README ¶
Back to All Sims (also for general info and executable downloads)
Introduction
This simplified basal ganglia (BG) network shows how dopamine bursts can reinforce Go (direct pathway) firing for actions that lead to reward, and dopamine dips reinforce NoGo (indirect pathway) firing for actions that do not lead to positive outcomes, producing Thorndike's classic Law of Effect for instrumental conditioning, and also providing a mechanism to learn and select among actions with different reward probabilities over multiple experiences.
This model is based on the connection between dopamine D1 and D2 receptor properties and the direct and indirect pathways as originally described in Frank (2005). The version used here, for simple demonstration purposes, uses the same core code as the PBWM algorithm described in the Executive Function Chapter, which was originally published as O'Reilly & Frank, 2006. It does not include the more detailed aspects of disinhibitory gating circuitry in the BG, which are explored in the published papers and available elsewhere (see below).
The overall function of the model is to evaluate a given action being considered in the frontal cortex, and decide whether to execute that action on the basis of a learned history of reward / punishment outcomes associated with the action. This history of reward / punishment is learned by the model as you run it, based on the frequency of positive vs. negative dopamine signals associated with each action option. For each trial, if there is a dopamine burst (i.e., from a positive outcome), then simulated dopamine D1 receptors in the direct pathway cause the weights to increase for the active neurons in the MatrixGo (direct) pathway of the striatum. Unlike glutamate or GABA, the effect of dopamine is not directly excitatory or inhibitory, but rather depends on the type of receptor and also how much excitatory input is already present. Phasic bursts of dopamine excites those D1 neurons that are already receiving excitatory synaptic input - and thus an increase in dopamine preferentially amplifies activity in active Go neurons (i.e., those that selected the PFC action in response to the stimulus Input); and this increase in activity is associated with LTP. So even though there is no supervised 'target' for training the model what to do, actions can get reinforced if they produce outcomes that are better than expected (i.e. dopamine levels go up and the Go neuron activity increases).
In contrast to the Go units, dopamine inhibits NoGo neurons due to its effects on D2 receptors, and hence these units undergo LTD during rewards. This opposite effect also plays out in the case of dopamine dips (i.e., from a worse-than-expected or even overtly negative outcome). In this case, NoGo neurons receiving cortical input (i.e. those representing the action that was selected) actually become more active when dopamine levels go down (due to removal of the inhibitory D2 effect . This increase in NoGo activity for active neurons causes those specific neurons that would suppress the selected action for this input stimulus to increase their weights, which causes this action to be avoided more in the future. This mechanism of "opponency" allows the basal ganglia to learn and represent both the benefits (predicted reward probability) and costs (predicted probability of negative outcomes) separately, where the level of dopamine in the system can be used to regulate the degree to which choices are made based on benefits or costs (hence affecting risk taking, consistent with effects of dopamine manipulations across species; for more details see Collins & Frank (2014).
Training
- Let's begin by running a few trials -- do
Init
andStep Trial
.
This will present one of the 6 input stimuli, which in turn activates a corresponding action representation in the PFCout layer, representing the superficial layers of PFC neurons (layers 2-3), in this case premotor cortex rather than prefrontal cortex. And then as you continue to cycle, you should see activation in the MatrixGo and MatrixNoGo layers (for visualization convenience, all the connections are localist to units in the same position as the input units). These layers represent the matrisome (aka Matrix, which sounds cooler) medium spiny neurons (MSN's) of the striatum, e.g., in the dorsal region, which are interconnected with the frontal cortical action planning brain areas (both at the level of premotor cortex but also more anteriorly for abstract decisions). As you continue to cycle, activation spreads to the GPeNoGo (globus pallidus external segment) and Thalamus layers. Note that in the actual BG system (and in our more detailed models), Go units don't directly excite the thalamus but instead they inhibit downstream neurons in the globus pallidus internal segment (GPi); GPi neurons are normally tonically active and send inhibitory projections to thalamus.
So instead of exciting thalamus, Go unit activity has a disinhibitory function, enabling the Thalamus to get active by removing tonic inhibition by GPi. In contrast, NoGo units have the opposite effect (they inhibit GPe which in turn inhibits GPi, which itself inhibits Thalamus...). This disinhibition circuitry through has its own set of computational functions (i.e., there is a method to all this madness), but for simplicity here, we abstract away from this circuitry, using one common GPiThal
layer that summarizes the functions of GPi and
Thalamus together. In this abstraction, Go units excite the GPiThal
which directly excites the PFC, whereas NoGo units excite GPe which in turn inhibits GPiThal. This allows us to represent the basic notion of opponency, whereby activity in the direct Go pathway competes with the NoGo pathway for each action by affecting the degree of thalamocortical excitability. You can explore more detailed emergent models with disinhibitory dynamics, which allow one to capture various physiological and behavioral data on Michael Frank's website -- these currently require the extra 7.1 "LTS" package available on the emergent website.
When a given GPiThal unit gets above a threshold level of activation (0.5) (capturing the disinhibition that would be present in the real system) to allow activation to flow from the PFCout layer to the PFCoutD
(D = deep) layer, i.e., deep layer "output" neurons in this frontal area. This is our current understanding of the net effect of BG disinhibition of the thalamus: the thalamus is bidirectionally interconnected with the deep layer PFC neurons, and disinhibiting it allows these deep neurons to become active. These PFC deep layer neurons then project to other areas in the frontal cortex and to other subcortical targets -- for example frontal eye field (FEF) neurons project to the superior colliculus and directly influence saccadic motor actions, while primary motor cortex deep neurons project all the way down to the spinal cord and drive muscle contraction patterns there. Thus, this transition between superficial to deep activation, under control of the BG disinhibition, is the neural correlate of deciding to execute a motor action. As we discuss more in the Executive Function Chapter, in most areas of frontal cortex, this deep-layer activation has more indirect effects that include maintaining a strong top-down activation signal on other areas of cortex, which ultimately guide and shape behavior according to more abstract action plans -- i.e., in most cases it is not as simple as directly activating a set of muscles in response to a stimulus input! Nevertheless, hierarchical extensions of the BG model, where multiple cascading PFC-BG loops interact, have been used to simulate more elaborated action selection processes in which the basic computational function of the BG is similar at each level.
- Proceed to
Step Trial
through more trials to see the variety of behavior that the model exhibits based on its initial random weights.
Learning
Dopamine (DA) from the SNc modulates the relative balance of activity in Go versus NoGo units via simulated D1 and D2 receptors. Dopamine effects are greatest on those striatal units that are already activated by corticostriatal glutamatergic input. Go units activated by the current stimulus and motor response are further excited by D1 receptor stimulation. In contrast, DA is uniformally inhibitory on NoGo units via D2 receptors. This differential effect of DA on Go and NoGo units, via D1 and D2 receptors, directly affects performance (i.e., more DA leads to more Go and associated response vigor, faster reaction times) and, critically, learning, as described above. Specifically, dopamine bursts reinforce Go learning and weaken NoGo learning, while dips have the opposite effects, and these make sense in terms of reinforcing actions associated with positive outcomes, and avoiding those associated with less-positive or negative ones. By integrating reinforcement history over multiple trials this system can also learn which actions are probabilistically more rewarding / punishing than others, allowing it to select the best option among available alternatives.
- In this simple model, we have predetermined the SNc dopamine signals for each action, as you can see by clicking on the
TrainEnv
main.BanditEnv
table, and then on theP
field which shows the probabilities associated with each action. That is, rather than have the model select among multiple actions and learn from their outcomes (as is done in the more elaborated BG models to simulate various other aspects of dopamine on learning and choice), here we just control the proportion of trials each action is associated with a positive or negative dopamine signal, and then monitor what the striatum learns about these probabilities for each action, which is sufficient to make the point.
This environment simulates a simplified version of the probabilistic selection task (Frank, Seeberger & O'Reilly, 2014), where human participants were asked to choose among different Japanese characters in a two-alternative forced-choice task and had to learn which characters were probabilistically more rewarding (e.g., for the A-B pair of characters, A was rewarded 80% of the time while B was only rewarded 20%). Note that participants (and models) can learn that A is the most rewarding, that B is the least rewarding, or both -- you can't tell just by looking at choices among A and B. After initial training on specific pairs of stimuli, we then tested on all different paired combinations (e.g., A is paired with other stimuli that have on average neutral 50% probability, and B is paired with those same stimuli; thus any bias in Go vs NoGo learning will show up as better performance approaching A or avoiding B in these test trials). Critically, this allows one to see the difference between a Go bias for rewarded stimuli vs. a NoGo bias for non-rewarded ones. Empirically, we found that Parkinson's patients off of their medications, who have reduced levels of dopamine, learned more NoGo than Go, while those on their medications learned more Go than NoGo, while age-matched controls were somewhere in between (Figure 1). This basic pattern has now been reported in various other experiments and tasks. We will see that we can explain these results in our simple model.
Figure 1: Data from Frank, Seeberger & O'Reilly (2004), showing that Parkinson's Disease (PD) patients OFF their medications were more likely to learn to avoid the B stimulus that was only rewarded 20% of the time, while PD patients ON their meds learned more to choose the A stimulus that was rewarded 80% of the time. Age-matched control Seniors were more balanced in learning across both cases. These results make sense in terms of PD OFF having low dopamine and favoring D2 dopamine-dip based learning, while PD ON has elevated dopamine from the meds that also "fill in the dips", producing a bias toward D1 burst learning and away from D2 dip learning.
- Do
Train
and click on theActAvg
variable in the NetView -- this shows the running-average activations over many trials, so you can see the effects of learning on the general activations of units in the network. You can click the checkboxes next toMin
andMax
at the top of the NetView to auto-scale the display range to see it better. You can also see a corresponding pattern in the weights into the Go and NoGo units. You can Click on theTrnEpcPlot
orWeights
tabs to make it run faster. The patterns should stabilize after the 30 epochs of training. You can alsoStep Trial
for a few trials with theAct
variable shown, to see how the network is performing on specific trials -- it should correspond with what you observed with theActAvg
averages.
Question 7.1: Describe the patterns of activation that emerge in the
MatrixGo
,MatrixNoGo
, andPFCoutD
layers over learning. Briefly explain why these patterns emerged given the probabilities of dopamine bursts and dips associated with each of the different stimuli / actions (a-f). (Note that in this simulation, the reward probabilities are highest for A, next highest for B and so forth -- unlike the labels in the empirical task where B was the least rewarded action).
You should have observed that the model learned a sensible action valuation representation given the relative dopamine outcomes associated with these actions, similar to the participants in the probabilistic selection task, who were reliably able to choose the more consistently rewarded stimuli over those that were less frequently rewarded. You should also have noticed that while the matrix units encode a more continuous representation of the reward probabilities, the net output of the system reflected a threshold-like behavior that chooses any action that is has more good than bad outcomes, while avoiding those with the opposite profile.
Simulating Parkinson's Disease and Dopamine Medications
- To simulate the reduction in dopamine present in unmedicated Parkinson's disease (PD), set the
BurstDaGain
value to .25 instead of 1 -- corresponding to the roughly 75% damage of DA neurons in PD, and thus weakening the effects of dopamine bursts, while leaving the dips alone (i.e., DA levels are not prevented from decreasing to zero in PD).Init
andTrain
the model again.
Question 7.2: How do the
ActAvg
results from this model withBurstDaGain
=.25 compare to that of the "intact" network from before, withBurstDaGain
=1, in the MatrixGo and NoGo pathways, and the PFCOutD output layer? How does this correspond with the results from PD patients OFF meds, as shown in Figure 1? Recall that the PFCOutD layer reflects the net impact of the learning on action valuation, so units that have highActAvg
correspond to those that the system would deem rewarding on average -- you should notice a difference in how rewarding an action needs to be before the system will reliably select it.
Next, we can simulate effects of DA medication given to PD patients -- for example levodopa increases the synthesis of dopamine. In addition to increasing dopamine availability, medications also continually and directly stimulate dopamine D2 receptors (so-called D2 agonists), which has the effect of blunting the impact of any dips in dopamine (i.e., even when dopamine levels go down, the drugs will continue to occupy D2 receptors and prevent NoGo units from getting excited and learning).
- Set
BurstDaGain
back to 1 to reflect an increase in dopamine availability, andDipDaGain
to .25 (or lower for more extreme effects) to simulate the effects of D2 agonists in continuing to stimulate D2 receptors and hence blocking the effects of dips.Init
andTrain
the model again.
Question 7.3: Now how do the
ActAvg
results compare with both previous runs? How does this correspond with the results from PD patients ON meds, as shown in Figure 1?
One interesting side-effect of PD meds is that a subset of patients develop a gambling habit as a result of taking these meds! This can be explained in the model in terms of the shift in the balance between Go vs. NoGo learning due to the meds -- all those failures to win count for less, and the rare wins count for more.
Although this very simple model can account for the key qualitative features of dopamine-based learning in the BG to promote adaptive action selection, and also for these fascinating patterns of effects in PD patients, there are a number of more complex issues that must be solved to produce a more realistic model of the full complexities of the decision making process that underlies complex behavior. To start, decision making is not just a simple Go/NoGo decision on a single action. The more elaborate models of the BG circuitry (available at Michael Frank's website allow for it to select among multiple actions (where action selection involves both Go and NoGo activity for multiple actions in parallel), and explore i) the differential roles of dopamine on learning vs. choice (i.e., risky decision making), ii) the function of the subthalamic nucleus and the 'hyperdirect' pathway for preventing impulsive actions in response to decision conflict, (iii) the role of cholinergic interneurons for optimizing learning as a function of uncertainty, and (iv) hierarchical interactions among multiple cortico-BG circuits for more advanced learning and abstraction of hierarchical task rules during decision making that supports generalization to new situations.
Beyond these functions, in the real world, the dopaminergic outcomes associated with actions almost never come immediately after the action is taken -- often multiple sequences of actions are required, with outcomes arriving some minutes, hours or even later! In our more complete PBWM (prefrontal-cortex basal-ganglia working memory) model covered in the Executive Function Chapter, we show how these same BG dynamics and learning mechanisms can support maintenance and updating of activation-based "working memory" representations in PFC, to bridge longer gaps of time. Furthermore, properties of brain systems that drive phasic dopamine firing, covered in the RL and PVLV models, help to transfer phasic dopamine signals from firing at the time of later outcomes, to earlier stimuli that reliably predict these later outcomes -- this is helpful for driving action learning to achieve sub-goals or milestones along the way toward a larger desired outcome. Furthermore, we'll see that a synaptic tagging-based trace learning mechanism is very effective in bridging these temporal gaps, and solves a number of different problems that other mechanisms cannot. Another critical element missing from this model is the ability to explicitly represent the nature of the outcomes of different actions, and to reason about these outcomes in relation to factors such as effort, difficulty and uncertainty -- these capabilities require the functions of the orbitofrontal cortex (OFC), anterior cingulate cortex (ACC), and other ventral / medial PFC brain areas, all working in conjunction with these basic BG and dopaminergic systems. Developing such models is at theforefront of current research.
References
Collins, A. G. E., & Frank, M. J. (2014). Opponent actor learning (OpAL): modeling interactive effects of striatal dopamine on reinforcement learning and choice incentive. Psychological Review, 121(3), 337–366. Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/25090423
Frank, M. J. (2005). Dynamic dopamine modulation in the basal ganglia: A neurocomputational account of cognitive deficits in medicated and non-medicated Parkinsonism. Journal Of Cognitive Neuroscience, 17, 51-72.
Frank, M. J., Seeberger, L. C., & O'Reilly, R. C. (2004). By carrot or by stick: Cognitive reinforcement learning in Parkinsonism. Science, 306(5703), 1940-1943.
O'Reilly, R. C., & Frank, M. J. (2006). Making working memory work: A computational model of learning in the prefrontal cortex and basal ganglia. Neural Computation, 18(2), 283-328. Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/16378516
Documentation ¶
Overview ¶
bg is a simplified basal ganglia (BG) network showing how dopamine bursts can reinforce *Go* (direct pathway) firing for actions that lead to reward, and dopamine dips reinforce *NoGo* (indirect pathway) firing for actions that do not lead to positive outcomes, producing Thorndike's classic *Law of Effect* for instrumental conditioning, and also providing a mechanism to learn and select among actions with different reward probabilities over multiple experiences.