pcore

pcore

This is a simple test of the pcore model of basal ganglia (BG) function -- see link for details on the algorithm and expected behavior.

This test model has all of the standard PFC layers, which are kept busy by predicting a sequence of input values on the In layer, via the deep predictive learning mechanism (InP is the pulvinar layer representing the prediction of In). This prediction task is completely orthogonal from the gating decision made by the BG, which is driven by the ACCPos and ACCNeg layers.

These ACC layers have PopCode representations of values, and the BG gating is trained to gate when Pos > Neg. Typically you do TrainRun or Step with the step set to Run, and then TestRun which will run through all combinations of ACCPos (outer loop) and ACCNeg (inner loop), with 25 samples of each value to get statistics (it takes a while). Click on TestTrialStats Plot to see the results -- you can click on ACCPos and ACCNeg to see those inputs, and then compare Gated with Should to see how the network performed.

results

Training data shows close match between Gated and Should (high Match proportion).

Testing data over ACC Pos (outer loop) and ACC Neg (inner loop) shows increasing probability of gating as Pos increases, and reduced firing, and slower RT, as Neg increases, closely matching the target Should behavior. 25 samples of each case are performed, so intermediate levels indicate probability of gating. Model shows appropriate probabilistic behavior on the marginal cases.