuron variables, accessed via NrnV function with flexible striding
	Neurons []float32 `` /* 141-byte string literal not displayed */

	// [view: -] [Layers][Neurons][MaxData]] entire network's allocation of neuron average avariables, accessed via NrnAvgV function with flexible striding
	NeuronAvgs []float32 `` /* 154-byte string literal not displayed */

	// [view: -] [Layers][Neurons] entire network's allocation of neuron index variables, accessed via NrnI function with flexible striding
	NeuronIxs []uint32 `` /* 138-byte string literal not displayed */

	// [view: -] [Layers][SendPrjns] pointers to all projections in the network, sender-based
	Prjns []*Prjn `view:"-" desc:"[Layers][SendPrjns] pointers to all projections in the network, sender-based"`

	// [view: -] [Layers][SendPrjns] array of projection parameters, in 1-to-1 correspondence with Prjns, sender-based
	PrjnParams []PrjnParams `view:"-" desc:"[Layers][SendPrjns] array of projection parameters, in 1-to-1 correspondence with Prjns, sender-based"`

	// [view: -] [Layers][SendPrjns][SendNeurons][RecvNeurons] entire network's allocation of synapse idx vars, organized sender-based, with flexible striding, accessed via SynI function
	SynapseIxs []uint32 `` /* 185-byte string literal not displayed */

	// [view: -] [Layers][SendPrjns][SendNeurons][RecvNeurons] entire network's allocation of synapses, organized sender-based, with flexible striding, accessed via SynV function
	Synapses []float32 `` /* 177-byte string literal not displayed */

	// [view: -] [Layers][SendPrjns][SendNeurons][RecvNeurons][MaxData] entire network's allocation of synapse Ca vars, organized sender-based, with flexible striding, accessed via SynCaV function
	SynapseCas []float32 `` /* 195-byte string literal not displayed */

	// [view: -] [Layers][SendPrjns][SendNeurons] starting offset and N cons for each sending neuron, for indexing into the Syns synapses, which are organized sender-based.
	PrjnSendCon []StartN `` /* 171-byte string literal not displayed */

	// [view: -] [Layers][RecvPrjns][RecvNeurons] starting offset and N cons for each recv neuron, for indexing into the RecvSynIdx array of indexes into the Syns synapses, which are organized sender-based.
	PrjnRecvCon []StartN `` /* 205-byte string literal not displayed */

	// [view: -] [Layers][RecvPrjns][RecvNeurons][MaxDelay][MaxData] conductance buffer for accumulating spikes -- subslices are allocated to each projection -- uses int-encoded float values for faster GPU atomic integration
	PrjnGBuf []int32 `` /* 223-byte string literal not displayed */

	// [view: -] [Layers][RecvPrjns][RecvNeurons][MaxData] synaptic conductance integrated over time per projection per recv neurons -- spikes come in via PrjnBuf -- subslices are allocated to each projection
	PrjnGSyns []float32 `` /* 207-byte string literal not displayed */

	// [view: -] [Layers][RecvPrjns] indexes into Prjns (organized by SendPrjn) organized by recv projections -- needed for iterating through recv prjns efficiently on GPU.
	RecvPrjnIdxs []uint32 `` /* 171-byte string literal not displayed */

	// [view: -] [Layers][RecvPrjns][RecvNeurons][Syns] indexes into Synapses for each recv neuron, organized into blocks according to PrjnRecvCon, for receiver-based access.
	RecvSynIdxs []uint32 `` /* 173-byte string literal not displayed */

	// [In / Targ Layers][Neurons][Data] external input values for all Input / Target / Compare layers in the network -- the ApplyExt methods write to this per layer, and it is then actually applied in one consistent method.
	Exts []float32 `` /* 224-byte string literal not displayed */

	// [view: -] context used only for accessing neurons for display -- NetIdxs.NData in here is copied from active context in NewState
	Ctx Context `` /* 134-byte string literal not displayed */

	// [view: -] random number generator for the network -- all random calls must use this -- set seed here for weight initialization values
	Rand erand.SysRand `` /* 139-byte string literal not displayed */

	// random seed to be set at the start of configuring the network and initializing the weights -- set this to get a different set of weights
	RndSeed int64 `` /* 156-byte string literal not displayed */

	// number of threads to use for parallel processing
	NThreads int `desc:"number of threads to use for parallel processing"`

	// [view: inline] GPU implementation
	GPU GPU `view:"inline" desc:"GPU implementation"`

	// [view: -] record function timer information
	RecFunTimes bool `view:"-" desc:"record function timer information"`

	// [view: -] timers for each major function (step of processing)
	FunTimes map[string]*timer.Time `view:"-" desc:"timers for each major function (step of processing)"`
}

NetworkBase manages the basic structural components of a network (layers). The main Network then can just have the algorithm-specific code.

func (*NetworkBase) AddLayer added in v1.4.5 

func (nt *NetworkBase) AddLayer(name string, shape []int, typ LayerTypes) *Layer

AddLayer adds a new layer with given name and shape to the network. 2D and 4D layer shapes are generally preferred but not essential -- see AddLayer2D and 4D for convenience methods for those. 4D layers enable pool (unit-group) level inhibition in Axon networks, for example. shape is in row-major format with outer-most dimensions first: e.g., 4D 3, 2, 4, 5 = 3 rows (Y) of 2 cols (X) of pools, with each unit group having 4 rows (Y) of 5 (X) units.

func (*NetworkBase) AddLayer2D added in v1.4.5 

func (nt *NetworkBase) AddLayer2D(name string, shapeY, shapeX int, typ LayerTypes) *Layer

AddLayer2D adds a new layer with given name and 2D shape to the network. 2D and 4D layer shapes are generally preferred but not essential.

func (*NetworkBase) AddLayer4D added in v1.4.5 

func (nt *NetworkBase) AddLayer4D(name string, nPoolsY, nPoolsX, nNeurY, nNeurX int, typ LayerTypes) *Layer

AddLayer4D adds a new layer with given name and 4D shape to the network. 4D layers enable pool (unit-group) level inhibition in Axon networks, for example. shape is in row-major format with outer-most dimensions first: e.g., 4D 3, 2, 4, 5 = 3 rows (Y) of 2 cols (X) of pools, with each pool having 4 rows (Y) of 5 (X) neurons.

func (*NetworkBase) AddLayerInit added in v1.4.5 

func (nt *NetworkBase) AddLayerInit(ly *Layer, name string, shape []int, typ LayerTypes)

AddLayerInit is implementation routine that takes a given layer and adds it to the network, and initializes and configures it properly.

func (*NetworkBase) AllGlobalVals added in v1.8.0 

func (nt *NetworkBase) AllGlobalVals(ctrKey string, vals map[string]float32)

AllGlobalVals adds to map of all Global variables and values. ctrKey is a key of counters to contextualize values.

func (*NetworkBase) AllGlobals added in v1.8.0 

func (nt *NetworkBase) AllGlobals() string

AllGlobals returns a listing of all Global variables and values.

func (*NetworkBase) AllLayerInhibs added in v1.7.11 

func (nt *NetworkBase) AllLayerInhibs() string

AllLayerInhibs returns a listing of all Layer Inhibition parameters in the Network

func (*NetworkBase) AllParams added in v1.4.5 

func (nt *NetworkBase) AllParams() string

AllParams returns a listing of all parameters in the Network.

func (*NetworkBase) AllPrjnScales added in v1.4.5 

func (nt *NetworkBase) AllPrjnScales() string

AllPrjnScales returns a listing of all PrjnScale parameters in the Network in all Layers, Recv projections. These are among the most important and numerous of parameters (in larger networks) -- this helps keep track of what they all are set to.

func (*NetworkBase) ApplyParams added in v1.4.5 

func (nt *NetworkBase) ApplyParams(pars *params.Sheet, setMsg bool) (bool, error)

ApplyParams applies given parameter style Sheet to layers and prjns in this network. Calls UpdateParams to ensure derived parameters are all updated. If setMsg is true, then a message is printed to confirm each parameter that is set. it always prints a message if a parameter fails to be set. returns true if any params were set, and error if there were any errors.

func (*NetworkBase) AxonLayerByName added in v1.7.12 

func (nt *NetworkBase) AxonLayerByName(name string) *Layer

LayByName returns a layer by looking it up by name in the layer map (nil if not found). Will create the layer map if it is nil or a different size than layers slice, but otherwise needs to be updated manually.

func (*NetworkBase) BidirConnectLayerNames added in v1.4.5 

func (nt *NetworkBase) BidirConnectLayerNames(low, high string, pat prjn.Pattern) (lowlay, highlay *Layer, fwdpj, backpj *Prjn, err error)

BidirConnectLayerNames establishes bidirectional projections between two layers, referenced by name, with low = the lower layer that sends a Forward projection to the high layer, and receives a Back projection in the opposite direction. Returns error if not successful. Does not yet actually connect the units within the layers -- that requires Build.

func (*NetworkBase) BidirConnectLayers added in v1.4.5 

func (nt *NetworkBase) BidirConnectLayers(low, high *Layer, pat prjn.Pattern) (fwdpj, backpj *Prjn)

BidirConnectLayers establishes bidirectional projections between two layers, with low = lower layer that sends a Forward projection to the high layer, and receives a Back projection in the opposite direction. Does not yet actually connect the units within the layers -- that requires Build.

func (*NetworkBase) BidirConnectLayersPy added in v1.4.5 

func (nt *NetworkBase) BidirConnectLayersPy(low, high *Layer, pat prjn.Pattern)

BidirConnectLayersPy establishes bidirectional projections between two layers, with low = lower layer that sends a Forward projection to the high layer, and receives a Back projection in the opposite direction. Does not yet actually connect the units within the layers -- that requires Build. Py = python version with no return vals.

func (*NetworkBase) Bounds added in v1.4.5 

func (nt *NetworkBase) Bounds() (min, max mat32.Vec3)

func (*NetworkBase) BoundsUpdt added in v1.4.5 

func (nt *NetworkBase) BoundsUpdt()

BoundsUpdt updates the Min / Max display bounds for 3D display

func (*NetworkBase) Build added in v1.4.5 

func (nt *NetworkBase) Build(simCtx *Context) error

Build constructs the layer and projection state based on the layer shapes and patterns of interconnectivity. Configures threading using heuristics based on final network size. Must set UseGPUOrder properly prior to calling. Configures the given Context object used in the simulation with the memory access strides for this network -- must be set properly -- see SetCtxStrides.

func (*NetworkBase) BuildGlobals added in v1.8.0 

func (nt *NetworkBase) BuildGlobals(ctx *Context)

BuildGlobals builds Globals vars, using params set in given context

func (*NetworkBase) BuildPrjnGBuf added in v1.7.2 

func (nt *NetworkBase) BuildPrjnGBuf()

BuildPrjnGBuf builds the PrjnGBuf, PrjnGSyns, based on the MaxDelay values in thePrjnParams, which should have been configured by this point. Called by default in InitWts()

func (*NetworkBase) ConnectLayerNames added in v1.4.5 

func (nt *NetworkBase) ConnectLayerNames(send, recv string, pat prjn.Pattern, typ PrjnTypes) (rlay, slay *Layer, pj *Prjn, err error)

ConnectLayerNames establishes a projection between two layers, referenced by name adding to the recv and send projection lists on each side of the connection. Returns error if not successful. Does not yet actually connect the units within the layers -- that requires Build.

func (*NetworkBase) ConnectLayers added in v1.4.5 

func (nt *NetworkBase) ConnectLayers(send, recv *Layer, pat prjn.Pattern, typ PrjnTypes) *Prjn

ConnectLayers establishes a projection between two layers, adding to the recv and send projection lists on each side of the connection. Does not yet actually connect the units within the layers -- that requires Build.

func (*NetworkBase) DeleteAll added in v1.4.5 

func (nt *NetworkBase) DeleteAll()

DeleteAll deletes all layers, prepares network for re-configuring and building

func (*NetworkBase) FunTimerStart added in v1.4.5 

func (nt *NetworkBase) FunTimerStart(fun string)

FunTimerStart starts function timer for given function name -- ensures creation of timer

func (*NetworkBase) FunTimerStop added in v1.4.5 

func (nt *NetworkBase) FunTimerStop(fun string)

FunTimerStop stops function timer -- timer must already exist

func (*NetworkBase) KeyLayerParams added in v1.7.11 

func (nt *NetworkBase) KeyLayerParams() string

KeyLayerParams returns a listing for all layers in the network, of the most important layer-level params (specific to each algorithm).

func (*NetworkBase) KeyPrjnParams added in v1.7.11 

func (nt *NetworkBase) KeyPrjnParams() string

KeyPrjnParams returns a listing for all Recv projections in the network, of the most important projection-level params (specific to each algorithm).

func (*NetworkBase) Label added in v1.4.5 

func (nt *NetworkBase) Label() string

func (*NetworkBase) LateralConnectLayer added in v1.4.5 

func (nt *NetworkBase) LateralConnectLayer(lay *Layer, pat prjn.Pattern) *Prjn

LateralConnectLayer establishes a self-projection within given layer. Does not yet actually connect the units within the layers -- that requires Build.

func (*NetworkBase) LateralConnectLayerPrjn added in v1.4.5 

func (nt *NetworkBase) LateralConnectLayerPrjn(lay *Layer, pat prjn.Pattern, pj *Prjn) *Prjn

LateralConnectLayerPrjn makes lateral self-projection using given projection. Does not yet actually connect the units within the layers -- that requires Build.

func (*NetworkBase) LayByNameTry added in v1.7.12 

func (nt *NetworkBase) LayByNameTry(name string) (*Layer, error)

LayByNameTry returns a layer by looking it up by name -- returns error message if layer is not found

func (*NetworkBase) Layer added in v1.4.5 

func (nt *NetworkBase) Layer(idx int) emer.Layer

func (*NetworkBase) LayerByName added in v1.4.5 

func (nt *NetworkBase) LayerByName(name string) emer.Layer

LayByName returns a layer by looking it up by name in the layer map (nil if not found). Will create the layer map if it is nil or a different size than layers slice, but otherwise needs to be updated manually.

func (*NetworkBase) LayerByNameTry added in v1.4.5 

func (nt *NetworkBase) LayerByNameTry(name string) (emer.Layer, error)

LayerByNameTry returns a layer by looking it up by name -- returns error message if layer is not found

func (*NetworkBase) LayerMapPar added in v1.7.24 

func (nt *NetworkBase) LayerMapPar(fun func(ly *Layer), funame string)

LayerMapPar applies function of given name to all layers using as many go routines as configured in NetThreads.Neurons.

func (*NetworkBase) LayerMapSeq added in v1.6.17 

func (nt *NetworkBase) LayerMapSeq(fun func(ly *Layer), funame string)

LayerMapSeq applies function of given name to all layers sequentially.

func (*NetworkBase) LayerVals added in v1.8.0 

func (nt *NetworkBase) LayerVals(li, di uint32) *LayerVals

LayerVal returns LayerVals for given layer and data parallel indexes

func (*NetworkBase) LayersByClass added in v1.4.5 

func (nt *NetworkBase) LayersByClass(classes ...string) []string

LayersByClass returns a list of layer names by given class(es). Lists are compiled when network Build() function called. The layer Type is always included as a Class, along with any other space-separated strings specified in Class for parameter styling, etc. If no classes are passed, all layer names in order are returned.

func (*NetworkBase) LayersByType added in v1.7.1 

func (nt *NetworkBase) LayersByType(layType ...LayerTypes) []string

LayersByType returns a list of layer names by given layer types. Lists are compiled when network Build() function called. The layer Type is always included as a Class, along with any other space-separated strings specified in Class for parameter styling, etc. If no classes are passed, all layer names in order are returned.

func (*NetworkBase) Layout added in v1.4.5 

func (nt *NetworkBase) Layout()

Layout computes the 3D layout of layers based on their relative position settings

func (*NetworkBase) MakeLayMap added in v1.4.5 

func (nt *NetworkBase) MakeLayMap()

MakeLayMap updates layer map based on current layers

func (*NetworkBase) MaxParallelData added in v1.8.0 

func (nt *NetworkBase) MaxParallelData() int

func (*NetworkBase) NLayers added in v1.4.5 

func (nt *NetworkBase) NLayers() int

func (*NetworkBase) NParallelData added in v1.8.0 

func (nt *NetworkBase) NParallelData() int

func (*NetworkBase) Name added in v1.4.5 

func (nt *NetworkBase) Name() string

emer.Network interface methods:

func (*NetworkBase) NeuronMapPar added in v1.7.24 

func (nt *NetworkBase) NeuronMapPar(ctx *Context, fun func(ly *Layer, ni uint32), funame string)

NeuronMapPar applies function of given name to all neurons using as many go routines as configured in NetThreads.Neurons.

func (*NetworkBase) NeuronMapSeq added in v1.7.24 

func (nt *NetworkBase) NeuronMapSeq(ctx *Context, fun func(ly *Layer, ni uint32), funame string)

NeuronMapSeq applies function of given name to all neurons sequentially.

func (*NetworkBase) NonDefaultParams added in v1.4.5 

func (nt *NetworkBase) NonDefaultParams() string

NonDefaultParams returns a listing of all parameters in the Network that are not at their default values -- useful for setting param styles etc.

func (*NetworkBase) OpenWtsCpp added in v1.4.5 

func (nt *NetworkBase) OpenWtsCpp(filename gi.FileName) error

OpenWtsCpp opens network weights (and any other state that adapts with learning) from old C++ emergent format. If filename has .gz extension, then file is gzip uncompressed.

func (*NetworkBase) OpenWtsJSON added in v1.4.5 

func (nt *NetworkBase) OpenWtsJSON(filename gi.FileName) error

OpenWtsJSON opens network weights (and any other state that adapts with learning) from a JSON-formatted file. If filename has .gz extension, then file is gzip uncompressed.

func (*NetworkBase) ParamsApplied added in v1.7.11 

func (nt *NetworkBase) ParamsApplied(sel *params.Sel)

ParamsApplied is just to satisfy History interface so reset can be applied

func (*NetworkBase) ParamsHistoryReset added in v1.7.11 

func (nt *NetworkBase) ParamsHistoryReset()

ParamsHistoryReset resets parameter application history

func (*NetworkBase) PrjnMapSeq added in v1.6.17 

func (nt *NetworkBase) PrjnMapSeq(fun func(pj *Prjn), funame string)

PrjnMapSeq applies function of given name to all projections sequentially.

func (*NetworkBase) ReadWtsCpp added in v1.4.5 

func (nt *NetworkBase) ReadWtsCpp(r io.Reader) error

ReadWtsCpp reads the weights from old C++ emergent format. Reads entire file into a temporary weights.Weights structure that is then passed to Layers etc using SetWts method.

func (*NetworkBase) ReadWtsJSON added in v1.4.5 

func (nt *NetworkBase) ReadWtsJSON(r io.Reader) error

ReadWtsJSON reads network weights from the receiver-side perspective in a JSON text format. Reads entire file into a temporary weights.Weights structure that is then passed to Layers etc using SetWts method.

func (*NetworkBase) ResetRndSeed added in v1.7.12 

func (nt *NetworkBase) ResetRndSeed()

ResetRndSeed sets random seed to saved RndSeed, ensuring that the network-specific random seed generator has been created.

func (*NetworkBase) SaveAllLayerInhibs added in v1.8.4 

func (nt *NetworkBase) SaveAllLayerInhibs(filename gi.FileName) error

SaveAllLayerInhibs saves list of all layer Inhibition parameters to given file

func (*NetworkBase) SaveAllParams added in v1.8.4 

func (nt *NetworkBase) SaveAllParams(filename gi.FileName) error

SaveAllParams saves list of all parameters in Network to given file.

func (*NetworkBase) SaveAllPrjnScales added in v1.8.4 

func (nt *NetworkBase) SaveAllPrjnScales(filename gi.FileName) error

SavePrjnScales saves a listing of all PrjnScale parameters in the Network in all Layers, Recv projections. These are among the most important and numerous of parameters (in larger networks) -- this helps keep track of what they all are set to.

func (*NetworkBase) SaveNonDefaultParams added in v1.8.4 

func (nt *NetworkBase) SaveNonDefaultParams(filename gi.FileName) error

SaveNonDefaultParams saves list of all non-default parameters in Network to given file.

func (*NetworkBase) SaveParamsSnapshot added in v1.8.4 

func (nt *NetworkBase) SaveParamsSnapshot(pars *netparams.Sets, cfg any, good bool) error

SaveParamsSnapshot saves various views of current parameters to either `params_good` if good = true (for current good reference params) or `params_2006_01_02` (year, month, day) datestamp, providing a snapshot of the simulation params for easy diffs and later reference. Also saves current Config and Params state.

func (*NetworkBase) SaveWtsJSON added in v1.4.5 

func (nt *NetworkBase) SaveWtsJSON(filename gi.FileName) error

SaveWtsJSON saves network weights (and any other state that adapts with learning) to a JSON-formatted file. If filename has .gz extension, then file is gzip compressed.

func (*NetworkBase) SetCtxStrides added in v1.8.0 

func (nt *NetworkBase) SetCtxStrides(simCtx *Context)

SetCtxStrides sets the given simulation context strides for accessing variables on this network -- these must be set properly before calling any compute methods with the context.

func (*NetworkBase) SetMaxData added in v1.8.0 

func (nt *NetworkBase) SetMaxData(simCtx *Context, maxData int)

SetMaxData sets the MaxData and current NData for both the Network and the Context

func (*NetworkBase) SetNThreads added in v1.7.24 

func (nt *NetworkBase) SetNThreads(nthr int)

SetNThreads sets number of threads to use for CPU parallel processing. pass 0 to use a default heuristic number based on current GOMAXPROCS processors and the number of neurons in the network (call after building)

func (*NetworkBase) SetRndSeed added in v1.7.12 

func (nt *NetworkBase) SetRndSeed(seed int64)

SetRndSeed sets random seed and calls ResetRndSeed

func (*NetworkBase) SetWts added in v1.4.5 

func (nt *NetworkBase) SetWts(nw *weights.Network) error

SetWts sets the weights for this network from weights.Network decoded values

func (*NetworkBase) StdVertLayout added in v1.4.5 

func (nt *NetworkBase) StdVertLayout()

StdVertLayout arranges layers in a standard vertical (z axis stack) layout, by setting the Rel settings

func (*NetworkBase) SynVarNames added in v1.8.0 

func (nt *NetworkBase) SynVarNames() []string

SynVarNames returns the names of all the variables on the synapses in this network. Not all projections need to support all variables, but must safely return 0's for unsupported ones. The order of this list determines NetView variable display order. This is typically a global list so do not modify!

func (*NetworkBase) SynVarProps added in v1.8.0 

func (nt *NetworkBase) SynVarProps() map[string]string

SynVarProps returns properties for variables

func (*NetworkBase) TimerReport added in v1.4.5 

func (nt *NetworkBase) TimerReport()

TimerReport reports the amount of time spent in each function, and in each thread

func (*NetworkBase) UnitVarNames added in v1.8.0 

func (nt *NetworkBase) UnitVarNames() []string

UnitVarNames returns a list of variable names available on the units in this network. Not all layers need to support all variables, but must safely return 0's for unsupported ones. The order of this list determines NetView variable display order. This is typically a global list so do not modify!

func (*NetworkBase) UnitVarProps added in v1.8.0 

func (nt *NetworkBase) UnitVarProps() map[string]string

UnitVarProps returns properties for variables

func (*NetworkBase) VarRange added in v1.4.5 

func (nt *NetworkBase) VarRange(varNm string) (min, max float32, err error)

VarRange returns the min / max values for given variable todo: support r. s. projection values

func (*NetworkBase) WriteWtsJSON added in v1.4.5 

func (nt *NetworkBase) WriteWtsJSON(w io.Writer) error

WriteWtsJSON writes the weights from this layer from the receiver-side perspective in a JSON text format. We build in the indentation logic to make it much faster and more efficient.

type NeuroModParams added in v1.7.0 

type NeuroModParams struct {

	// dopamine receptor-based effects of dopamine modulation on excitatory and inhibitory conductances: D1 is excitatory, D2 is inhibitory as a function of increasing dopamine
	DAMod DAModTypes `` /* 176-byte string literal not displayed */

	// valence coding of this layer -- may affect specific layer types but does not directly affect neuromodulators currently
	Valence ValenceTypes `desc:"valence coding of this layer -- may affect specific layer types but does not directly affect neuromodulators currently"`

	// [viewif: DAMod!=NoDAMod] multiplicative factor on overall DA modulation specified by DAMod -- resulting overall gain factor is: 1 + DAModGain * DA, where DA is appropriate DA-driven factor
	DAModGain float32 `` /* 194-byte string literal not displayed */

	// modulate the sign of the learning rate factor according to the DA sign, taking into account the DAMod sign reversal for D2Mod, also using BurstGain and DipGain to modulate DA value -- otherwise, only the magnitude of the learning rate is modulated as a function of raw DA magnitude according to DALRateMod (without additional gain factors)
	DALRateSign slbool.Bool `` /* 346-byte string literal not displayed */

	// [viewif: !DALRateSign] [min: 0] [max: 1] if not using DALRateSign, this is the proportion of maximum learning rate that Abs(DA) magnitude can modulate -- e.g., if 0.2, then DA = 0 = 80% of std learning rate, 1 = 100%
	DALRateMod float32 `` /* 220-byte string literal not displayed */

	// [min: 0] [max: 1] proportion of maximum learning rate that ACh can modulate -- e.g., if 0.2, then ACh = 0 = 80% of std learning rate, 1 = 100%
	AChLRateMod float32 `` /* 147-byte string literal not displayed */

	// [def: 0,5] [min: 0] amount of extra Gi inhibition added in proportion to 1 - ACh level -- makes ACh disinhibitory
	AChDisInhib float32 `min:"0" def:"0,5" desc:"amount of extra Gi inhibition added in proportion to 1 - ACh level -- makes ACh disinhibitory"`

	// [def: 1] [min: 0] multiplicative gain factor applied to positive dopamine signals -- this operates on the raw dopamine signal prior to any effect of D2 receptors in reversing its sign!
	BurstGain float32 `` /* 189-byte string literal not displayed */

	// [def: 1] [min: 0] multiplicative gain factor applied to negative dopamine signals -- this operates on the raw dopamine signal prior to any effect of D2 receptors in reversing its sign! should be small for acq, but roughly equal to burst for ext
	DipGain float32 `` /* 249-byte string literal not displayed */
	// contains filtered or unexported fields
}

NeuroModParams specifies the effects of neuromodulators on neural activity and learning rate. These can apply to any neuron type, and are applied in the core cycle update equations.

func (*NeuroModParams) DAGain added in v1.7.11 

func (nm *NeuroModParams) DAGain(da float32) float32

DAGain returns DA dopamine value with Burst / Dip Gain factors applied

func (*NeuroModParams) DASign added in v1.7.16 

func (nm *NeuroModParams) DASign() float32

DASign returns the sign of dopamine effects: D2Mod = -1, else 1

func (*NeuroModParams) Defaults added in v1.7.0 

func (nm *NeuroModParams) Defaults()

func (*NeuroModParams) GGain added in v1.7.0 

func (nm *NeuroModParams) GGain(da float32) float32

GGain returns effective Ge and Gi gain factor given dopamine (DA) +/- burst / dip value (0 = tonic level). factor is 1 for no modulation, otherwise higher or lower.

func (*NeuroModParams) GiFmACh added in v1.7.0 

func (nm *NeuroModParams) GiFmACh(ach float32) float32

GIFmACh returns amount of extra inhibition to add based on disinhibitory effects of ACh -- no inhibition when ACh = 1, extra when < 1.

func (*NeuroModParams) IsBLAExt added in v1.7.18 

func (nm *NeuroModParams) IsBLAExt() bool

IsBLAExt returns true if this is Positive, D2 or Negative D1 -- BLA extinction

func (*NeuroModParams) LRMod added in v1.7.0 

func (nm *NeuroModParams) LRMod(da, ach float32) float32

LRMod returns overall learning rate modulation factor due to neuromodulation from given dopamine (DA) and ACh inputs. If DALRateMod is true and DAMod == D1Mod or D2Mod, then the sign is a function of the DA

func (*NeuroModParams) LRModFact added in v1.7.0 

func (nm *NeuroModParams) LRModFact(pct, val float32) float32

LRModFact returns learning rate modulation factor for given inputs.

func (*NeuroModParams) Update added in v1.7.0 

func (nm *NeuroModParams) Update()

type NeuronAvgVarStrides added in v1.8.0 

type NeuronAvgVarStrides struct {

	// neuron level
	Neuron uint32 `desc:"neuron level"`

	// variable level
	Var uint32 `desc:"variable level"`
	// contains filtered or unexported fields
}

NeuronAvgVarStrides encodes the stride offsets for neuron variable access into network float32 array. Data is always the inner-most variable.

func (*NeuronAvgVarStrides) Idx added in v1.8.0 

func (ns *NeuronAvgVarStrides) Idx(neurIdx uint32, nvar NeuronAvgVars) uint32

Idx returns the index into network float32 array for given neuron and variable

func (*NeuronAvgVarStrides) SetNeuronOuter added in v1.8.0 

func (ns *NeuronAvgVarStrides) SetNeuronOuter()

SetNeuronOuter sets strides with neurons as outer loop: [Neurons][Vars], which is optimal for CPU-based computation.

func (*NeuronAvgVarStrides) SetVarOuter added in v1.8.0 

func (ns *NeuronAvgVarStrides) SetVarOuter(nneur int)

SetVarOuter sets strides with vars as outer loop: [Vars][Neurons], which is optimal for GPU-based computation.

type NeuronAvgVars added in v1.8.0 

type NeuronAvgVars int32

NeuronAvgVars are mostly neuron variables involved in longer-term average activity which is aggregated over time and not specific to each input data state, along with any other state that is not input data specific. 

const (
	// ActAvg is average activation (of minus phase activation state) over long time intervals (time constant = Dt.LongAvgTau) -- useful for finding hog units and seeing overall distribution of activation
	ActAvg NeuronAvgVars = iota

	// AvgPct is ActAvg as a proportion of overall layer activation -- this is used for synaptic scaling to match TrgAvg activation -- updated at SlowInterval intervals
	AvgPct

	// TrgAvg is neuron's target average activation as a proportion of overall layer activation, assigned during weight initialization, driving synaptic scaling relative to AvgPct
	TrgAvg

	// DTrgAvg is change in neuron's target average activation as a result of unit-wise error gradient -- acts like a bias weight.  MPI needs to share these across processors.
	DTrgAvg

	// AvgDif is AvgPct - TrgAvg -- i.e., the error in overall activity level relative to set point for this neuron, which drives synaptic scaling -- updated at SlowInterval intervals
	AvgDif

	// GeBase is baseline level of Ge, added to GeRaw, for intrinsic excitability
	GeBase

	// GiBase is baseline level of Gi, added to GiRaw, for intrinsic excitability
	GiBase

	NeuronAvgVarsN
)

func (NeuronAvgVars) Desc added in v1.8.14 

func (i NeuronAvgVars) Desc() string

func (*NeuronAvgVars) FromString added in v1.8.0 

func (i *NeuronAvgVars) FromString(s string) error

func (NeuronAvgVars) MarshalJSON added in v1.8.0 

func (ev NeuronAvgVars) MarshalJSON() ([]byte, error)

func (NeuronAvgVars) String added in v1.8.0 

func (i NeuronAvgVars) String() string

func (*NeuronAvgVars) UnmarshalJSON added in v1.8.0 

func (ev *NeuronAvgVars) UnmarshalJSON(b []byte) error

type NeuronFlags added in v1.6.4 

type NeuronFlags int32

NeuronFlags are bit-flags encoding relevant binary state for neurons 

const (
	// NeuronOff flag indicates that this neuron has been turned off (i.e., lesioned)
	NeuronOff NeuronFlags = 1

	// NeuronHasExt means the neuron has external input in its Ext field
	NeuronHasExt NeuronFlags = 2

	// NeuronHasTarg means the neuron has external target input in its Target field
	NeuronHasTarg NeuronFlags = 4

	// NeuronHasCmpr means the neuron has external comparison input in its Target field -- used for computing
	// comparison statistics but does not drive neural activity ever
	NeuronHasCmpr NeuronFlags = 8
)

The neuron flags

func (NeuronFlags) Desc added in v1.8.14 

func (i NeuronFlags) Desc() string

func (NeuronFlags) String added in v1.6.4 

func (i NeuronFlags) String() string

type NeuronIdxStrides added in v1.8.0 

type NeuronIdxStrides struct {

	// neuron level
	Neuron uint32 `desc:"neuron level"`

	// index value level
	Index uint32 `desc:"index value level"`
	// contains filtered or unexported fields
}

NeuronIdxStrides encodes the stride offsets for neuron index access into network uint32 array.

func (*NeuronIdxStrides) Idx added in v1.8.0 

func (ns *NeuronIdxStrides) Idx(neurIdx uint32, idx NeuronIdxs) uint32

Idx returns the index into network uint32 array for given neuron, index value

func (*NeuronIdxStrides) SetIdxOuter added in v1.8.0 

func (ns *NeuronIdxStrides) SetIdxOuter(nneur int)

SetIdxOuter sets strides with indexes as outer dimension: [Idxs][Neurons] (outer to inner), which is optimal for GPU-based computation.

func (*NeuronIdxStrides) SetNeuronOuter added in v1.8.0 

func (ns *NeuronIdxStrides) SetNeuronOuter()

SetNeuronOuter sets strides with neurons as outer dimension: [Neurons[[Idxs] (outer to inner), which is optimal for CPU-based computation.

type NeuronIdxs added in v1.8.0 

type NeuronIdxs int32

NeuronIdxs are the neuron indexes and other uint32 values. There is only one of these per neuron -- not data parallel. note: Flags are encoded in Vars because they are data parallel and writable, whereas indexes are read-only. 

const (
	// NrnNeurIdx is the index of this neuron within its owning layer
	NrnNeurIdx NeuronIdxs = iota

	// NrnLayIdx is the index of the layer that this neuron belongs to,
	// needed for neuron-level parallel code.
	NrnLayIdx

	// NrnSubPool is the index of the sub-level inhibitory pool for this neuron
	// (only for 4D shapes, the pool (unit-group / hypercolumn) structure level).
	// Indicies start at 1 -- 0 is layer-level pool (is 0 if no sub-pools).
	NrnSubPool

	NeuronIdxsN
)

func (NeuronIdxs) Desc added in v1.8.14 

func (i NeuronIdxs) Desc() string

func (*NeuronIdxs) FromString added in v1.8.0 

func (i *NeuronIdxs) FromString(s string) error

func (NeuronIdxs) MarshalJSON added in v1.8.0 

func (ev NeuronIdxs) MarshalJSON() ([]byte, error)

func (NeuronIdxs) String added in v1.8.0 

func (i NeuronIdxs) String() string

func (*NeuronIdxs) UnmarshalJSON added in v1.8.0 

func (ev *NeuronIdxs) UnmarshalJSON(b []byte) error

type NeuronVarStrides added in v1.8.0 

type NeuronVarStrides struct {

	// neuron level
	Neuron uint32 `desc:"neuron level"`

	// variable level
	Var uint32 `desc:"variable level"`
	// contains filtered or unexported fields
}

NeuronVarStrides encodes the stride offsets for neuron variable access into network float32 array. Data is always the inner-most variable.

func (*NeuronVarStrides) Idx added in v1.8.0 

func (ns *NeuronVarStrides) Idx(neurIdx, di uint32, nvar NeuronVars) uint32

Idx returns the index into network float32 array for given neuron, data, and variable

func (*NeuronVarStrides) SetNeuronOuter added in v1.8.0 

func (ns *NeuronVarStrides) SetNeuronOuter(ndata int)

SetNeuronOuter sets strides with neurons as outer loop: [Neurons][Vars][Data], which is optimal for CPU-based computation.

func (*NeuronVarStrides) SetVarOuter added in v1.8.0 

func (ns *NeuronVarStrides) SetVarOuter(nneur, ndata int)

SetVarOuter sets strides with vars as outer loop: [Vars][Neurons][Data], which is optimal for GPU-based computation.

type NeuronVars 

type NeuronVars int32

NeuronVars are the neuron variables representing current active state, specific to each input data state. See NeuronAvgVars for vars shared across data. 

const (

	// Spike is whether neuron has spiked or not on this cycle (0 or 1)
	Spike NeuronVars = iota

	// Spiked is 1 if neuron has spiked within the last 10 cycles (msecs), corresponding to a nominal max spiking rate of 100 Hz, 0 otherwise -- useful for visualization and computing activity levels in terms of average spiked levels.
	Spiked

	// Act is rate-coded activation value reflecting instantaneous estimated rate of spiking, based on 1 / ISIAvg.  This drives feedback inhibition in the FFFB function (todo: this will change when better inhibition is implemented), and is integrated over time for ActInt which is then used for performance statistics and layer average activations, etc.  Should not be used for learning or other computations.
	Act

	// ActInt is integrated running-average activation value computed from Act with time constant Act.Dt.IntTau, to produce a longer-term integrated value reflecting the overall activation state across the ThetaCycle time scale, as the overall response of network to current input state -- this is copied to ActM and ActP at the ends of the minus and plus phases, respectively, and used in computing performance-level statistics (which are typically based on ActM).  Should not be used for learning or other computations.
	ActInt

	// ActM is ActInt activation state at end of third quarter, representing the posterior-cortical minus phase activation -- used for statistics and monitoring network performance. Should not be used for learning or other computations.
	ActM

	// ActP is ActInt activation state at end of fourth quarter, representing the posterior-cortical plus_phase activation -- used for statistics and monitoring network performance.  Should not be used for learning or other computations.
	ActP

	// Ext is external input: drives activation of unit from outside influences (e.g., sensory input)
	Ext

	// Target is the target value: drives learning to produce this activation value
	Target

	// Ge is total excitatory conductance, including all forms of excitation (e.g., NMDA) -- does *not* include Gbar.E
	Ge

	// Gi is total inhibitory synaptic conductance -- the net inhibitory input to the neuron -- does *not* include Gbar.I
	Gi

	// Gk is total potassium conductance, typically reflecting sodium-gated potassium currents involved in adaptation effects -- does *not* include Gbar.K
	Gk

	// Inet is net current produced by all channels -- drives update of Vm
	Inet

	// Vm is membrane potential -- integrates Inet current over time
	Vm

	// VmDend is dendritic membrane potential -- has a slower time constant, is not subject to the VmR reset after spiking
	VmDend

	// ISI is current inter-spike-interval -- counts up since last spike.  Starts at -1 when initialized.
	ISI

	// ISIAvg is average inter-spike-interval -- average time interval between spikes, integrated with ISITau rate constant (relatively fast) to capture something close to an instantaneous spiking rate.  Starts at -1 when initialized, and goes to -2 after first spike, and is only valid after the second spike post-initialization.
	ISIAvg

	// CaSpkP is continuous cascaded integration of CaSpkM at PTau time constant (typically 40), representing neuron-level purely spiking version of plus, LTP direction of weight change and capturing the function of CaMKII in the Kinase learning rule. Used for specialized learning and computational functions, statistics, instead of Act.
	CaSpkP

	// CaSpkD is continuous cascaded integration CaSpkP at DTau time constant (typically 40), representing neuron-level purely spiking version of minus, LTD direction of weight change and capturing the function of DAPK1 in the Kinase learning rule. Used for specialized learning and computational functions, statistics, instead of Act.
	CaSpkD

	// CaSyn is spike-driven calcium trace for synapse-level Ca-driven learning: exponential integration of SpikeG * Spike at SynTau time constant (typically 30).  Synapses integrate send.CaSyn * recv.CaSyn across M, P, D time integrals for the synaptic trace driving credit assignment in learning. Time constant reflects binding time of Glu to NMDA and Ca buffering postsynaptically, and determines time window where pre * post spiking must overlap to drive learning.
	CaSyn

	// CaSpkM is spike-driven calcium trace used as a neuron-level proxy for synpatic credit assignment factor based on continuous time-integrated spiking: exponential integration of SpikeG * Spike at MTau time constant (typically 5).  Simulates a calmodulin (CaM) like signal at the most abstract level.
	CaSpkM

	// CaSpkPM is minus-phase snapshot of the CaSpkP value -- similar to ActM but using a more directly spike-integrated value.
	CaSpkPM

	// CaLrn is recv neuron calcium signal used to drive temporal error difference component of standard learning rule, combining NMDA (NmdaCa) and spiking-driven VGCC (VgccCaInt) calcium sources (vs. CaSpk* which only reflects spiking component).  This is integrated into CaM, CaP, CaD, and temporal derivative is CaP - CaD (CaMKII - DAPK1).  This approximates the backprop error derivative on net input, but VGCC component adds a proportion of recv activation delta as well -- a balance of both works best.  The synaptic-level trace multiplier provides the credit assignment factor, reflecting coincident activity and potentially integrated over longer multi-trial timescales.
	CaLrn

	// NrnCaM is integrated CaLrn at MTau timescale (typically 5), simulating a calmodulin (CaM) like signal, which then drives CaP, CaD for delta signal driving error-driven learning.
	NrnCaM

	// NrnCaP is cascaded integration of CaM at PTau time constant (typically 40), representing the plus, LTP direction of weight change and capturing the function of CaMKII in the Kinase learning rule.
	NrnCaP

	// NrnCaD is cascaded integratoin of CaP at DTau time constant (typically 40), representing the minus, LTD direction of weight change and capturing the function of DAPK1 in the Kinase learning rule.
	NrnCaD

	// CaDiff is difference between CaP - CaD -- this is the error signal that drives error-driven learning.
	CaDiff

	// Attn is Attentional modulation factor, which can be set by special layers such as the TRC -- multiplies Ge
	Attn

	// RLRate is recv-unit based learning rate multiplier, reflecting the sigmoid derivative computed from the CaSpkD of recv unit, and the normalized difference CaSpkP - CaSpkD / MAX(CaSpkP - CaSpkD).
	RLRate

	// SpkMaxCa is Ca integrated like CaSpkP but only starting at MaxCycStart cycle, to prevent inclusion of carryover spiking from prior theta cycle trial -- the PTau time constant otherwise results in significant carryover.  This is the input to SpkMax
	SpkMaxCa

	// SpkMax is maximum CaSpkP across one theta cycle time window (max of SpkMaxCa) -- used for specialized algorithms that have more phasic behavior within a single trial, e.g., BG Matrix layer gating.  Also useful for visualization of peak activity of neurons.
	SpkMax

	// SpkPrv is final CaSpkD activation state at end of previous theta cycle.  used for specialized learning mechanisms that operate on delayed sending activations.
	SpkPrv

	// SpkSt1 is the activation state at specific time point within current state processing window (e.g., 50 msec for beta cycle within standard theta cycle), as saved by SpkSt1() function.  Used for example in hippocampus for CA3, CA1 learning
	SpkSt1

	// SpkSt2 is the activation state at specific time point within current state processing window (e.g., 100 msec for beta cycle within standard theta cycle), as saved by SpkSt2() function.  Used for example in hippocampus for CA3, CA1 learning
	SpkSt2

	// GeNoiseP is accumulating poisson probability factor for driving excitatory noise spiking -- multiply times uniform random deviate at each time step, until it gets below the target threshold based on lambda.
	GeNoiseP

	// GeNoise is integrated noise excitatory conductance, added into Ge
	GeNoise

	// GiNoiseP is accumulating poisson probability factor for driving inhibitory noise spiking -- multiply times uniform random deviate at each time step, until it gets below the target threshold based on lambda.
	GiNoiseP

	// GiNoise is integrated noise inhibotyr conductance, added into Gi
	GiNoise

	// GeExt is extra excitatory conductance added to Ge -- from Ext input, GeCtxt etc
	GeExt

	// GeRaw is raw excitatory conductance (net input) received from senders = current raw spiking drive
	GeRaw

	// GeSyn is time-integrated total excitatory synaptic conductance, with an instantaneous rise time from each spike (in GeRaw) and exponential decay with Dt.GeTau, aggregated over projections -- does *not* include Gbar.E
	GeSyn

	// GiRaw is raw inhibitory conductance (net input) received from senders  = current raw spiking drive
	GiRaw

	// GiSyn is time-integrated total inhibitory synaptic conductance, with an instantaneous rise time from each spike (in GiRaw) and exponential decay with Dt.GiTau, aggregated over projections -- does *not* include Gbar.I.  This is added with computed FFFB inhibition to get the full inhibition in Gi
	GiSyn

	// GeInt is integrated running-average activation value computed from Ge with time constant Act.Dt.IntTau, to produce a longer-term integrated value reflecting the overall Ge level across the ThetaCycle time scale (Ge itself fluctuates considerably) -- useful for stats to set strength of connections etc to get neurons into right range of overall excitatory drive
	GeInt

	// GeIntNorm is normalized GeInt value (divided by the layer maximum) -- this is used for learning in layers that require learning on subthreshold activity
	GeIntNorm

	// GiInt is integrated running-average activation value computed from GiSyn with time constant Act.Dt.IntTau, to produce a longer-term integrated value reflecting the overall synaptic Gi level across the ThetaCycle time scale (Gi itself fluctuates considerably) -- useful for stats to set strength of connections etc to get neurons into right range of overall inhibitory drive
	GiInt

	// GModRaw is raw modulatory conductance, received from GType = ModulatoryG projections
	GModRaw

	// GModSyn is syn integrated modulatory conductance, received from GType = ModulatoryG projections
	GModSyn

	// GMaintRaw is raw maintenance conductance, received from GType = MaintG projections
	GMaintRaw

	// GMaintSyn is syn integrated maintenance conductance, integrated using MaintNMDA params.
	GMaintSyn

	// SSGi is SST+ somatostatin positive slow spiking inhibition
	SSGi

	// SSGiDend is amount of SST+ somatostatin positive slow spiking inhibition applied to dendritic Vm (VmDend)
	SSGiDend

	// Gak is conductance of A-type K potassium channels
	Gak

	// MahpN is accumulating voltage-gated gating value for the medium time scale AHP
	MahpN

	// SahpCa is slowly accumulating calcium value that drives the slow AHP
	SahpCa

	// SahpN is sAHP gating value
	SahpN

	// GknaMed is conductance of sodium-gated potassium channel (KNa) medium dynamics (Slick) -- produces accommodation / adaptation of firing
	GknaMed

	// GknaSlow is conductance of sodium-gated potassium channel (KNa) slow dynamics (Slack) -- produces accommodation / adaptation of firing
	GknaSlow

	// GnmdaSyn is integrated NMDA recv synaptic current -- adds GeRaw and decays with time constant
	GnmdaSyn

	// Gnmda is net postsynaptic (recv) NMDA conductance, after Mg V-gating and Gbar -- added directly to Ge as it has the same reversal potential
	Gnmda

	// GnmdaMaint is net postsynaptic maintenance NMDA conductance, computed from GMaintSyn and GMaintRaw, after Mg V-gating and Gbar -- added directly to Ge as it has the same reversal potential
	GnmdaMaint

	// GnmdaLrn is learning version of integrated NMDA recv synaptic current -- adds GeRaw and decays with time constant -- drives NmdaCa that then drives CaM for learning
	GnmdaLrn

	// NmdaCa is NMDA calcium computed from GnmdaLrn, drives learning via CaM
	NmdaCa

	// GgabaB is net GABA-B conductance, after Vm gating and Gbar + Gbase -- applies to Gk, not Gi, for GIRK, with .1 reversal potential.
	GgabaB

	// GABAB is GABA-B / GIRK activation -- time-integrated value with rise and decay time constants
	GABAB

	// GABABx is GABA-B / GIRK internal drive variable -- gets the raw activation and decays
	GABABx

	// Gvgcc is conductance (via Ca) for VGCC voltage gated calcium channels
	Gvgcc

	// VgccM is activation gate of VGCC channels
	VgccM

	// VgccH inactivation gate of VGCC channels
	VgccH

	// VgccCa is instantaneous VGCC calcium flux -- can be driven by spiking or directly from Gvgcc
	VgccCa

	// VgccCaInt time-integrated VGCC calcium flux -- this is actually what drives learning
	VgccCaInt

	// SKCaIn is intracellular calcium store level, available to be released with spiking as SKCaR, which can bind to SKCa receptors and drive K current. replenishment is a function of spiking activity being below a threshold
	SKCaIn

	// SKCaR released amount of intracellular calcium, from SKCaIn, as a function of spiking events.  this can bind to SKCa channels and drive K currents.
	SKCaR

	// SKCaM is Calcium-gated potassium channel gating factor, driven by SKCaR via a Hill equation as in chans.SKPCaParams.
	SKCaM

	// Gsk is Calcium-gated potassium channel conductance as a function of Gbar * SKCaM.
	Gsk

	// Burst is 5IB bursting activation value, computed by thresholding regular CaSpkP value in Super superficial layers
	Burst

	// BurstPrv is previous Burst bursting activation from prior time step -- used for context-based learning
	BurstPrv

	// CtxtGe is context (temporally delayed) excitatory conductance, driven by deep bursting at end of the plus phase, for CT layers.
	CtxtGe

	// CtxtGeRaw is raw update of context (temporally delayed) excitatory conductance, driven by deep bursting at end of the plus phase, for CT layers.
	CtxtGeRaw

	// CtxtGeOrig is original CtxtGe value prior to any decay factor -- updates at end of plus phase.
	CtxtGeOrig

	// NrnFlags are bit flags for binary state variables, which are converted to / from uint32.
	// These need to be in Vars because they can be differential per data (for ext inputs)
	// and are writable (indexes are read only).
	NrnFlags

	NeuronVarsN
)

func (NeuronVars) Desc added in v1.8.14 

func (i NeuronVars) Desc() string

func (*NeuronVars) FromString added in v1.8.0 

func (i *NeuronVars) FromString(s string) error

func (NeuronVars) MarshalJSON added in v1.8.0 

func (ev NeuronVars) MarshalJSON() ([]byte, error)

func (NeuronVars) String added in v1.8.0 

func (i NeuronVars) String() string

func (*NeuronVars) UnmarshalJSON added in v1.8.0 

func (ev *NeuronVars) UnmarshalJSON(b []byte) error

type PVLV added in v1.7.11 

type PVLV struct {

	// number of possible positive US states and corresponding drives -- the first is always reserved for novelty / curiosity.  Must be set programmatically via SetNUSs method, which allocates corresponding parameters.
	NPosUSs uint32 `` /* 231-byte string literal not displayed */

	// number of possible negative US states -- [0] is reserved for accumulated time, [1] the accumulated effort cost.  Must be set programmatically via SetNUSs method, which allocates corresponding parameters.
	NNegUSs uint32 `` /* 223-byte string literal not displayed */

	// parameters and state for built-in drives that form the core motivations of agent, controlled by lateral hypothalamus and associated body state monitoring such as glucose levels and thirst.
	Drive DriveParams `` /* 195-byte string literal not displayed */

	// [view: inline] urgency (increasing pressure to do something) and parameters for updating it. Raw urgency is incremented by same units as effort, but is only reset with a positive US.
	Urgency UrgencyParams `` /* 188-byte string literal not displayed */

	// controls how positive and negative USs are weighted and integrated to compute an overall PV primary value.
	USs USParams `desc:"controls how positive and negative USs are weighted and integrated to compute an overall PV primary value."`

	// [view: inline] lateral habenula (LHb) parameters and state, which drives dipping / pausing in dopamine when the predicted positive outcome > actual, or actual negative outcome > predicted.  Can also drive bursting for the converse, and via matrix phasic firing
	LHb LHbParams `` /* 266-byte string literal not displayed */

	// parameters for giving up based on PV pos - neg difference
	GiveUp GiveUpParams `desc:"parameters for giving up based on PV pos - neg difference"`
}

PVLV represents the core brainstem-level (hypothalamus) bodily drives and resulting dopamine from US (unconditioned stimulus) inputs, as computed by the PVLV model of primary value (PV) and learned value (LV), describing the functions of the Amygala, Ventral Striatum, VTA and associated midbrain nuclei (LDT, LHb, RMTg). Core LHb (lateral habenula) and VTA (ventral tegmental area) dopamine are computed in equations using inputs from specialized network layers (LDTLayer driven by BLA, CeM layers, VSPatchLayer). The Drives, Effort, US and resulting LHb PV dopamine computation all happens at the at the start of each trial (NewState, Step). The LV / CS dopamine is computed cycle-by-cycle by the VTA layer using parameters set by the VTA layer. Renders USLayer, PVLayer, DrivesLayer representations based on state updated here.

func (*PVLV) AddTimeEffort added in v1.8.16 

func (pp *PVLV) AddTimeEffort(ctx *Context, di uint32, effort float32)

AddTimeEffort adds a unit of time and an increment of effort

func (*PVLV) DAFmPVs added in v1.8.17 

func (pp *PVLV) DAFmPVs(pvPos, pvNeg, vsPatchPos float32) (burst, dip, da, rew float32)

DAFmPVs computes the overall PV DA in terms of LHb burst and dip activity from given pvPos, pvNeg, and vsPatchPos values. Also returns the net "reward" value as the discounted PV value, separate from the vsPatchPos prediction error factor.

func (*PVLV) Defaults added in v1.7.11 

func (pp *PVLV) Defaults()

func (*PVLV) DriveUpdt added in v1.7.11 

func (pp *PVLV) DriveUpdt(ctx *Context, di uint32)

DriveUpdt is used when auto-updating drive levels based on US consumption, which partially satisfies (decrements) corresponding drive, and on time passing, where drives adapt to their overall baseline levels.

func (*PVLV) EffortUrgencyUpdt added in v1.7.18 

func (pp *PVLV) EffortUrgencyUpdt(ctx *Context, di uint32, effort float32)

EffortUrgencyUpdt updates the Effort or Urgency based on given effort increment. Effort is incremented when VSMatrixHasGated (i.e., goal engaged) and Urgency updates otherwise (when not goal engaged) Call this at the start of the trial, in ApplyPVLV method, after NewState.

func (*PVLV) GiveUpFmPV added in v1.8.16 

func (pp *PVLV) GiveUpFmPV(ctx *Context, di uint32, pvNeg float32, rnd erand.Rand) bool

GiveUpFmPV determines whether to give up on current goal based on balance between estimated PVpos and accumulated PVneg. returns true if give up triggered.

func (*PVLV) HasPosUS added in v1.7.11 

func (pp *PVLV) HasPosUS(ctx *Context, di uint32) bool

HasPosUS returns true if there is at least one non-zero positive US

func (*PVLV) InitDrives added in v1.7.11 

func (pp *PVLV) InitDrives(ctx *Context, di uint32)

InitDrives initializes all the Drives to baseline values (default = 0)

func (*PVLV) InitUS added in v1.7.11 

func (pp *PVLV) InitUS(ctx *Context, di uint32)

InitUS initializes all the USs to zero

func (*PVLV) NewState added in v1.7.18 

func (pp *PVLV) NewState(ctx *Context, di uint32, rnd erand.Rand)

NewState is called at very start of new state (trial) of processing. sets HadRew = HasRew from last trial -- used to then reset various things after reward.

func (*PVLV) PVDA added in v1.8.16 

func (pp *PVLV) PVDA(ctx *Context, di uint32, rnd erand.Rand)

PVDA computes the PV (primary value) based dopamine based on current state information, at the start of a trial. PV DA is computed by the VS (ventral striatum) and the LHb / RMTg, and the resulting values are stored in global variables. Called after updating USs, Effort, Drives at start of trial step, in Step.

func (*PVLV) PVneg added in v1.8.16 

func (pp *PVLV) PVneg(ctx *Context, di uint32) (pvNegSum, pvNeg float32)

PVneg returns the summed weighted negative value of current negative US state, where each US is multiplied by a weighting factor and summed (usNegSum) and the normalized version of this sum (PVneg = overall negative PV) as 1 / (1 + (PVnegGain * PVnegSum))

func (*PVLV) PVnegEstFmUSs added in v1.8.17 

func (pp *PVLV) PVnegEstFmUSs(uss []float32) (pvNegSum, pvNeg float32)

PVnegEstFmUSs returns the estimated negative PV value based on given externally-provided US values. This can be used to compute estimates to compare network performance.

func (*PVLV) PVpos added in v1.8.16 

func (pp *PVLV) PVpos(ctx *Context, di uint32) (pvPosSum, pvPos float32)

PVpos returns the summed weighted positive value of current positive US state, where each US is multiplied by its current drive and weighting factor (pvPosSum), and the normalized version of this sum (PVpos = overall positive PV) as 1 / (1 + (PVposGain * pvPosSum))

func (*PVLV) PVposEst added in v1.8.16 

func (pp *PVLV) PVposEst(ctx *Context, di uint32) (pvPosSum, pvPos float32)

PVposEst returns the estimated positive PV value based on drives and OFCposUSPT maint and VSMatrix gating

func (*PVLV) PVposEstFmUSs added in v1.8.16 

func (pp *PVLV) PVposEstFmUSs(ctx *Context, di uint32, uss []float32) (pvPosSum, pvPos float32)

PVposEstFmUSs returns the estimated positive PV value based on drives and given US values. This can be used to compute estimates to compare network performance.

func (*PVLV) PVposEstFmUSsDrives added in v1.8.17 

func (pp *PVLV) PVposEstFmUSsDrives(uss, drives []float32) (pvPosSum, pvPos float32)

PVposEstFmUSsDrives returns the estimated positive PV value based on given externally-provided drives and US values. This can be used to compute estimates to compare network performance.

func (*PVLV) PVsFmUSs added in v1.8.16 

func (pp *PVLV) PVsFmUSs(ctx *Context, di uint32)

PVsFmUSs updates the current PV summed, weighted, normalized values from the underlying US values.

func (*PVLV) Reset added in v1.7.18 

func (pp *PVLV) Reset(ctx *Context, di uint32)

Reset resets all PVLV state

func (*PVLV) ResetGiveUp added in v1.8.16 

func (pp *PVLV) ResetGiveUp(ctx *Context, di uint32)

ResetGiveUp resets all the give-up related global values.

func (*PVLV) ResetGoalState added in v1.8.16 

func (pp *PVLV) ResetGoalState(ctx *Context, di uint32)

ResetGoalState resets all the goal-engaged global values. Critically, this is only called after goal accomplishment, not after goal gating -- prevents "shortcutting" by re-gating.

func (*PVLV) SetDrive added in v1.7.11 

func (pp *PVLV) SetDrive(ctx *Context, di uint32, dr uint32, val float32)

PVLVSetDrive sets given Drive to given value

func (*PVLV) SetDrives added in v1.8.16 

func (pp *PVLV) SetDrives(ctx *Context, di uint32, curiosity float32, drives ...float32)

SetDrives is used when directly controlling drive levels externally. curiosity sets the strength for the curiosity drive and drives are strengths of the remaining sim-specified drives, in order. any drives not so specified are at the InitDrives baseline level.

func (*PVLV) SetNUSs added in v1.8.16 

func (pp *PVLV) SetNUSs(ctx *Context, nPos, nNeg int)

SetNUSs sets the number of _additional_ simulation-specific positive and negative USs (primary value outcomes). This must be called _before_ network Build, which allocates global values that depend on these numbers. Any change must also call network.BuildGlobals. 1 PosUS (curiosity / novelty) and 2 NegUSs (time, effort) are managed automatically by the PVLV code; any additional USs specified here need to be managed by the simulation via the SetUS method. Positive USs each have corresponding Drives.

func (*PVLV) SetUS added in v1.8.16 

func (pp *PVLV) SetUS(ctx *Context, di uint32, valence ValenceTypes, usIdx int, magnitude float32)

SetUS sets the given _simulation specific_ unconditioned stimulus (US) state for PVLV algorithm. usIdx = 0 is first additional US, etc. The US then drives activity of relevant PVLV-rendered inputs, and dopamine. By default, negative USs only set the overall ctx.NeuroMod.HasRew flag when they exceed the NegUSOutcomeThr, thus triggering a full-blown US learning event. Otherwise, they accumulate as in the case of effort and time, and can then trigger giving up as a function of the total accumulated negative valence.

func (*PVLV) Step added in v1.8.16 

func (pp *PVLV) Step(ctx *Context, di uint32, rnd erand.Rand)

Step does one step (trial) after applying USs, Drives, and updating Effort. It should be the final call in ApplyPVLV. Calls PVDA which does all US, PV, LHb, GiveUp updating.

func (*PVLV) TimeEffortReset added in v1.8.16 

func (pp *PVLV) TimeEffortReset(ctx *Context, di uint32)

TimeEffortReset resets the raw time and effort back to zero, at start of new gating event

func (*PVLV) USnegIdx added in v1.8.16 

func (pp *PVLV) USnegIdx(simUsIdx int) int

USnegIdx adds 2 to the given _simulation specific_ negative US index to get the actual US index, where the first pool is reserved for time, and the second for effort

func (*PVLV) USposIdx added in v1.8.16 

func (pp *PVLV) USposIdx(simUsIdx int) int

USposIdx adds 1 to the given _simulation specific_ positive US index to get the actual US / Drive index, where the first pool is reserved for curiosity / novelty.

func (*PVLV) Update added in v1.7.11 

func (pp *PVLV) Update()

func (*PVLV) VSPatchNewState added in v1.8.16 

func (pp *PVLV) VSPatchNewState(ctx *Context, di uint32)

VSPatchNewState does VSPatch processing in NewState: saves to Prev, updates global VSPatchPos and VSPatchPosSum. uses max across recorded VSPatch activity levels.

type Pool 

type Pool struct {

	// starting and ending (exlusive) layer-wise indexes for the list of neurons in this pool
	StIdx, EdIdx uint32 `inactive:"+" desc:"starting and ending (exlusive) layer-wise indexes for the list of neurons in this pool"`

	// [view: -] layer index in global layer list
	LayIdx uint32 `view:"-" desc:"layer index in global layer list"`

	// [view: -] data parallel index (innermost index per layer)
	DataIdx uint32 `view:"-" desc:"data parallel index (innermost index per layer)"`

	// [view: -] pool index in global pool list: [Layer][Pool][Data]
	PoolIdx uint32 `view:"-" desc:"pool index in global pool list: [Layer][Pool][Data]"`

	// is this a layer-wide pool?  if not, it represents a sub-pool of units within a 4D layer
	IsLayPool slbool.Bool `inactive:"+" desc:"is this a layer-wide pool?  if not, it represents a sub-pool of units within a 4D layer"`

	// for special types where relevant (e.g., MatrixLayer, BGThalLayer), indicates if the pool was gated
	Gated slbool.Bool `inactive:"+" desc:"for special types where relevant (e.g., MatrixLayer, BGThalLayer), indicates if the pool was gated"`

	// fast-slow FFFB inhibition values
	Inhib fsfffb.Inhib `inactive:"+" desc:"fast-slow FFFB inhibition values"`

	// average and max values for relevant variables in this pool, at different time scales
	AvgMax PoolAvgMax `desc:"average and max values for relevant variables in this pool, at different time scales"`

	// [view: inline] absolute value of AvgDif differences from actual neuron ActPct relative to TrgAvg
	AvgDif AvgMaxI32 `inactive:"+" view:"inline" desc:"absolute value of AvgDif differences from actual neuron ActPct relative to TrgAvg"`
	// contains filtered or unexported fields
}

Pool contains computed values for FS-FFFB inhibition, and various other state values for layers and pools (unit groups) that can be subject to inhibition

func (*Pool) AvgMaxUpdate added in v1.8.0 

func (pl *Pool) AvgMaxUpdate(ctx *Context, ni, di uint32)

AvgMaxUpdate updates the AvgMax values based on current neuron values

func (*Pool) Init 

func (pl *Pool) Init()

Init is callled during InitActs

func (*Pool) NNeurons added in v1.5.12 

func (pl *Pool) NNeurons() int

NNeurons returns the number of neurons in the pool: EdIdx - StIdx

func (*Pool) TestVals added in v1.8.0 

func (pl *Pool) TestVals(layKey string, vals map[string]float32)

TestVals returns a map of CaSpkD.Avg, which provides an integrated summary of pool activity for testing

type PoolAvgMax added in v1.7.0 

type PoolAvgMax struct {

	// [view: inline] avg and maximum CaSpkP (continuously updated at roughly 40 msec integration window timescale, ends up capturing potentiation, plus-phase signal) -- this is the primary variable to use for tracking overall pool activity
	CaSpkP AvgMaxPhases `` /* 252-byte string literal not displayed */

	// [view: inline] avg and maximum CaSpkD longer-term depression / DAPK1 signal in layer
	CaSpkD AvgMaxPhases `inactive:"+" view:"inline" desc:"avg and maximum CaSpkD longer-term depression / DAPK1 signal in layer"`

	// [view: inline] avg and maximum SpkMax value (based on CaSpkP) -- reflects peak activity at any point across the cycle
	SpkMax AvgMaxPhases `` /* 136-byte string literal not displayed */

	// [view: inline] avg and maximum Act firing rate value
	Act AvgMaxPhases `inactive:"+" view:"inline" desc:"avg and maximum Act firing rate value"`

	// [view: inline] avg and maximum GeInt integrated running-average excitatory conductance value
	GeInt AvgMaxPhases `inactive:"+" view:"inline" desc:"avg and maximum GeInt integrated running-average excitatory conductance value"`

	// [view: inline] avg and maximum GiInt integrated running-average inhibitory conductance value
	GiInt AvgMaxPhases `inactive:"+" view:"inline" desc:"avg and maximum GiInt integrated running-average inhibitory conductance value"`
}

PoolAvgMax contains the average and maximum values over a Pool of neurons for different variables of interest, at Cycle, Minus and Plus phase timescales. All of the cycle level values are updated at the *start* of the cycle based on values from the prior cycle -- thus are 1 cycle behind in general.

func (*PoolAvgMax) Calc added in v1.7.9 

func (am *PoolAvgMax) Calc(refIdx int32)

Calc does Calc on Cycle level, and re-inits

func (*PoolAvgMax) CycleToMinus added in v1.7.0 

func (am *PoolAvgMax) CycleToMinus()

CycleToMinus grabs current Cycle values into the Minus phase values

func (*PoolAvgMax) CycleToPlus added in v1.7.0 

func (am *PoolAvgMax) CycleToPlus()

CycleToPlus grabs current Cycle values into the Plus phase values

func (*PoolAvgMax) Init added in v1.7.0 

func (am *PoolAvgMax) Init()

Init does Init on Cycle vals-- for update start. always left init'd so generally unnecessary

func (*PoolAvgMax) SetN added in v1.7.9 

func (am *PoolAvgMax) SetN(n int32)

SetN sets the N for aggregation

func (*PoolAvgMax) Zero added in v1.7.9 

func (am *PoolAvgMax) Zero()

Zero does full reset on everything -- for InitActs

type PopCodeParams added in v1.7.11 

type PopCodeParams struct {

	// use popcode encoding of variable(s) that this layer represents
	On slbool.Bool `desc:"use popcode encoding of variable(s) that this layer represents"`

	// [def: 0.1] [viewif: On] Ge multiplier for driving excitatory conductance based on PopCode -- multiplies normalized activation values
	Ge float32 `` /* 137-byte string literal not displayed */

	// [def: -0.1] [viewif: On] minimum value representable -- for GaussBump, typically include extra to allow mean with activity on either side to represent the lowest value you want to encode
	Min float32 `` /* 191-byte string literal not displayed */

	// [def: 1.1] [viewif: On] maximum value representable -- for GaussBump, typically include extra to allow mean with activity on either side to represent the lowest value you want to encode
	Max float32 `` /* 190-byte string literal not displayed */

	// [def: 1,0.5] [viewif: On] activation multiplier for values at Min end of range, where values at Max end have an activation of 1 -- if this is &lt; 1, then there is a rate code proportional to the value in addition to the popcode pattern -- see also MinSigma, MaxSigma
	MinAct float32 `` /* 272-byte string literal not displayed */

	// [def: 0.1,0.08] [viewif: On] sigma parameter of a gaussian specifying the tuning width of the coarse-coded units, in normalized 0-1 range -- for Min value -- if MinSigma &lt; MaxSigma then more units are activated for Max values vs. Min values, proportionally
	MinSigma float32 `` /* 264-byte string literal not displayed */

	// [def: 0.1,0.12] [viewif: On] sigma parameter of a gaussian specifying the tuning width of the coarse-coded units, in normalized 0-1 range -- for Min value -- if MinSigma &lt; MaxSigma then more units are activated for Max values vs. Min values, proportionally
	MaxSigma float32 `` /* 264-byte string literal not displayed */

	// [viewif: On] ensure that encoded and decoded value remains within specified range
	Clip slbool.Bool `viewif:"On" desc:"ensure that encoded and decoded value remains within specified range"`
}

PopCodeParams provides an encoding of scalar value using population code, where a single continuous (scalar) value is encoded as a gaussian bump across a population of neurons (1 dimensional). It can also modulate rate code and number of neurons active according to the value. This is for layers that represent values as in the PVLV system (from Context.PVLV). Both normalized activation values (1 max) and Ge conductance values can be generated.

func (*PopCodeParams) ClipVal added in v1.7.11 

func (pc *PopCodeParams) ClipVal(val float32) float32

ClipVal returns clipped (clamped) value in min / max range

func (*PopCodeParams) Defaults added in v1.7.11 

func (pc *PopCodeParams) Defaults()

func (*PopCodeParams) EncodeGe added in v1.7.11 

func (pc *PopCodeParams) EncodeGe(i, n uint32, val float32) float32

EncodeGe returns Ge value for given value, for neuron index i out of n total neurons. n must be 2 or more.

func (*PopCodeParams) EncodeVal added in v1.7.11 

func (pc *PopCodeParams) EncodeVal(i, n uint32, val float32) float32

EncodeVal returns value for given value, for neuron index i out of n total neurons. n must be 2 or more.

func (*PopCodeParams) ProjectParam added in v1.7.11 

func (pc *PopCodeParams) ProjectParam(minParam, maxParam, clipVal float32) float32

ProjectParam projects given min / max param value onto val within range

func (*PopCodeParams) SetRange added in v1.7.11 

func (pc *PopCodeParams) SetRange(min, max, minSigma, maxSigma float32)

SetRange sets the min, max and sigma values

func (*PopCodeParams) Update added in v1.7.11 

func (pc *PopCodeParams) Update()

type Prjn 

type Prjn struct {
	PrjnBase

	// all prjn-level parameters -- these must remain constant once configured
	Params *PrjnParams `desc:"all prjn-level parameters -- these must remain constant once configured"`
}

axon.Prjn is a basic Axon projection with synaptic learning parameters

func (*Prjn) AllParams 

func (pj *Prjn) AllParams() string

AllParams returns a listing of all parameters in the Layer

func (*Prjn) AsAxon 

func (pj *Prjn) AsAxon() *Prjn

AsAxon returns this prjn as a axon.Prjn -- all derived prjns must redefine this to return the base Prjn type, so that the AxonPrjn interface does not need to include accessors to all the basic stuff.

func (*Prjn) Class added in v1.7.0 

func (pj *Prjn) Class() string

func (*Prjn) DWt 

func (pj *Prjn) DWt(ctx *Context, si uint32)

DWt computes the weight change (learning), based on synaptically-integrated spiking, computed at the Theta cycle interval. This is the trace version for hidden units, and uses syn CaP - CaD for targets.

func (*Prjn) DWtSubMean added in v1.2.23 

func (pj *Prjn) DWtSubMean(ctx *Context, ri uint32)

DWtSubMean subtracts the mean from any projections that have SubMean > 0. This is called on *receiving* projections, prior to WtFmDwt.

func (*Prjn) Defaults 

func (pj *Prjn) Defaults()

func (*Prjn) InitGBuffs added in v1.5.10 

func (pj *Prjn) InitGBuffs()

InitGBuffs initializes the per-projection synaptic conductance buffers. This is not typically needed (called during InitWts, InitActs) but can be called when needed. Must be called to completely initialize prior activity, e.g., full Glong clearing.

func (*Prjn) InitSynCa added in v1.8.0 

func (pj *Prjn) InitSynCa(ctx *Context, syni, di uint32)

InitSynCa initializes synaptic calcium variables

func (*Prjn) InitWtSym 

func (pj *Prjn) InitWtSym(ctx *Context, rpj *Prjn)

InitWtSym initializes weight symmetry. Is given the reciprocal projection where the Send and Recv layers are reversed (see LayerBase RecipToRecvPrjn)

func (*Prjn) InitWts 

func (pj *Prjn) InitWts(ctx *Context, nt *Network)

InitWts initializes weight values according to SWt params, enforcing current constraints.

func (*Prjn) InitWtsSyn 

func (pj *Prjn) InitWtsSyn(ctx *Context, syni uint32, rnd erand.Rand, mean, spct float32)

InitWtsSyn initializes weight values based on WtInit randomness parameters for an individual synapse. It also updates the linear weight value based on the sigmoidal weight value.

func (*Prjn) LRateMod added in v1.6.13 

func (pj *Prjn) LRateMod(mod float32)

LRateMod sets the LRate modulation parameter for Prjns, which is for dynamic modulation of learning rate (see also LRateSched). Updates the effective learning rate factor accordingly.

func (*Prjn) LRateSched added in v1.6.13 

func (pj *Prjn) LRateSched(sched float32)

LRateSched sets the schedule-based learning rate multiplier. See also LRateMod. Updates the effective learning rate factor accordingly.

func (*Prjn) Object added in v1.7.0 

func (pj *Prjn) Object() any

Object returns the object with parameters to be set by emer.Params

func (*Prjn) PrjnType added in v1.7.0 

func (pj *Prjn) PrjnType() PrjnTypes

PrjnType returns axon specific cast of pj.Typ prjn type

func (*Prjn) ReadWtsJSON 

func (pj *Prjn) ReadWtsJSON(r io.Reader) error

ReadWtsJSON reads the weights from this projection from the receiver-side perspective in a JSON text format. This is for a set of weights that were saved *for one prjn only* and is not used for the network-level ReadWtsJSON, which reads into a separate structure -- see SetWts method.

func (*Prjn) SWtFmWt added in v1.2.45 

func (pj *Prjn) SWtFmWt(ctx *Context)

SWtFmWt updates structural, slowly-adapting SWt value based on accumulated DSWt values, which are zero-summed with additional soft bounding relative to SWt limits.

func (*Prjn) SWtRescale added in v1.2.45 

func (pj *Prjn) SWtRescale(ctx *Context)

SWtRescale rescales the SWt values to preserve the target overall mean value, using subtractive normalization.

func (*Prjn) SendSpike 

func (pj *Prjn) SendSpike(ctx *Context, ni, di, maxData uint32)

SendSpike sends a spike from the sending neuron at index sendIdx into the GBuf buffer on the receiver side. The buffer on the receiver side is a ring buffer, which is used for modelling the time delay between sending and receiving spikes.

func (*Prjn) SetSWtsFunc added in v1.2.75 

func (pj *Prjn) SetSWtsFunc(ctx *Context, swtFun func(si, ri int, send, recv *etensor.Shape) float32)

SetSWtsFunc initializes structural SWt values using given function based on receiving and sending unit indexes.

func (*Prjn) SetSWtsRPool added in v1.2.75 

func (pj *Prjn) SetSWtsRPool(ctx *Context, swts etensor.Tensor)

SetSWtsRPool initializes SWt structural weight values using given tensor of values which has unique values for each recv neuron within a given pool.

func (*Prjn) SetSynVal 

func (pj *Prjn) SetSynVal(varNm string, sidx, ridx int, val float32) error

SetSynVal sets value of given variable name on the synapse between given send, recv unit indexes (1D, flat indexes) returns error for access errors.

func (*Prjn) SetWts 

func (pj *Prjn) SetWts(pw *weights.Prjn) error

SetWts sets the weights for this projection from weights.Prjn decoded values

func (*Prjn) SetWtsFunc 

func (pj *Prjn) SetWtsFunc(ctx *Context, wtFun func(si, ri int, send, recv *etensor.Shape) float32)

SetWtsFunc initializes synaptic Wt value using given function based on receiving and sending unit indexes. Strongly suggest calling SWtRescale after.

func (*Prjn) SlowAdapt added in v1.2.37 

func (pj *Prjn) SlowAdapt(ctx *Context)

SlowAdapt does the slow adaptation: SWt learning and SynScale

func (*Prjn) SynCaRecv added in v1.7.9 

func (pj *Prjn) SynCaRecv(ctx *Context, ni, di uint32, updtThr float32)

SynCaRecv updates synaptic calcium based on spiking, for SynSpkTheta mode. Optimized version only updates at point of spiking. This pass goes through in recv order, filtering on recv spike, and skips when sender spiked, as those were already done in Send version.

func (*Prjn) SynCaReset added in v1.8.0 

func (pj *Prjn) SynCaReset(ctx *Context)

SynCaReset resets SynCa values -- called during SlowAdapt

func (*Prjn) SynCaSend added in v1.7.9 

func (pj *Prjn) SynCaSend(ctx *Context, ni, di uint32, updtThr float32)

SynCaSend updates synaptic calcium based on spiking, for SynSpkTheta mode. Optimized version only updates at point of spiking. This pass goes through in sending order, filtering on sending spike. Sender will update even if recv neuron spiked -- recv will skip sender spike cases.

func (*Prjn) SynFail added in v1.2.92 

func (pj *Prjn) SynFail(ctx *Context)

SynFail updates synaptic weight failure only -- normally done as part of DWt and WtFmDWt, but this call can be used during testing to update failing synapses.

func (*Prjn) SynScale added in v1.2.23 

func (pj *Prjn) SynScale(ctx *Context)

SynScale performs synaptic scaling based on running average activation vs. targets. Layer-level AvgDifFmTrgAvg function must be called first.

func (*Prjn) Update added in v1.7.0 

func (pj *Prjn) Update()

Update is interface that does local update of struct vals

func (*Prjn) UpdateParams 

func (pj *Prjn) UpdateParams()

UpdateParams updates all params given any changes that might have been made to individual values

func (*Prjn) WriteWtsJSON 

func (pj *Prjn) WriteWtsJSON(w io.Writer, depth int)

WriteWtsJSON writes the weights from this projection from the receiver-side perspective in a JSON text format. We build in the indentation logic to make it much faster and more efficient.

func (*Prjn) WtFmDWt 

func (pj *Prjn) WtFmDWt(ctx *Context, ni uint32)

WtFmDWt computes the weight change (learning), based on synaptically-integrated spiking, computed at the Theta cycle interval. This is the trace version for hidden units, and uses syn CaP - CaD for targets.

type PrjnBase added in v1.4.14 

type PrjnBase struct {

	// [view: -] we need a pointer to ourselves as an AxonPrjn, which can always be used to extract the true underlying type of object when prjn is embedded in other structs -- function receivers do not have this ability so this is necessary.
	AxonPrj AxonPrjn `` /* 267-byte string literal not displayed */

	// inactivate this projection -- allows for easy experimentation
	Off bool `desc:"inactivate this projection -- allows for easy experimentation"`

	// Class is for applying parameter styles, can be space separated multple tags
	Cls string `desc:"Class is for applying parameter styles, can be space separated multple tags"`

	// can record notes about this projection here
	Notes string `desc:"can record notes about this projection here"`

	// sending layer for this projection
	Send *Layer `desc:"sending layer for this projection"`

	// receiving layer for this projection
	Recv *Layer `desc:"receiving layer for this projection"`

	// [tableview: -] pattern of connectivity
	Pat prjn.Pattern `tableview:"-" desc:"pattern of connectivity"`

	// type of projection -- Forward, Back, Lateral, or extended type in specialized algorithms -- matches against .Cls parameter styles (e.g., .Back etc)
	Typ PrjnTypes `` /* 154-byte string literal not displayed */

	// [tableview: -] default parameters that are applied prior to user-set parameters -- these are useful for specific functionality in specialized brain areas (e.g., PVLV, BG etc) not associated with a prjn type, which otherwise is used to hard-code initial default parameters -- typically just set to a literal map.
	DefParams params.Params `` /* 317-byte string literal not displayed */

	// [tableview: -] provides a history of parameters applied to the layer
	ParamsHistory params.HistoryImpl `tableview:"-" desc:"provides a history of parameters applied to the layer"`

	// [view: inline] [tableview: -] average and maximum number of recv connections in the receiving layer
	RecvConNAvgMax minmax.AvgMax32 `tableview:"-" inactive:"+" view:"inline" desc:"average and maximum number of recv connections in the receiving layer"`

	// [view: inline] [tableview: -] average and maximum number of sending connections in the sending layer
	SendConNAvgMax minmax.AvgMax32 `tableview:"-" inactive:"+" view:"inline" desc:"average and maximum number of sending connections in the sending layer"`

	// [view: -] start index into global Synapse array: [Layer][SendPrjns][Synapses]
	SynStIdx uint32 `view:"-" desc:"start index into global Synapse array: [Layer][SendPrjns][Synapses]"`

	// [view: -] number of synapses in this projection
	NSyns uint32 `view:"-" desc:"number of synapses in this projection"`

	// [view: -] [RecvNeurons] starting offset and N cons for each recv neuron, for indexing into the RecvSynIdx array of indexes into the Syns synapses, which are organized sender-based.  This is locally-managed during build process, but also copied to network global PrjnRecvCons slice for GPU usage.
	RecvCon []StartN `` /* 301-byte string literal not displayed */

	// [view: -] [SendNeurons][SendCon.N RecvNeurons] index into Syns synaptic state for each sending unit and connection within that, for the sending projection which does not own the synapses, and instead indexes into recv-ordered list
	RecvSynIdx []uint32 `` /* 236-byte string literal not displayed */

	// [view: -] [RecvNeurons][RecvCon.N SendingNeurons] for each recv synapse, this is index of *sending* neuron  It is generally preferable to use the Synapse SendIdx where needed, instead of this slice, because then the memory access will be close by other values on the synapse.
	RecvConIdx []uint32 `` /* 281-byte string literal not displayed */

	// [view: -] [SendNeurons] starting offset and N cons for each sending neuron, for indexing into the Syns synapses, which are organized sender-based.  This is locally-managed during build process, but also copied to network global PrjnSendCons slice for GPU usage.
	SendCon []StartN `` /* 267-byte string literal not displayed */

	// [view: -] [SendNeurons[[SendCon.N RecvNeurons] index of other neuron that receives the sender's synaptic input, ordered by the sending layer's order of units as the outer loop, and SendCon.N receiving units within that.  It is generally preferable to use the Synapse RecvIdx where needed, instead of this slice, because then the memory access will be close by other values on the synapse.
	SendConIdx []uint32 `` /* 394-byte string literal not displayed */

	// [view: -] [RecvNeurons][Params.Com.MaxDelay][MaxData] Ge or Gi conductance ring buffer for each neuron, accessed through Params.Com.ReadIdx, WriteIdx -- scale * weight is added with Com delay offset -- a subslice from network PrjnGBuf. Uses int-encoded float values for faster GPU atomic integration
	GBuf []int32 `` /* 305-byte string literal not displayed */

	// [view: -] [RecvNeurons][MaxData] projection-level synaptic conductance values, integrated by prjn before being integrated at the neuron level, which enables the neuron to perform non-linear integration as needed -- a subslice from network PrjnGSyn.
	GSyns []float32 `` /* 254-byte string literal not displayed */
}

PrjnBase contains the basic structural information for specifying a projection of synaptic connections between two layers, and maintaining all the synaptic connection-level data. The same struct token is added to the Recv and Send layer prjn lists, and it manages everything about the connectivity, and methods on the Prjn handle all the relevant computation. The Base does not have algorithm-specific methods and parameters, so it can be easily reused for different algorithms, and cleanly separates the algorithm-specific code. Any dependency on the algorithm-level Prjn can be captured in the AxonPrjn interface, accessed via the AxonPrj field.

func (*PrjnBase) AddClass added in v1.8.4 

func (pj *PrjnBase) AddClass(cls string)

func (*PrjnBase) ApplyDefParams added in v1.7.18 

func (pj *PrjnBase) ApplyDefParams()

ApplyDefParams applies DefParams default parameters if set Called by Prjn.Defaults()

func (*PrjnBase) ApplyParams added in v1.4.14 

func (pj *PrjnBase) ApplyParams(pars *params.Sheet, setMsg bool) (bool, error)

ApplyParams applies given parameter style Sheet to this projection. Calls UpdateParams if anything set to ensure derived parameters are all updated. If setMsg is true, then a message is printed to confirm each parameter that is set. it always prints a message if a parameter fails to be set. returns true if any params were set, and error if there were any errors.

func (*PrjnBase) Build added in v1.7.0 

func (pj *PrjnBase) Build() error

Build constructs the full connectivity among the layers. Calls Validate and returns error if invalid. Pat.Connect is called to get the pattern of the connection. Then the connection indexes are configured according to that pattern. Does NOT allocate synapses -- these are set by Network from global slice.

func (*PrjnBase) Class added in v1.4.14 

func (pj *PrjnBase) Class() string

func (*PrjnBase) Connect added in v1.4.14 

func (pj *PrjnBase) Connect(slay, rlay *Layer, pat prjn.Pattern, typ PrjnTypes)

Connect sets the connectivity between two layers and the pattern to use in interconnecting them

func (*PrjnBase) Init added in v1.4.14 

func (pj *PrjnBase) Init(prjn emer.Prjn)

Init MUST be called to initialize the prjn's pointer to itself as an emer.Prjn which enables the proper interface methods to be called.

func (*PrjnBase) IsOff added in v1.4.14 

func (pj *PrjnBase) IsOff() bool

func (*PrjnBase) Label added in v1.4.14 

func (pj *PrjnBase) Label() string

func (*PrjnBase) Name added in v1.4.14 

func (pj *PrjnBase) Name() string

func (*PrjnBase) NonDefaultParams added in v1.4.14 

func (pj *PrjnBase) NonDefaultParams() string

NonDefaultParams returns a listing of all parameters in the Layer that are not at their default values -- useful for setting param styles etc.

func (*PrjnBase) ParamsApplied added in v1.7.11 

func (pj *PrjnBase) ParamsApplied(sel *params.Sel)

ParamsApplied is just to satisfy History interface so reset can be applied

func (*PrjnBase) ParamsHistoryReset added in v1.7.11 

func (pj *PrjnBase) ParamsHistoryReset()

ParamsHistoryReset resets parameter application history

func (*PrjnBase) Pattern added in v1.4.14 

func (pj *PrjnBase) Pattern() prjn.Pattern

func (*PrjnBase) PrjnTypeName added in v1.4.14 

func (pj *PrjnBase) PrjnTypeName() string

func (*PrjnBase) RecvLay added in v1.4.14 

func (pj *PrjnBase) RecvLay() emer.Layer

func (*PrjnBase) RecvSynIdxs added in v1.7.24 

func (pj *PrjnBase) RecvSynIdxs(ri uint32) []uint32

RecvSynIdxs returns the receiving synapse indexes for given recv unit index within the receiving layer, to be iterated over for recv-based processing.

func (*PrjnBase) SendLay added in v1.4.14 

func (pj *PrjnBase) SendLay() emer.Layer

func (*PrjnBase) SetClass added in v1.7.0 

func (pj *PrjnBase) SetClass(cls string) emer.Prjn

func (*PrjnBase) SetConStartN added in v1.7.2 

func (pj *PrjnBase) SetConStartN(con *[]StartN, avgmax *minmax.AvgMax32, tn *etensor.Int32) uint32

SetConStartN sets the *Con StartN values given n tensor from Pat. Returns total number of connections for this direction.

func (*PrjnBase) SetOff added in v1.4.14 

func (pj *PrjnBase) SetOff(off bool)

SetOff individual projection. Careful: Layer.SetOff(true) will reactivate all prjns of that layer, so prjn-level lesioning should always be done last.

func (*PrjnBase) SetPattern added in v1.7.0 

func (pj *PrjnBase) SetPattern(pat prjn.Pattern) emer.Prjn

func (*PrjnBase) SetType added in v1.7.0 

func (pj *PrjnBase) SetType(typ emer.PrjnType) emer.Prjn

func (*PrjnBase) String added in v1.4.14 

func (pj *PrjnBase) String() string

String satisfies fmt.Stringer for prjn

func (*PrjnBase) Syn1DNum added in v1.7.0 

func (pj *PrjnBase) Syn1DNum() int

Syn1DNum returns the number of synapses for this prjn as a 1D array. This is the max idx for SynVal1D and the number of vals set by SynVals.

func (*PrjnBase) SynIdx added in v1.7.0 

func (pj *PrjnBase) SynIdx(sidx, ridx int) int

SynIdx returns the index of the synapse between given send, recv unit indexes (1D, flat indexes, layer relative). Returns -1 if synapse not found between these two neurons. Requires searching within connections for sending unit.

func (*PrjnBase) SynVal added in v1.7.0 

func (pj *PrjnBase) SynVal(varNm string, sidx, ridx int) float32

SynVal returns value of given variable name on the synapse between given send, recv unit indexes (1D, flat indexes). Returns mat32.NaN() for access errors (see SynValTry for error message)

func (*PrjnBase) SynVal1D added in v1.7.0 

func (pj *PrjnBase) SynVal1D(varIdx int, synIdx int) float32

SynVal1D returns value of given variable index (from SynVarIdx) on given SynIdx. Returns NaN on invalid index. This is the core synapse var access method used by other methods.

func (*PrjnBase) SynVal1DDi added in v1.8.0 

func (pj *PrjnBase) SynVal1DDi(varIdx int, synIdx int, di int) float32

SynVal1DDi returns value of given variable index (from SynVarIdx) on given SynIdx. Returns NaN on invalid index. This is the core synapse var access method used by other methods. Includes Di data parallel index for data-parallel synaptic values.

func (*PrjnBase) SynValDi added in v1.8.0 

func (pj *PrjnBase) SynValDi(varNm string, sidx, ridx int, di int) float32

SynValDi returns value of given variable name on the synapse between given send, recv unit indexes (1D, flat indexes). Returns mat32.NaN() for access errors (see SynValTry for error message) Includes Di data parallel index for data-parallel synaptic values.

func (*PrjnBase) SynVals added in v1.7.0 

func (pj *PrjnBase) SynVals(vals *[]float32, varNm string) error

SynVals sets values of given variable name for each synapse, using the natural ordering of the synapses (sender based for Axon), into given float32 slice (only resized if not big enough). Returns error on invalid var name.

func (*PrjnBase) SynVarIdx added in v1.7.0 

func (pj *PrjnBase) SynVarIdx(varNm string) (int, error)

SynVarIdx returns the index of given variable within the synapse, according to *this prjn's* SynVarNames() list (using a map to lookup index), or -1 and error message if not found.

func (*PrjnBase) SynVarNames added in v1.7.0 

func (pj *PrjnBase) SynVarNames() []string

func (*PrjnBase) SynVarNum added in v1.7.0 

func (pj *PrjnBase) SynVarNum() int

SynVarNum returns the number of synapse-level variables for this prjn. This is needed for extending indexes in derived types.

func (*PrjnBase) SynVarProps added in v1.7.0 

func (pj *PrjnBase) SynVarProps() map[string]string

SynVarProps returns properties for variables

func (*PrjnBase) Type added in v1.4.14 

func (pj *PrjnBase) Type() emer.PrjnType

func (*PrjnBase) TypeName added in v1.4.14 

func (pj *PrjnBase) TypeName() string

func (*PrjnBase) Validate added in v1.4.14 

func (pj *PrjnBase) Validate(logmsg bool) error

Validate tests for non-nil settings for the projection -- returns error message or nil if no problems (and logs them if logmsg = true)

type PrjnGTypes added in v1.7.0 

type PrjnGTypes int32

PrjnGTypes represents the conductance (G) effects of a given projection, including excitatory, inhibitory, and modulatory. 

const (
	// Excitatory projections drive Ge conductance on receiving neurons,
	// which send to GiRaw and GiSyn neuron variables.
	ExcitatoryG PrjnGTypes = iota

	// Inhibitory projections drive Gi inhibitory conductance,
	// which send to GiRaw and GiSyn neuron variables.
	InhibitoryG

	// Modulatory projections have a multiplicative effect on other inputs,
	// which send to GModRaw and GModSyn neuron variables.
	ModulatoryG

	// Maintenance projections drive unique set of NMDA channels that support
	// strong active maintenance abilities.
	// Send to GMaintRaw and GMaintSyn neuron variables.
	MaintG

	// Context projections are for inputs to CT layers, which update
	// only at the end of the plus phase, and send to CtxtGe.
	ContextG

	PrjnGTypesN
)

The projection conductance types

func (PrjnGTypes) Desc added in v1.8.14 

func (i PrjnGTypes) Desc() string

func (*PrjnGTypes) FromString added in v1.7.0 

func (i *PrjnGTypes) FromString(s string) error

func (PrjnGTypes) MarshalJSON added in v1.7.0 

func (ev PrjnGTypes) MarshalJSON() ([]byte, error)

func (PrjnGTypes) String added in v1.7.0 

func (i PrjnGTypes) String() string

func (*PrjnGTypes) UnmarshalJSON added in v1.7.0 

func (ev *PrjnGTypes) UnmarshalJSON(b []byte) error

type PrjnIdxs added in v1.7.0 

type PrjnIdxs struct {
	PrjnIdx    uint32 // index of the projection in global prjn list: [Layer][SendPrjns]
	RecvLay    uint32 // index of the receiving layer in global list of layers
	RecvNeurSt uint32 // starting index of neurons in recv layer -- so we don't need layer to get to neurons
	RecvNeurN  uint32 // number of neurons in recv layer
	SendLay    uint32 // index of the sending layer in global list of layers
	SendNeurSt uint32 // starting index of neurons in sending layer -- so we don't need layer to get to neurons
	SendNeurN  uint32 // number of neurons in send layer
	SynapseSt  uint32 // start index into global Synapse array: [Layer][SendPrjns][Synapses]
	SendConSt  uint32 // start index into global PrjnSendCon array: [Layer][SendPrjns][SendNeurons]
	RecvConSt  uint32 // start index into global PrjnRecvCon array: [Layer][RecvPrjns][RecvNeurons]
	RecvSynSt  uint32 // start index into global sender-based Synapse index array: [Layer][SendPrjns][Synapses]
	GBufSt     uint32 // start index into global PrjnGBuf global array: [Layer][RecvPrjns][RecvNeurons][MaxDelay+1]
	GSynSt     uint32 // start index into global PrjnGSyn global array: [Layer][RecvPrjns][RecvNeurons]
	// contains filtered or unexported fields
}

PrjnIdxs contains prjn-level index information into global memory arrays

func (*PrjnIdxs) RecvNIdxToLayIdx added in v1.7.2 

func (pi *PrjnIdxs) RecvNIdxToLayIdx(ni uint32) uint32

RecvNIdxToLayIdx converts a neuron's index in network level global list of all neurons to receiving layer-specific index-- e.g., for accessing GBuf and GSyn values. Just subtracts RecvNeurSt -- docu-function basically..

func (*PrjnIdxs) SendNIdxToLayIdx added in v1.7.2 

func (pi *PrjnIdxs) SendNIdxToLayIdx(ni uint32) uint32

SendNIdxToLayIdx converts a neuron's index in network level global list of all neurons to sending layer-specific index. Just subtracts SendNeurSt -- docu-function basically..

type PrjnParams added in v1.7.0 

type PrjnParams struct {

	// functional type of prjn -- determines functional code path for specialized layer types, and is synchronized with the Prjn.Typ value
	PrjnType PrjnTypes `` /* 138-byte string literal not displayed */

	// [view: -] recv and send neuron-level projection index array access info
	Idxs PrjnIdxs `view:"-" desc:"recv and send neuron-level projection index array access info"`

	// [view: inline] synaptic communication parameters: delay, probability of failure
	Com SynComParams `view:"inline" desc:"synaptic communication parameters: delay, probability of failure"`

	// [view: inline] projection scaling parameters for computing GScale: modulates overall strength of projection, using both absolute and relative factors, with adaptation option to maintain target max conductances
	PrjnScale PrjnScaleParams `` /* 215-byte string literal not displayed */

	// [view: add-fields] slowly adapting, structural weight value parameters, which control initial weight values and slower outer-loop adjustments
	SWts SWtParams `` /* 147-byte string literal not displayed */

	// [view: add-fields] synaptic-level learning parameters for learning in the fast LWt values.
	Learn LearnSynParams `view:"add-fields" desc:"synaptic-level learning parameters for learning in the fast LWt values."`

	// [view: inline] conductance scaling values
	GScale GScaleVals `view:"inline" desc:"conductance scaling values"`

	// [view: inline] [viewif: PrjnType=[RWPrjn,TDPredPrjn]] Params for RWPrjn and TDPredPrjn for doing dopamine-modulated learning for reward prediction: Da * Send activity. Use in RWPredLayer or TDPredLayer typically to generate reward predictions. If the Da sign is positive, the first recv unit learns fully; for negative, second one learns fully.  Lower lrate applies for opposite cases.  Weights are positive-only.
	RLPred RLPredPrjnParams `` /* 418-byte string literal not displayed */

	// [view: inline] [viewif: PrjnType=MatrixPrjn] for trace-based learning in the MatrixPrjn. A trace of synaptic co-activity is formed, and then modulated by dopamine whenever it occurs.  This bridges the temporal gap between gating activity and subsequent activity, and is based biologically on synaptic tags. Trace is reset at time of reward based on ACh level from CINs.
	Matrix MatrixPrjnParams `` /* 374-byte string literal not displayed */

	// [view: inline] [viewif: PrjnType=BLAPrjn] Basolateral Amygdala projection parameters.
	BLA BLAPrjnParams `viewif:"PrjnType=BLAPrjn" view:"inline" desc:"Basolateral Amygdala projection parameters."`

	// [view: inline] [viewif: PrjnType=HipPrjn] Hip bench parameters.
	Hip HipPrjnParams `viewif:"PrjnType=HipPrjn" view:"inline" desc:"Hip bench parameters."`
	// contains filtered or unexported fields
}

PrjnParams contains all of the prjn parameters. These values must remain constant over the course of computation. On the GPU, they are loaded into a uniform.

func (*PrjnParams) AllParams added in v1.7.0 

func (pj *PrjnParams) AllParams() string

func (*PrjnParams) BLADefaults added in v1.7.18 

func (pj *PrjnParams) BLADefaults()

func (*PrjnParams) CTCtxtPrjnDefaults added in v1.7.18 

func (pj *PrjnParams) CTCtxtPrjnDefaults()

func (*PrjnParams) DWtFmDiDWtSyn added in v1.8.0 

func (pj *PrjnParams) DWtFmDiDWtSyn(ctx *Context, syni uint32)

DWtFmDiDWtSyn updates DWt from data parallel DiDWt values

func (*PrjnParams) DWtSyn added in v1.7.0 

func (pj *PrjnParams) DWtSyn(ctx *Context, syni, si, ri, di uint32, layPool, subPool *Pool, isTarget bool)

DWtSyn is the overall entry point for weight change (learning) at given synapse. It selects appropriate function based on projection type. rpl is the receiving layer SubPool

func (*PrjnParams) DWtSynBLA added in v1.7.18 

func (pj *PrjnParams) DWtSynBLA(ctx *Context, syni, si, ri, di uint32, layPool, subPool *Pool)

DWtSynBLA computes the weight change (learning) at given synapse for BLAPrjn type. Like the BG Matrix learning rule, a synaptic tag "trace" is established at CS onset (ACh) and learning at US / extinction is a function of trace * delta from US activity (temporal difference), which limits learning.

func (*PrjnParams) DWtSynCortex added in v1.7.0 

func (pj *PrjnParams) DWtSynCortex(ctx *Context, syni, si, ri, di uint32, layPool, subPool *Pool, isTarget bool)

DWtSynCortex computes the weight change (learning) at given synapse for cortex. Uses synaptically-integrated spiking, computed at the Theta cycle interval. This is the trace version for hidden units, and uses syn CaP - CaD for targets.

func (*PrjnParams) DWtSynHip added in v1.8.5 

func (pj *PrjnParams) DWtSynHip(ctx *Context, syni, si, ri, di uint32, layPool, subPool *Pool, isTarget bool)

DWtSynHip computes the weight change (learning) at given synapse for cortex + Hip (CPCA Hebb learning). Uses synaptically-integrated spiking, computed at the Theta cycle interval. This is the trace version for hidden units, and uses syn CaP - CaD for targets. Adds proportional CPCA learning rule for hip-specific prjns

func (*PrjnParams) DWtSynMatrix added in v1.7.0 

func (pj *PrjnParams) DWtSynMatrix(ctx *Context, syni, si, ri, di uint32, layPool, subPool *Pool)

DWtSynMatrix computes the weight change (learning) at given synapse, for the MatrixPrjn type.

func (*PrjnParams) DWtSynRWPred added in v1.7.0 

func (pj *PrjnParams) DWtSynRWPred(ctx *Context, syni, si, ri, di uint32, layPool, subPool *Pool)

DWtSynRWPred computes the weight change (learning) at given synapse, for the RWPredPrjn type

func (*PrjnParams) DWtSynTDPred added in v1.7.0 

func (pj *PrjnParams) DWtSynTDPred(ctx *Context, syni, si, ri, di uint32, layPool, subPool *Pool)

DWtSynTDPred computes the weight change (learning) at given synapse, for the TDRewPredPrjn type

func (*PrjnParams) DWtSynVSPatch added in v1.7.11 

func (pj *PrjnParams) DWtSynVSPatch(ctx *Context, syni, si, ri, di uint32, layPool, subPool *Pool)

DWtSynVSPatch computes the weight change (learning) at given synapse, for the VSPatchPrjn type. Currently only supporting the Pos D1 type.

func (*PrjnParams) Defaults added in v1.7.0 

func (pj *PrjnParams) Defaults()

func (*PrjnParams) DoSynCa added in v1.7.11 

func (pj *PrjnParams) DoSynCa() bool

DoSynCa returns false if should not do synaptic-level calcium updating. Done by default in Cortex, not for some other special projection types.

func (*PrjnParams) GatherSpikes added in v1.7.2 

func (pj *PrjnParams) GatherSpikes(ctx *Context, ly *LayerParams, ni, di uint32, gRaw float32, gSyn *float32)

GatherSpikes integrates G*Raw and G*Syn values for given neuron from the given Prjn-level GRaw value, first integrating projection-level GSyn value.

func (*PrjnParams) HipDefaults added in v1.8.5 

func (pj *PrjnParams) HipDefaults()

func (*PrjnParams) IsExcitatory added in v1.7.0 

func (pj *PrjnParams) IsExcitatory() bool

func (*PrjnParams) IsInhib added in v1.7.0 

func (pj *PrjnParams) IsInhib() bool

func (*PrjnParams) MatrixDefaults added in v1.7.11 

func (pj *PrjnParams) MatrixDefaults()

func (*PrjnParams) RLPredDefaults added in v1.7.18 

func (pj *PrjnParams) RLPredDefaults()

func (*PrjnParams) SetFixedWts added in v1.7.11 

func (pj *PrjnParams) SetFixedWts()

SetFixedWts sets parameters for fixed, non-learning weights with a default of Mean = 0.8, Var = 0 strength

func (*PrjnParams) SynCaSyn added in v1.7.24 

func (pj *PrjnParams) SynCaSyn(ctx *Context, syni uint32, ni, di uint32, otherCaSyn, updtThr float32)

SynCaSyn updates synaptic calcium based on spiking, for SynSpkTheta mode. Optimized version only updates at point of spiking, threaded over neurons.

func (*PrjnParams) SynRecvLayIdx added in v1.7.2 

func (pj *PrjnParams) SynRecvLayIdx(ctx *Context, syni uint32) uint32

SynRecvLayIdx converts the Synapse RecvIdx of recv neuron's index in network level global list of all neurons to receiving layer-specific index.

func (*PrjnParams) SynSendLayIdx added in v1.7.2 

func (pj *PrjnParams) SynSendLayIdx(ctx *Context, syni uint32) uint32

SynSendLayIdx converts the Synapse SendIdx of sending neuron's index in network level global list of all neurons to sending layer-specific index.

func (*PrjnParams) Update added in v1.7.0 

func (pj *PrjnParams) Update()

func (*PrjnParams) VSPatchDefaults added in v1.7.18 

func (pj *PrjnParams) VSPatchDefaults()

func (*PrjnParams) WtFmDWtSyn added in v1.7.0 

func (pj *PrjnParams) WtFmDWtSyn(ctx *Context, syni uint32)

WtFmDWtSyn is the overall entry point for updating weights from weight changes.

func (*PrjnParams) WtFmDWtSynCortex added in v1.7.0 

func (pj *PrjnParams) WtFmDWtSynCortex(ctx *Context, syni uint32)

WtFmDWtSynCortex updates weights from dwt changes

func (*PrjnParams) WtFmDWtSynNoLimits added in v1.7.0 

func (pj *PrjnParams) WtFmDWtSynNoLimits(ctx *Context, syni uint32)

WtFmDWtSynNoLimits -- weight update without limits

type PrjnScaleParams added in v1.2.45 

type PrjnScaleParams struct {

	// [min: 0] [Defaults: Forward=1, Back=0.2] relative scaling that shifts balance between different projections -- this is subject to normalization across all other projections into receiving neuron, and determines the GScale.Target for adapting scaling
	Rel float32 `` /* 255-byte string literal not displayed */

	// [def: 1] [min: 0] absolute multiplier adjustment factor for the prjn scaling -- can be used to adjust for idiosyncrasies not accommodated by the standard scaling based on initial target activation level and relative scaling factors -- any adaptation operates by directly adjusting scaling factor from the initially computed value
	Abs float32 `` /* 334-byte string literal not displayed */
	// contains filtered or unexported fields
}

PrjnScaleParams are projection scaling parameters: modulates overall strength of projection, using both absolute and relative factors.

func (*PrjnScaleParams) Defaults added in v1.2.45 

func (ws *PrjnScaleParams) Defaults()

func (*PrjnScaleParams) FullScale added in v1.2.45 

func (ws *PrjnScaleParams) FullScale(savg, snu, ncon float32) float32

FullScale returns full scaling factor, which is product of Abs * Rel * SLayActScale

func (*PrjnScaleParams) SLayActScale added in v1.2.45 

func (ws *PrjnScaleParams) SLayActScale(savg, snu, ncon float32) float32

SLayActScale computes scaling factor based on sending layer activity level (savg), number of units in sending layer (snu), and number of recv connections (ncon). Uses a fixed sem_extra standard-error-of-the-mean (SEM) extra value of 2 to add to the average expected number of active connections to receive, for purposes of computing scaling factors with partial connectivity For 25% layer activity, binomial SEM = sqrt(p(1-p)) = .43, so 3x = 1.3 so 2 is a reasonable default.

func (*PrjnScaleParams) Update added in v1.2.45 

func (ws *PrjnScaleParams) Update()

type PrjnTypes added in v1.7.0 

type PrjnTypes int32

PrjnTypes is an axon-specific prjn type enum, that encompasses all the different algorithm types supported. Class parameter styles automatically key off of these types. The first entries must be kept synchronized with the emer.PrjnType. 

const (
	// Forward is a feedforward, bottom-up projection from sensory inputs to higher layers
	ForwardPrjn PrjnTypes = iota

	// Back is a feedback, top-down projection from higher layers back to lower layers
	BackPrjn

	// Lateral is a lateral projection within the same layer / area
	LateralPrjn

	// Inhib is an inhibitory projection that drives inhibitory
	// synaptic conductances instead of the default excitatory ones.
	InhibPrjn

	// CTCtxt are projections from Superficial layers to CT layers that
	// send Burst activations drive updating of CtxtGe excitatory conductance,
	// at end of plus (51B Bursting) phase.  Biologically, this projection
	// comes from the PT layer 5IB neurons, but it is simpler to use the
	// Super neurons directly, and PT are optional for most network types.
	// These projections also use a special learning rule that
	// takes into account the temporal delays in the activation states.
	// Can also add self context from CT for deeper temporal context.
	CTCtxtPrjn

	// RWPrjn does dopamine-modulated learning for reward prediction:
	// Da * Send.CaSpkP (integrated current spiking activity).
	// Uses RLPredPrjn parameters.
	// Use in RWPredLayer typically to generate reward predictions.
	// If the Da sign is positive, the first recv unit learns fully;
	// for negative, second one learns fully.  Lower lrate applies for
	// opposite cases.  Weights are positive-only.
	RWPrjn

	// TDPredPrjn does dopamine-modulated learning for reward prediction:
	// DWt = Da * Send.SpkPrv (activity on *previous* timestep)
	// Uses RLPredPrjn parameters.
	// Use in TDPredLayer typically to generate reward predictions.
	// If the Da sign is positive, the first recv unit learns fully;
	// for negative, second one learns fully.  Lower lrate applies for
	// opposite cases.  Weights are positive-only.
	TDPredPrjn

	// BLAPrjn implements the PVLV BLA learning rule:
	// dW = ACh * X_t-1 * (Y_t - Y_t-1)
	// The recv delta is across trials, where the US should activate on trial
	// boundary, to enable sufficient time for gating through to OFC, so
	// BLA initially learns based on US present - US absent.
	// It can also learn based on CS onset if there is a prior CS that predicts that.
	BLAPrjn

	HipPrjn

	// VSPatchPrjn implements the VSPatch learning rule:
	// dW = ACh * DA * X * Y
	// where DA is D1 vs. D2 modulated DA level, X = sending activity factor,
	// Y = receiving activity factor, and ACh provides overall modulation.
	VSPatchPrjn

	// MatrixPrjn supports trace-based learning, where an initial
	// trace of synaptic co-activity is formed, and then modulated
	// by subsequent phasic dopamine & ACh when an outcome occurs.
	// This bridges the temporal gap between gating activity
	// and subsequent outcomes, and is based biologically on synaptic tags.
	// Trace is reset at time of reward based on ACh level (from CINs in biology).
	MatrixPrjn

	PrjnTypesN
)

The projection types

func (PrjnTypes) Desc added in v1.8.14 

func (i PrjnTypes) Desc() string

func (*PrjnTypes) FromString added in v1.7.0 

func (i *PrjnTypes) FromString(s string) error

func (PrjnTypes) MarshalJSON added in v1.7.0 

func (ev PrjnTypes) MarshalJSON() ([]byte, error)

func (PrjnTypes) String added in v1.7.0 

func (i PrjnTypes) String() string

func (*PrjnTypes) UnmarshalJSON added in v1.7.0 

func (ev *PrjnTypes) UnmarshalJSON(b []byte) error

type PulvParams added in v1.7.0 

type PulvParams struct {

	// [def: 0.1] [min: 0.0] multiplier on driver input strength, multiplies CaSpkP from driver layer to produce Ge excitatory input to Pulv unit.
	DriveScale float32 `` /* 144-byte string literal not displayed */

	// [def: 0.6] [min: 0.01] Level of Max driver layer CaSpkP at which the drivers fully drive the burst phase activation.  If there is weaker driver input, then (Max/FullDriveAct) proportion of the non-driver inputs remain and this critically prevents the network from learning to turn activation off, which is difficult and severely degrades learning.
	FullDriveAct float32 `` /* 352-byte string literal not displayed */

	// index of layer that generates the driving activity into this one -- set via SetBuildConfig(DriveLayName) setting
	DriveLayIdx int32 `` /* 132-byte string literal not displayed */
	// contains filtered or unexported fields
}

PulvParams provides parameters for how the plus-phase (outcome) state of Pulvinar thalamic relay cell neurons is computed from the corresponding driver neuron Burst activation (or CaSpkP if not Super)

func (*PulvParams) Defaults added in v1.7.0 

func (tp *PulvParams) Defaults()

func (*PulvParams) DriveGe added in v1.7.0 

func (tp *PulvParams) DriveGe(act float32) float32

DriveGe returns effective excitatory conductance to use for given driver input Burst activation

func (*PulvParams) NonDrivePct added in v1.7.0 

func (tp *PulvParams) NonDrivePct(drvMax float32) float32

NonDrivePct returns the multiplier proportion of the non-driver based Ge to keep around, based on FullDriveAct and the max activity in driver layer.

func (*PulvParams) Update added in v1.7.0 

func (tp *PulvParams) Update()

type PushOff added in v1.8.2 

type PushOff struct {

	// offset
	Off uint32 `desc:"offset"`
	// contains filtered or unexported fields
}

PushOff has push constants for setting offset into compute shader

type RLPredPrjnParams added in v1.7.0 

type RLPredPrjnParams struct {

	// how much to learn on opposite DA sign coding neuron (0..1)
	OppSignLRate float32 `desc:"how much to learn on opposite DA sign coding neuron (0..1)"`

	// tolerance on DA -- if below this abs value, then DA goes to zero and there is no learning -- prevents prediction from exactly learning to cancel out reward value, retaining a residual valence of signal
	DaTol float32 `` /* 208-byte string literal not displayed */
	// contains filtered or unexported fields
}

RLPredPrjnParams does dopamine-modulated learning for reward prediction: Da * Send.Act Used by RWPrjn and TDPredPrjn within corresponding RWPredLayer or TDPredLayer to generate reward predictions based on its incoming weights, using linear activation function. Has no weight bounds or limits on sign etc.

func (*RLPredPrjnParams) Defaults added in v1.7.0 

func (pj *RLPredPrjnParams) Defaults()

func (*RLPredPrjnParams) Update added in v1.7.0 

func (pj *RLPredPrjnParams) Update()

type RLRateParams added in v1.6.13 

type RLRateParams struct {

	// [def: true] use learning rate modulation
	On slbool.Bool `def:"true" desc:"use learning rate modulation"`

	// [def: 0.05,1] [viewif: On] minimum learning rate multiplier for sigmoidal act (1-act) factor -- prevents lrate from going too low for extreme values.  Set to 1 to disable Sigmoid derivative factor, which is default for Target layers.
	SigmoidMin float32 `` /* 238-byte string literal not displayed */

	// [viewif: On] modulate learning rate as a function of plus - minus differences
	Diff slbool.Bool `viewif:"On" desc:"modulate learning rate as a function of plus - minus differences"`

	// [def: 0.1] [viewif: On&&Diff] threshold on Max(CaSpkP, CaSpkD) below which Min lrate applies -- must be > 0 to prevent div by zero
	SpkThr float32 `` /* 135-byte string literal not displayed */

	// [def: 0.02] [viewif: On&&Diff] threshold on recv neuron error delta, i.e., |CaSpkP - CaSpkD| below which lrate is at Min value
	DiffThr float32 `` /* 131-byte string literal not displayed */

	// [def: 0.001] [viewif: On&&Diff] for Diff component, minimum learning rate value when below ActDiffThr
	Min float32 `viewif:"On&&Diff" def:"0.001" desc:"for Diff component, minimum learning rate value when below ActDiffThr"`
	// contains filtered or unexported fields
}

RLRateParams are recv neuron learning rate modulation parameters. Has two factors: the derivative of the sigmoid based on CaSpkD activity levels, and based on the phase-wise differences in activity (Diff).

func (*RLRateParams) Defaults added in v1.6.13 

func (rl *RLRateParams) Defaults()

func (*RLRateParams) RLRateDiff added in v1.6.13 

func (rl *RLRateParams) RLRateDiff(scap, scad float32) float32

RLRateDiff returns the learning rate as a function of difference between CaSpkP and CaSpkD values

func (*RLRateParams) RLRateSigDeriv added in v1.6.13 

func (rl *RLRateParams) RLRateSigDeriv(act float32, laymax float32) float32

RLRateSigDeriv returns the sigmoid derivative learning rate factor as a function of spiking activity, with mid-range values having full learning and extreme values a reduced learning rate: deriv = act * (1 - act) The activity should be CaSpkP and the layer maximum is used to normalize that to a 0-1 range.

func (*RLRateParams) Update added in v1.6.13 

func (rl *RLRateParams) Update()

type RWDaParams added in v1.7.0 

type RWDaParams struct {

	// tonic baseline Ge level for DA = 0 -- +/- are between 0 and 2*TonicGe -- just for spiking display of computed DA value
	TonicGe float32 `desc:"tonic baseline Ge level for DA = 0 -- +/- are between 0 and 2*TonicGe -- just for spiking display of computed DA value"`

	// idx of RWPredLayer to get reward prediction from -- set during Build from BuildConfig RWPredLayName
	RWPredLayIdx int32 `inactive:"+" desc:"idx of RWPredLayer to get reward prediction from -- set during Build from BuildConfig RWPredLayName"`
	// contains filtered or unexported fields
}

RWDaParams computes a dopamine (DA) signal using simple Rescorla-Wagner learning dynamic (i.e., PV learning in the PVLV framework).

func (*RWDaParams) Defaults added in v1.7.0 

func (rp *RWDaParams) Defaults()

func (*RWDaParams) GeFmDA added in v1.7.0 

func (rp *RWDaParams) GeFmDA(da float32) float32

GeFmDA returns excitatory conductance from DA dopamine value

func (*RWDaParams) Update added in v1.7.0 

func (rp *RWDaParams) Update()

type RWPredParams added in v1.7.0 

type RWPredParams struct {

	// default 0.1..0.99 range of predictions that can be represented -- having a truncated range preserves some sensitivity in dopamine at the extremes of good or poor performance
	PredRange minmax.F32 `` /* 180-byte string literal not displayed */
}

RWPredParams parameterizes reward prediction for a simple Rescorla-Wagner learning dynamic (i.e., PV learning in the PVLV framework).

func (*RWPredParams) Defaults added in v1.7.0 

func (rp *RWPredParams) Defaults()

func (*RWPredParams) Update added in v1.7.0 

func (rp *RWPredParams) Update()

type RandFunIdx added in v1.7.7 

type RandFunIdx uint32
const (
	RandFunActPGe RandFunIdx = iota
	RandFunActPGi
	RandFunIdxN
)

We use this enum to store a unique index for each function that requires random number generation. If you add a new function, you need to add a new enum entry here. RandFunIdxN is the total number of random functions. It autoincrements due to iota.

type SWtAdaptParams added in v1.2.45 

type SWtAdaptParams struct {

	// if true, adaptation is active -- if false, SWt values are not updated, in which case it is generally good to have Init.SPct=0 too.
	On slbool.Bool `` /* 137-byte string literal not displayed */

	// [def: 0.1,0.01,0.001,0.0002] [viewif: On] learning rate multiplier on the accumulated DWt values (which already have fast LRate applied) to incorporate into SWt during slow outer loop updating -- lower values impose stronger constraints, for larger networks that need more structural support, e.g., 0.001 is better after 1,000 epochs in large models.  0.1 is fine for smaller models.
	LRate float32 `` /* 388-byte string literal not displayed */

	// [def: 1] [viewif: On] amount of mean to subtract from SWt delta when updating -- generally best to set to 1
	SubMean float32 `viewif:"On" def:"1" desc:"amount of mean to subtract from SWt delta when updating -- generally best to set to 1"`

	// [def: 6] [viewif: On] gain of sigmoidal constrast enhancement function used to transform learned, linear LWt values into Wt values
	SigGain float32 `` /* 135-byte string literal not displayed */
}

SWtAdaptParams manages adaptation of SWt values

func (*SWtAdaptParams) Defaults added in v1.2.45 

func (sp *SWtAdaptParams) Defaults()

func (*SWtAdaptParams) Update added in v1.2.45 

func (sp *SWtAdaptParams) Update()

type SWtInitParams added in v1.2.45 

type SWtInitParams struct {

	// [def: 0,1,0.5] [min: 0] [max: 1] how much of the initial random weights are captured in the SWt values -- rest goes into the LWt values.  1 gives the strongest initial biasing effect, for larger models that need more structural support. 0.5 should work for most models where stronger constraints are not needed.
	SPct float32 `` /* 315-byte string literal not displayed */

	// [def: 0.5,0.4] target mean weight values across receiving neuron's projection -- the mean SWt values are constrained to remain at this value.  some projections may benefit from lower mean of .4
	Mean float32 `` /* 199-byte string literal not displayed */

	// [def: 0.25] initial variance in weight values, prior to constraints.
	Var float32 `def:"0.25" desc:"initial variance in weight values, prior to constraints."`

	// [def: true] symmetrize the initial weight values with those in reciprocal projection -- typically true for bidirectional excitatory connections
	Sym slbool.Bool `` /* 149-byte string literal not displayed */
}

SWtInitParams for initial SWt values

func (*SWtInitParams) Defaults added in v1.2.45 

func (sp *SWtInitParams) Defaults()

func (*SWtInitParams) RndVar added in v1.2.45 

func (sp *SWtInitParams) RndVar(rnd erand.Rand) float32

RndVar returns the random variance in weight value (zero mean) based on Var param

func (*SWtInitParams) Update added in v1.2.45 

func (sp *SWtInitParams) Update()

type SWtParams added in v1.2.45 

type SWtParams struct {

	// [view: inline] initialization of SWt values
	Init SWtInitParams `view:"inline" desc:"initialization of SWt values"`

	// [view: inline] adaptation of SWt values in response to LWt learning
	Adapt SWtAdaptParams `view:"inline" desc:"adaptation of SWt values in response to LWt learning"`

	// [def: {'Min':0.2,'Max':0.8}] [view: inline] range limits for SWt values
	Limit minmax.F32 `def:"{'Min':0.2,'Max':0.8}" view:"inline" desc:"range limits for SWt values"`
}

SWtParams manages structural, slowly adapting weight values (SWt), in terms of initialization and updating over course of learning. SWts impose initial and slowly adapting constraints on neuron connectivity to encourage differentiation of neuron representations and overall good behavior in terms of not hogging the representational space. The TrgAvg activity constraint is not enforced through SWt -- it needs to be more dynamic and supported by the regular learned weights.

func (*SWtParams) ClipSWt added in v1.2.45 

func (sp *SWtParams) ClipSWt(swt float32) float32

ClipSWt returns SWt value clipped to valid range

func (*SWtParams) ClipWt added in v1.2.75 

func (sp *SWtParams) ClipWt(wt float32) float32

ClipWt returns Wt value clipped to 0-1 range

func (*SWtParams) Defaults added in v1.2.45 

func (sp *SWtParams) Defaults()

func (*SWtParams) InitWtsSyn added in v1.3.5 

func (sp *SWtParams) InitWtsSyn(ctx *Context, syni uint32, rnd erand.Rand, mean, spct float32)

InitWtsSyn initializes weight values based on WtInit randomness parameters for an individual synapse. It also updates the linear weight value based on the sigmoidal weight value.

func (*SWtParams) LWtFmWts added in v1.2.47 

func (sp *SWtParams) LWtFmWts(wt, swt float32) float32

LWtFmWts returns linear, learning LWt from wt and swt. LWt is set to reproduce given Wt relative to given SWt base value.

func (*SWtParams) LinFmSigWt added in v1.2.45 

func (sp *SWtParams) LinFmSigWt(wt float32) float32

LinFmSigWt returns linear weight from sigmoidal contrast-enhanced weight. wt is centered at 1, and normed in range +/- 1 around that, return value is in 0-1 range, centered at .5

func (*SWtParams) SigFmLinWt added in v1.2.45 

func (sp *SWtParams) SigFmLinWt(lw float32) float32

SigFmLinWt returns sigmoidal contrast-enhanced weight from linear weight, centered at 1 and normed in range +/- 1 around that in preparation for multiplying times SWt

func (*SWtParams) Update added in v1.2.45 

func (sp *SWtParams) Update()

func (*SWtParams) WtFmDWt added in v1.2.45 

func (sp *SWtParams) WtFmDWt(wt, lwt *float32, dwt, swt float32)

WtFmDWt updates the synaptic weights from accumulated weight changes. wt is the sigmoidal contrast-enhanced weight and lwt is the linear weight value.

func (*SWtParams) WtVal added in v1.2.45 

func (sp *SWtParams) WtVal(swt, lwt float32) float32

WtVal returns the effective Wt value given the SWt and LWt values

type SpikeNoiseParams added in v1.2.94 

type SpikeNoiseParams struct {

	// add noise simulating background spiking levels
	On slbool.Bool `desc:"add noise simulating background spiking levels"`

	// [def: 100] [viewif: On] mean frequency of excitatory spikes -- typically 50Hz but multiple inputs increase rate -- poisson lambda parameter, also the variance
	GeHz float32 `` /* 163-byte string literal not displayed */

	// [viewif: On] [min: 0] excitatory conductance per spike -- .001 has minimal impact, .01 can be strong, and .15 is needed to influence timing of clamped inputs
	Ge float32 `` /* 162-byte string literal not displayed */

	// [def: 200] [viewif: On] mean frequency of inhibitory spikes -- typically 100Hz fast spiking but multiple inputs increase rate -- poisson lambda parameter, also the variance
	GiHz float32 `` /* 177-byte string literal not displayed */

	// [viewif: On] [min: 0] excitatory conductance per spike -- .001 has minimal impact, .01 can be strong, and .15 is needed to influence timing of clamped inputs
	Gi float32 `` /* 162-byte string literal not displayed */

	// [view: -] Exp(-Interval) which is the threshold for GeNoiseP as it is updated
	GeExpInt float32 `view:"-" json:"-" xml:"-" desc:"Exp(-Interval) which is the threshold for GeNoiseP as it is updated"`

	// [view: -] Exp(-Interval) which is the threshold for GiNoiseP as it is updated
	GiExpInt float32 `view:"-" json:"-" xml:"-" desc:"Exp(-Interval) which is the threshold for GiNoiseP as it is updated"`
	// contains filtered or unexported fields
}

SpikeNoiseParams parameterizes background spiking activity impinging on the neuron, simulated using a poisson spiking process.

func (*SpikeNoiseParams) Defaults added in v1.2.94 

func (an *SpikeNoiseParams) Defaults()

func (*SpikeNoiseParams) PGe added in v1.2.94 

func (an *SpikeNoiseParams) PGe(ctx *Context, p *float32, ni uint32) float32

PGe updates the GeNoiseP probability, multiplying a uniform random number [0-1] and returns Ge from spiking if a spike is triggered

func (*SpikeNoiseParams) PGi added in v1.2.94 

func (an *SpikeNoiseParams) PGi(ctx *Context, p *float32, ni uint32) float32

PGi updates the GiNoiseP probability, multiplying a uniform random number [0-1] and returns Gi from spiking if a spike is triggered

func (*SpikeNoiseParams) Update added in v1.2.94 

func (an *SpikeNoiseParams) Update()

type SpikeParams 

type SpikeParams struct {

	// [def: 0.5] threshold value Theta (Q) for firing output activation (.5 is more accurate value based on AdEx biological parameters and normalization
	Thr float32 `` /* 152-byte string literal not displayed */

	// [def: 0.3] post-spiking membrane potential to reset to, produces refractory effect if lower than VmInit -- 0.3 is apropriate biologically-based value for AdEx (Brette & Gurstner, 2005) parameters.  See also RTau
	VmR float32 `` /* 217-byte string literal not displayed */

	// [def: 3] [min: 1] post-spiking explicit refractory period, in cycles -- prevents Vm updating for this number of cycles post firing -- Vm is reduced in exponential steps over this period according to RTau, being fixed at Tr to VmR exactly
	Tr int32 `` /* 242-byte string literal not displayed */

	// [def: 1.6667] time constant for decaying Vm down to VmR -- at end of Tr it is set to VmR exactly -- this provides a more realistic shape of the post-spiking Vm which is only relevant for more realistic channels that key off of Vm -- does not otherwise affect standard computation
	RTau float32 `` /* 285-byte string literal not displayed */

	// [def: true] if true, turn on exponential excitatory current that drives Vm rapidly upward for spiking as it gets past its nominal firing threshold (Thr) -- nicely captures the Hodgkin Huxley dynamics of Na and K channels -- uses Brette & Gurstner 2005 AdEx formulation
	Exp slbool.Bool `` /* 274-byte string literal not displayed */

	// [def: 0.02] [viewif: Exp] slope in Vm (2 mV = .02 in normalized units) for extra exponential excitatory current that drives Vm rapidly upward for spiking as it gets past its nominal firing threshold (Thr) -- nicely captures the Hodgkin Huxley dynamics of Na and K channels -- uses Brette & Gurstner 2005 AdEx formulation
	ExpSlope float32 `` /* 325-byte string literal not displayed */

	// [def: 0.9] [viewif: Exp] membrane potential threshold for actually triggering a spike when using the exponential mechanism
	ExpThr float32 `` /* 127-byte string literal not displayed */

	// [def: 180] [min: 1] for translating spiking interval (rate) into rate-code activation equivalent, what is the maximum firing rate associated with a maximum activation value of 1
	MaxHz float32 `` /* 182-byte string literal not displayed */

	// [def: 5] [min: 1] constant for integrating the spiking interval in estimating spiking rate
	ISITau float32 `def:"5" min:"1" desc:"constant for integrating the spiking interval in estimating spiking rate"`

	// [view: -] rate = 1 / tau
	ISIDt float32 `view:"-" desc:"rate = 1 / tau"`

	// [view: -] rate = 1 / tau
	RDt float32 `view:"-" desc:"rate = 1 / tau"`
	// contains filtered or unexported fields
}

SpikeParams contains spiking activation function params. Implements a basic thresholded Vm model, and optionally the AdEx adaptive exponential function (adapt is KNaAdapt)

func (*SpikeParams) ActFmISI 

func (sk *SpikeParams) ActFmISI(isi, timeInc, integ float32) float32

ActFmISI computes rate-code activation from estimated spiking interval

func (*SpikeParams) ActToISI 

func (sk *SpikeParams) ActToISI(act, timeInc, integ float32) float32

ActToISI compute spiking interval from a given rate-coded activation, based on time increment (.001 = 1msec default), Act.Dt.Integ

func (*SpikeParams) AvgFmISI 

func (sk *SpikeParams) AvgFmISI(avg float32, isi float32) float32

AvgFmISI returns updated spiking ISI from current isi interval value

func (*SpikeParams) Defaults 

func (sk *SpikeParams) Defaults()

func (*SpikeParams) Update 

func (sk *SpikeParams) Update()

type StartN added in v1.7.2 

type StartN struct {

	// starting offset
	Start uint32 `desc:"starting offset"`

	// number of items -- [Start:Start+N]
	N uint32 `desc:"number of items -- [Start:Start+N]"`
	// contains filtered or unexported fields
}

StartN holds a starting offset index and a number of items arranged from Start to Start+N (exclusive). This is not 16 byte padded and only for use on CPU side.

type SynComParams 

type SynComParams struct {

	// type of conductance (G) communicated by this projection
	GType PrjnGTypes `desc:"type of conductance (G) communicated by this projection"`

	// [def: 2] [min: 0] additional synaptic delay in msec for inputs arriving at this projection.  Must be <= MaxDelay which is set during network building based on MaxDelay of any existing Prjn in the network.  Delay = 0 means a spike reaches receivers in the next Cycle, which is the minimum time (1 msec).  Biologically, subtract 1 from biological synaptic delay values to set corresponding Delay value.
	Delay uint32 `` /* 405-byte string literal not displayed */

	// maximum value of Delay -- based on MaxDelay values when the BuildGBuf function was called when the network was built -- cannot set it longer than this, except by calling BuildGBuf on network after changing MaxDelay to a larger value in any projection in the network.
	MaxDelay uint32 `` /* 286-byte string literal not displayed */

	// probability of synaptic transmission failure -- if > 0, then weights are turned off at random as a function of PFail (times 1-SWt if PFailSwt)
	PFail float32 `` /* 149-byte string literal not displayed */

	// if true, then probability of failure is inversely proportional to SWt structural / slow weight value (i.e., multiply PFail * (1-SWt)))
	PFailSWt slbool.Bool `` /* 141-byte string literal not displayed */

	// [view: -] delay length = actual length of the GBuf buffer per neuron = Delay+1 -- just for speed
	DelLen uint32 `view:"-" desc:"delay length = actual length of the GBuf buffer per neuron = Delay+1 -- just for speed"`
	// contains filtered or unexported fields
}

SynComParams are synaptic communication parameters: used in the Prjn parameters. Includes delay and probability of failure, and Inhib for inhibitory connections, and modulatory projections that have multiplicative-like effects.

func (*SynComParams) Defaults 

func (sc *SynComParams) Defaults()

func (*SynComParams) Fail 

func (sc *SynComParams) Fail(ctx *Context, syni uint32, swt float32)

Fail updates failure status of given weight, given SWt value

func (*SynComParams) FloatFromGBuf added in v1.7.9 

func (sc *SynComParams) FloatFromGBuf(ival int32) float32

FloatFromGBuf converts the given int32 value produced via FloatToGBuf back into a float32 (divides by factor). If the value is negative, a panic is triggered indicating there was numerical overflow in the aggregation. If this occurs, the FloatToIntFactor needs to be decreased.

func (*SynComParams) FloatToGBuf added in v1.7.9 

func (sc *SynComParams) FloatToGBuf(val float32) int32

FloatToGBuf converts the given floating point value to a large int32 for accumulating in GBuf. Note: more efficient to bake factor into scale factor per prjn.

func (*SynComParams) FloatToIntFactor added in v1.7.9 

func (sc *SynComParams) FloatToIntFactor() float32

FloatToIntFactor returns the factor used for converting float32 to int32 in GBuf encoding. Because total G is constrained via scaling factors to be around ~1, it is safe to use a factor that uses most of the available bits, leaving enough room to prevent overflow when adding together the different vals. For encoding, bake this into scale factor in SendSpike, and cast the result to int32.

func (*SynComParams) ReadIdx added in v1.7.2 

func (sc *SynComParams) ReadIdx(rnIdx, di uint32, cycTot int32, nRecvNeurs, maxData uint32) uint32

ReadIdx returns index for reading existing spikes from the GBuf buffer, based on the layer-based recv neuron index, data parallel idx, and the ReadOff offset from the CyclesTotal.

func (*SynComParams) ReadOff added in v1.7.2 

func (sc *SynComParams) ReadOff(cycTot int32) uint32

ReadOff returns offset for reading existing spikes from the GBuf buffer, based on Context CyclesTotal counter which increments each cycle. This is logically the zero position in the ring buffer.

func (*SynComParams) RingIdx added in v1.7.2 

func (sc *SynComParams) RingIdx(i uint32) uint32

RingIdx returns the wrap-around ring index for given raw index. For writing and reading spikes to GBuf buffer, based on Context.CyclesTotal counter. RN: 0 1 2 <- recv neuron indexes DI: 0 1 2 0 1 2 0 1 2 <- delay indexes C0: ^ v <- cycle 0, ring index: ^ = write, v = read C1: ^ v <- cycle 1, shift over by 1 -- overwrite last read C2: v ^ <- cycle 2: read out value stored on C0 -- index wraps around

func (*SynComParams) Update 

func (sc *SynComParams) Update()

func (*SynComParams) WriteIdx added in v1.7.2 

func (sc *SynComParams) WriteIdx(rnIdx, di uint32, cycTot int32, nRecvNeurs, maxData uint32) uint32

WriteIdx returns actual index for writing new spikes into the GBuf buffer, based on the layer-based recv neuron index, data parallel idx, and the WriteOff offset computed from the CyclesTotal.

func (*SynComParams) WriteIdxOff added in v1.7.2 

func (sc *SynComParams) WriteIdxOff(rnIdx, di, wrOff uint32, nRecvNeurs, maxData uint32) uint32

WriteIdxOff returns actual index for writing new spikes into the GBuf buffer, based on the layer-based recv neuron index and the given WriteOff offset.

func (*SynComParams) WriteOff added in v1.7.2 

func (sc *SynComParams) WriteOff(cycTot int32) uint32

WriteOff returns offset for writing new spikes into the GBuf buffer, based on Context CyclesTotal counter which increments each cycle. This is logically the last position in the ring buffer.

func (*SynComParams) WtFail 

func (sc *SynComParams) WtFail(ctx *Context, swt float32) bool

WtFail returns true if synapse should fail, as function of SWt value (optionally)

func (*SynComParams) WtFailP 

func (sc *SynComParams) WtFailP(swt float32) float32

WtFailP returns probability of weight (synapse) failure given current SWt value

type SynapseCaStrides added in v1.8.0 

type SynapseCaStrides struct {

	// synapse level
	Synapse uint64 `desc:"synapse level"`

	// variable level
	Var uint64 `desc:"variable level"`
}

SynapseCaStrides encodes the stride offsets for synapse variable access into network float32 array. Data is always the inner-most variable.

func (*SynapseCaStrides) Idx added in v1.8.0 

func (ns *SynapseCaStrides) Idx(synIdx, di uint32, nvar SynapseCaVars) uint64

Idx returns the index into network float32 array for given synapse, data, and variable

func (*SynapseCaStrides) SetSynapseOuter added in v1.8.0 

func (ns *SynapseCaStrides) SetSynapseOuter(ndata int)

SetSynapseOuter sets strides with synapses as outer loop: [Synapses][Vars][Data], which is optimal for CPU-based computation.

func (*SynapseCaStrides) SetVarOuter added in v1.8.0 

func (ns *SynapseCaStrides) SetVarOuter(nsyn, ndata int)

SetVarOuter sets strides with vars as outer loop: [Vars][Synapses][Data], which is optimal for GPU-based computation.

type SynapseCaVars added in v1.8.0 

type SynapseCaVars int32

SynapseCaVars are synapse variables for calcium involved in learning, which are data parallel input specific. 

const (
	// CaM is first stage running average (mean) Ca calcium level (like CaM = calmodulin), feeds into CaP
	CaM SynapseCaVars = iota

	// CaP is shorter timescale integrated CaM value, representing the plus, LTP direction of weight change and capturing the function of CaMKII in the Kinase learning rule
	CaP

	// CaD is longer timescale integrated CaP value, representing the minus, LTD direction of weight change and capturing the function of DAPK1 in the Kinase learning rule
	CaD

	// CaUpT is time in CyclesTotal of last updating of Ca values at the synapse level, for optimized synaptic-level Ca integration -- converted to / from uint32
	CaUpT

	// Tr is trace of synaptic activity over time -- used for credit assignment in learning.  In MatrixPrjn this is a tag that is then updated later when US occurs.
	Tr

	// DTr is delta (change in) Tr trace of synaptic activity over time
	DTr

	// DiDWt is delta weight for each data parallel index (Di) -- this is directly computed from the Ca values (in cortical version) and then aggregated into the overall DWt (which may be further integrated across MPI nodes), which then drives changes in Wt values
	DiDWt

	SynapseCaVarsN
)

func (SynapseCaVars) Desc added in v1.8.14 

func (i SynapseCaVars) Desc() string

func (*SynapseCaVars) FromString added in v1.8.0 

func (i *SynapseCaVars) FromString(s string) error

func (SynapseCaVars) MarshalJSON added in v1.8.0 

func (ev SynapseCaVars) MarshalJSON() ([]byte, error)

func (SynapseCaVars) String added in v1.8.0 

func (i SynapseCaVars) String() string

func (*SynapseCaVars) UnmarshalJSON added in v1.8.0 

func (ev *SynapseCaVars) UnmarshalJSON(b []byte) error

type SynapseIdxStrides added in v1.8.0 

type SynapseIdxStrides struct {

	// synapse level
	Synapse uint32 `desc:"synapse level"`

	// index value level
	Index uint32 `desc:"index value level"`
	// contains filtered or unexported fields
}

SynapseIdxStrides encodes the stride offsets for synapse index access into network uint32 array.

func (*SynapseIdxStrides) Idx added in v1.8.0 

func (ns *SynapseIdxStrides) Idx(synIdx uint32, idx SynapseIdxs) uint32

Idx returns the index into network uint32 array for given synapse, index value

func (*SynapseIdxStrides) SetIdxOuter added in v1.8.0 

func (ns *SynapseIdxStrides) SetIdxOuter(nsyn int)

SetIdxOuter sets strides with indexes as outer dimension: [Idxs][Synapses] (outer to inner), which is optimal for GPU-based computation.

func (*SynapseIdxStrides) SetSynapseOuter added in v1.8.0 

func (ns *SynapseIdxStrides) SetSynapseOuter()

SetSynapseOuter sets strides with synapses as outer dimension: [Synapses][Idxs] (outer to inner), which is optimal for CPU-based computation.

type SynapseIdxs added in v1.8.0 

type SynapseIdxs int32

SynapseIdxs are the neuron indexes and other uint32 values (flags, etc). There is only one of these per neuron -- not data parallel. 

const (
	// SynRecvIdx is receiving neuron index in network's global list of neurons
	SynRecvIdx SynapseIdxs = iota

	// SynSendIdx is sending neuron index in network's global list of neurons
	SynSendIdx

	// SynPrjnIdx is projection index in global list of projections organized as [Layers][RecvPrjns]
	SynPrjnIdx

	SynapseIdxsN
)

func (SynapseIdxs) Desc added in v1.8.14 

func (i SynapseIdxs) Desc() string

func (*SynapseIdxs) FromString added in v1.8.0 

func (i *SynapseIdxs) FromString(s string) error

func (SynapseIdxs) MarshalJSON added in v1.8.0 

func (ev SynapseIdxs) MarshalJSON() ([]byte, error)

func (SynapseIdxs) String added in v1.8.0 

func (i SynapseIdxs) String() string

func (*SynapseIdxs) UnmarshalJSON added in v1.8.0 

func (ev *SynapseIdxs) UnmarshalJSON(b []byte) error

type SynapseVarStrides added in v1.8.0 

type SynapseVarStrides struct {

	// synapse level
	Synapse uint32 `desc:"synapse level"`

	// variable level
	Var uint32 `desc:"variable level"`
	// contains filtered or unexported fields
}

SynapseVarStrides encodes the stride offsets for synapse variable access into network float32 array.

func (*SynapseVarStrides) Idx added in v1.8.0 

func (ns *SynapseVarStrides) Idx(synIdx uint32, nvar SynapseVars) uint32

Idx returns the index into network float32 array for given synapse, and variable

func (*SynapseVarStrides) SetSynapseOuter added in v1.8.0 

func (ns *SynapseVarStrides) SetSynapseOuter()

SetSynapseOuter sets strides with synapses as outer loop: [Synapses][Vars], which is optimal for CPU-based computation.

func (*SynapseVarStrides) SetVarOuter added in v1.8.0 

func (ns *SynapseVarStrides) SetVarOuter(nsyn int)

SetVarOuter sets strides with vars as outer loop: [Vars][Synapses], which is optimal for GPU-based computation.

type SynapseVars 

type SynapseVars int32

SynapseVars are the neuron variables representing current synaptic state, specifically weights. 

const (
	// Wt is effective synaptic weight value, determining how much conductance one spike drives on the receiving neuron, representing the actual number of effective AMPA receptors in the synapse.  Wt = SWt * WtSig(LWt), where WtSig produces values between 0-2 based on LWt, centered on 1.
	Wt SynapseVars = iota

	// LWt is rapidly learning, linear weight value -- learns according to the lrate specified in the connection spec.  Biologically, this represents the internal biochemical processes that drive the trafficking of AMPA receptors in the synaptic density.  Initially all LWt are .5, which gives 1 from WtSig function.
	LWt

	// SWt is slowly adapting structural weight value, which acts as a multiplicative scaling factor on synaptic efficacy: biologically represents the physical size and efficacy of the dendritic spine.  SWt values adapt in an outer loop along with synaptic scaling, with constraints to prevent runaway positive feedback loops and maintain variance and further capacity to learn.  Initial variance is all in SWt, with LWt set to .5, and scaling absorbs some of LWt into SWt.
	SWt

	// DWt is delta (change in) synaptic weight, from learning -- updates LWt which then updates Wt.
	DWt

	// DSWt is change in SWt slow synaptic weight -- accumulates DWt
	DSWt

	SynapseVarsN
)

func (SynapseVars) Desc added in v1.8.14 

func (i SynapseVars) Desc() string

func (*SynapseVars) FromString added in v1.8.0 

func (i *SynapseVars) FromString(s string) error

func (SynapseVars) MarshalJSON added in v1.8.0 

func (ev SynapseVars) MarshalJSON() ([]byte, error)

func (SynapseVars) String added in v1.8.0 

func (i SynapseVars) String() string

func (*SynapseVars) UnmarshalJSON added in v1.8.0 

func (ev *SynapseVars) UnmarshalJSON(b []byte) error

type TDDaParams added in v1.7.0 

type TDDaParams struct {

	// tonic baseline Ge level for DA = 0 -- +/- are between 0 and 2*TonicGe -- just for spiking display of computed DA value
	TonicGe float32 `desc:"tonic baseline Ge level for DA = 0 -- +/- are between 0 and 2*TonicGe -- just for spiking display of computed DA value"`

	// idx of TDIntegLayer to get reward prediction from -- set during Build from BuildConfig TDIntegLayName
	TDIntegLayIdx int32 `inactive:"+" desc:"idx of TDIntegLayer to get reward prediction from -- set during Build from BuildConfig TDIntegLayName"`
	// contains filtered or unexported fields
}

TDDaParams are params for dopamine (DA) signal as the temporal difference (TD) between the TDIntegLayer activations in the minus and plus phase.

func (*TDDaParams) Defaults added in v1.7.0 

func (tp *TDDaParams) Defaults()

func (*TDDaParams) GeFmDA added in v1.7.0 

func (tp *TDDaParams) GeFmDA(da float32) float32

GeFmDA returns excitatory conductance from DA dopamine value

func (*TDDaParams) Update added in v1.7.0 

func (tp *TDDaParams) Update()

type TDIntegParams added in v1.7.0 

type TDIntegParams struct {

	// discount factor -- how much to discount the future prediction from TDPred
	Discount float32 `desc:"discount factor -- how much to discount the future prediction from TDPred"`

	// gain factor on TD rew pred activations
	PredGain float32 `desc:"gain factor on TD rew pred activations"`

	// idx of TDPredLayer to get reward prediction from -- set during Build from BuildConfig TDPredLayName
	TDPredLayIdx int32 `inactive:"+" desc:"idx of TDPredLayer to get reward prediction from -- set during Build from BuildConfig TDPredLayName"`
	// contains filtered or unexported fields
}

TDIntegParams are params for reward integrator layer

func (*TDIntegParams) Defaults added in v1.7.0 

func (tp *TDIntegParams) Defaults()

func (*TDIntegParams) Update added in v1.7.0 

func (tp *TDIntegParams) Update()

type TopoInhibParams added in v1.2.85 

type TopoInhibParams struct {

	// use topographic inhibition
	On slbool.Bool `desc:"use topographic inhibition"`

	// [viewif: On] half-width of topographic inhibition within layer
	Width int32 `viewif:"On" desc:"half-width of topographic inhibition within layer"`

	// [viewif: On] normalized gaussian sigma as proportion of Width, for gaussian weighting
	Sigma float32 `viewif:"On" desc:"normalized gaussian sigma as proportion of Width, for gaussian weighting"`

	// [viewif: On] half-width of topographic inhibition within layer
	Wrap slbool.Bool `viewif:"On" desc:"half-width of topographic inhibition within layer"`

	// [viewif: On] overall inhibition multiplier for topographic inhibition (generally <= 1)
	Gi float32 `viewif:"On" desc:"overall inhibition multiplier for topographic inhibition (generally <= 1)"`

	// [viewif: On] overall inhibitory contribution from feedforward inhibition -- multiplies average Ge from pools or Ge from neurons
	FF float32 `` /* 133-byte string literal not displayed */

	// [viewif: On] overall inhibitory contribution from feedback inhibition -- multiplies average activation from pools or Act from neurons
	FB float32 `` /* 139-byte string literal not displayed */

	// [viewif: On] feedforward zero point for Ge per neuron (summed Ge is compared to N * FF0) -- below this level, no FF inhibition is computed, above this it is FF * (Sum Ge - N * FF0)
	FF0 float32 `` /* 186-byte string literal not displayed */

	// weight value at width -- to assess the value of Sigma
	WidthWt float32 `inactive:"+" desc:"weight value at width -- to assess the value of Sigma"`
	// contains filtered or unexported fields
}

TopoInhibParams provides for topographic gaussian inhibition integrating over neighborhood. TODO: not currently being used

func (*TopoInhibParams) Defaults added in v1.2.85 

func (ti *TopoInhibParams) Defaults()

func (*TopoInhibParams) GiFmGeAct added in v1.2.85 

func (ti *TopoInhibParams) GiFmGeAct(ge, act, ff0 float32) float32

func (*TopoInhibParams) Update added in v1.2.85 

func (ti *TopoInhibParams) Update()

type TraceParams added in v1.5.1 

type TraceParams struct {

	// [def: 1,2,4] time constant for integrating trace over theta cycle timescales -- governs the decay rate of syanptic trace
	Tau float32 `` /* 126-byte string literal not displayed */

	// [def: 0,1] amount of the mean dWt to subtract, producing a zero-sum effect -- 1.0 = full zero-sum dWt -- only on non-zero DWts.  typically set to 0 for standard trace learning projections, although some require it for stability over the long haul.  can use SetSubMean to set to 1 after significant early learning has occurred with 0.  Some special prjn types (e.g., Hebb) benefit from SubMean = 1 always
	SubMean float32 `` /* 409-byte string literal not displayed */

	// threshold for learning, depending on different algorithms -- in Matrix and VSPatch it applies to normalized GeIntNorm value -- setting this relatively high encourages sparser representations
	LearnThr float32 `` /* 197-byte string literal not displayed */

	// [view: -] rate = 1 / tau
	Dt float32 `view:"-" json:"-" xml:"-" inactive:"+" desc:"rate = 1 / tau"`
}

TraceParams manages learning rate parameters

func (*TraceParams) Defaults added in v1.5.1 

func (tp *TraceParams) Defaults()

func (*TraceParams) TrFmCa added in v1.5.1 

func (tp *TraceParams) TrFmCa(tr float32, ca float32) float32

TrFmCa returns updated trace factor as function of a synaptic calcium update factor and current trace

func (*TraceParams) Update added in v1.5.1 

func (tp *TraceParams) Update()

type TrgAvgActParams added in v1.2.45 

type TrgAvgActParams struct {

	// whether to use target average activity mechanism to scale synaptic weights
	On slbool.Bool `desc:"whether to use target average activity mechanism to scale synaptic weights"`

	// if this is > 0, then each neuron's GiBase is initialized as this proportion of TrgRange.Max - TrgAvg -- gives neurons differences in intrinsic inhibition / leak as a starting bias
	GiBaseInit float32 `` /* 186-byte string literal not displayed */

	// [def: 0.02] [viewif: On] learning rate for adjustments to Trg value based on unit-level error signal.  Population TrgAvg values are renormalized to fixed overall average in TrgRange. Generally, deviating from the default doesn't make much difference.
	ErrLRate float32 `` /* 255-byte string literal not displayed */

	// [def: 0.005,0.0002] [viewif: On] rate parameter for how much to scale synaptic weights in proportion to the AvgDif between target and actual proportion activity -- this determines the effective strength of the constraint, and larger models may need more than the weaker default value.
	SynScaleRate float32 `` /* 289-byte string literal not displayed */

	// [def: 0,1] [viewif: On] amount of mean trg change to subtract -- 1 = full zero sum.  1 works best in general -- but in some cases it may be better to start with 0 and then increase using network SetSubMean method at a later point.
	SubMean float32 `` /* 235-byte string literal not displayed */

	// [def: true] [viewif: On] permute the order of TrgAvg values within layer -- otherwise they are just assigned in order from highest to lowest for easy visualization -- generally must be true if any topographic weights are being used
	Permute slbool.Bool `` /* 236-byte string literal not displayed */

	// [viewif: On] use pool-level target values if pool-level inhibition and 4D pooled layers are present -- if pool sizes are relatively small, then may not be useful to distribute targets just within pool
	Pool slbool.Bool `` /* 206-byte string literal not displayed */

	// [def: {'Min':0.5,'Max':2}] [viewif: On] range of target normalized average activations -- individual neurons are assigned values within this range to TrgAvg, and clamped within this range.
	TrgRange minmax.F32 `` /* 193-byte string literal not displayed */
	// contains filtered or unexported fields
}

TrgAvgActParams govern the target and actual long-term average activity in neurons. Target value is adapted by neuron-wise error and difference in actual vs. target. drives synaptic scaling at a slow timescale (Network.SlowInterval).

func (*TrgAvgActParams) Defaults added in v1.2.45 

func (ta *TrgAvgActParams) Defaults()

func (*TrgAvgActParams) Update added in v1.2.45 

func (ta *TrgAvgActParams) Update()

type USParams added in v1.8.16 

type USParams struct {

	// [def: 0.5] threshold for a negative US increment, _after_ multiplying by the USnegGains factor for that US (to allow for normalized input magnitudes that may translate into different magnitude of effects), to drive a phasic ACh response and associated VSMatrix gating and dopamine firing -- i.e., a full negative US outcome event (global NegUSOutcome flag is set)
	NegUSOutcomeThr float32 `` /* 369-byte string literal not displayed */

	// [def: 2] gain factor applied to sum of weighted, drive-scaled positive USs to compute PVpos primary value summary -- multiplied prior to 1/(1+x) normalization.  Use this to adjust the overall scaling of PVpos reward within 0-1 normalized range (see also PVnegGain).  Each USpos is assumed to be in 0-1 range, default 1.
	PVposGain float32 `` /* 325-byte string literal not displayed */

	// [def: 1] gain factor applied to sum of weighted negative USs to compute PVneg primary value summary -- multiplied prior to 1/(1+x) normalization.  Use this to adjust overall scaling of PVneg within 0-1 normalized range (see also PVposGain).
	PVnegGain float32 `` /* 246-byte string literal not displayed */

	// gain factor for each individual negative US, multiplied prior to 1/(1+x) normalization of each term for activating the OFCnegUS pools.  These gains are _not_ applied in computing summary PVneg value (see PVnegWts), and generally must be larger than the weights to leverage the dynamic range within each US pool.
	USnegGains []float32 `` /* 318-byte string literal not displayed */

	// weight factor applied to each separate positive US on the way to computing the overall PVpos summary value, to control the weighting of each US relative to the others. Each pos US is also multiplied by its dynamic Drive factor as well.  Use PVposGain to control the overall scaling of the PVpos value.
	PVposWts []float32 `` /* 308-byte string literal not displayed */

	// weight factor applied to each separate negative US on the way to computing the overall PVneg summary value, to control the weighting of each US relative to the others.  The first pool is Time, second is Effort, and these are typically weighted lower (.02) than salient simulation-specific USs (1).
	PVnegWts []float32 `` /* 304-byte string literal not displayed */

	// computed estimated US values, based on OFCposUSPT and VSMatrix gating, in PVposEst
	USposEst []float32 `inactive:"+" desc:"computed estimated US values, based on OFCposUSPT and VSMatrix gating, in PVposEst"`
}

USParams control how positive and negative USs are weighted and integrated to compute an overall PV primary value.

func (*USParams) Alloc added in v1.8.16 

func (us *USParams) Alloc(nPos, nNeg int)

func (*USParams) Defaults added in v1.8.16 

func (us *USParams) Defaults()

func (*USParams) NegUSOutcome added in v1.8.16 

func (us *USParams) NegUSOutcome(ctx *Context, di uint32, usIdx int, mag float32) bool

NegUSOutcome returns true if given magnitude of negative US increment is sufficient to drive a full-blown outcome event, clearing goals, driving DA etc. usIdx is actual index (0 = effort)

func (*USParams) USnegFromRaw added in v1.8.16 

func (us *USParams) USnegFromRaw(ctx *Context, di uint32)

USnegFromRaw sets normalized NegUS values from Raw values

func (*USParams) USnegToZero added in v1.8.16 

func (us *USParams) USnegToZero(ctx *Context, di uint32)

USnegToZero sets all values of USneg, USNegRaw to zero

func (*USParams) USposToZero added in v1.8.16 

func (us *USParams) USposToZero(ctx *Context, di uint32)

USposToZero sets all values of USpos to zero

func (*USParams) Update added in v1.8.16 

func (us *USParams) Update()

type UrgencyParams added in v1.8.16 

type UrgencyParams struct {

	// value of raw urgency where the urgency activation level is 50%
	U50 float32 `desc:"value of raw urgency where the urgency activation level is 50%"`

	// [def: 4] exponent on the urge factor -- valid numbers are 1,2,4,6
	Power int32 `def:"4" desc:"exponent on the urge factor -- valid numbers are 1,2,4,6"`

	// [def: 0.2] threshold for urge -- cuts off small baseline values
	Thr float32 `def:"0.2" desc:"threshold for urge -- cuts off small baseline values"`
}

UrgencyParams has urgency (increasing pressure to do something) and parameters for updating it. Raw urgency integrates effort when _not_ goal engaged while effort (negative US 0) integrates when a goal _is_ engaged.

func (*UrgencyParams) AddEffort added in v1.8.16 

func (ur *UrgencyParams) AddEffort(ctx *Context, di uint32, inc float32)

AddEffort adds an effort increment of urgency and updates the Urge factor

func (*UrgencyParams) Defaults added in v1.8.16 

func (ur *UrgencyParams) Defaults()

func (*UrgencyParams) Reset added in v1.8.16 

func (ur *UrgencyParams) Reset(ctx *Context, di uint32)

Reset resets the raw urgency back to zero -- at start of new gating event

func (*UrgencyParams) Update added in v1.8.16 

func (ur *UrgencyParams) Update()

func (*UrgencyParams) Urge added in v1.8.16 

func (ur *UrgencyParams) Urge(ctx *Context, di uint32) float32

Urge computes normalized Urge value from Raw

func (*UrgencyParams) UrgeFun added in v1.8.16 

func (ur *UrgencyParams) UrgeFun(urgency float32) float32

UrgeFun is the urgency function: urgency / (urgency + 1) where urgency = (Raw / U50)^Power

type VSPatchParams added in v1.7.11 

type VSPatchParams struct {

	// [def: 3] multiplier applied after Thr threshold
	Gain float32 `def:"3" desc:"multiplier applied after Thr threshold"`

	// [def: 0.15] initial value for overall threshold, which adapts over time -- stored in LayerVals.ActAvgVals.AdaptThr
	ThrInit float32 `def:"0.15" desc:"initial value for overall threshold, which adapts over time -- stored in LayerVals.ActAvgVals.AdaptThr"`

	// [def: 0,0.002] learning rate for the threshold -- moves in proportion to same predictive error signal that drives synaptic learning
	ThrLRate float32 `` /* 137-byte string literal not displayed */

	// [def: 10] extra gain factor for non-reward trials, which is the most critical
	ThrNonRew float32 `def:"10" desc:"extra gain factor for non-reward trials, which is the most critical"`
}

VSPatchParams parameters for VSPatch learning

func (*VSPatchParams) Defaults added in v1.7.11 

func (vp *VSPatchParams) Defaults()

func (*VSPatchParams) ThrVal added in v1.8.1 

func (vp *VSPatchParams) ThrVal(act, thr float32) float32

ThrVal returns the thresholded value, gain-multiplied value for given VSPatch activity level

func (*VSPatchParams) Update added in v1.7.11 

func (vp *VSPatchParams) Update()

type VTAParams added in v1.8.16 

type VTAParams struct {

	// [def: 0.75] gain on CeM activity difference (CeMPos - CeMNeg) for generating LV CS-driven dopamine values
	CeMGain float32 `def:"0.75" desc:"gain on CeM activity difference (CeMPos - CeMNeg) for generating LV CS-driven dopamine values"`

	// [def: 1.25] gain on computed LHb DA (Burst - Dip) -- for controlling DA levels
	LHbGain float32 `def:"1.25" desc:"gain on computed LHb DA (Burst - Dip) -- for controlling DA levels"`
	// contains filtered or unexported fields
}

VTAParams are for computing overall VTA DA based on LHb PVDA (primary value -- at US time, computed at start of each trial and stored in LHbPVDA global value) and Amygdala (CeM) CS / learned value (LV) activations, which update every cycle.

func (*VTAParams) Defaults added in v1.8.16 

func (vt *VTAParams) Defaults()

func (*VTAParams) Update added in v1.8.16 

func (vt *VTAParams) Update()

func (*VTAParams) VTADA added in v1.8.16 

func (vt *VTAParams) VTADA(ctx *Context, di uint32, ach float32, hasRew bool)

VTADA computes the final DA value from LHb values ACh value from LDT is passed as a parameter.

type ValenceTypes added in v1.7.11 

type ValenceTypes int32

ValenceTypes are types of valence coding: positive or negative. 

const (
	// Positive valence codes for outcomes aligned with drives / goals.
	Positive ValenceTypes = iota

	// Negative valence codes for harmful or aversive outcomes.
	Negative

	ValenceTypesN
)

func (ValenceTypes) Desc added in v1.8.14 

func (i ValenceTypes) Desc() string

func (*ValenceTypes) FromString added in v1.7.11 

func (i *ValenceTypes) FromString(s string) error

func (ValenceTypes) String added in v1.7.11 

func (i ValenceTypes) String() string

Source Files

View all Source files

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL
go.dev uses cookies from Google to deliver and enhance the quality of its services and to analyze traffic. Learn more.