The model illustrates how novelty is initially processed by the basal ganglia and then transferred to the cerebellum.
The model learns a spinal-cord > spinal-cord mapping, from the input on the lower-right to the output on the lower-left. It will be trained to associate a 'Go' signal with an output of 1 and 'No Go' signal with output of 0.
Initially both signals have no learned output; they both give an output of 0. We can see this by pressing the 'Test: No Dopamine' button. In this mode the 'Basal Ganglia (GPi)' neuron is clamped to 0 and so it exerts no influence on the network. In this mode, prior to training, notice that the same output is produced for activation of 'Go' and 'No Go' neurons; both produce zero output.
We can then train the model to produce the desired mapping from 'Go' to 1 and 'No Go' to 0, initially using the dopamine system. On a trial of length 200 ('trial length' is number of iterations), the network is exposed to both patterns for 100 iterations. For each of the two patterns an appropriate target value is set in the 'Target' neuron. The 'Dopamine' neuron then registers the difference between the target value and the output value, i.e. the error. This error then influences and 'corrects' the output via left Thalamus-Cortex-Spinal Cord connections.
Dopamine activity is visible in the blue time-series, and output activity in the red time series.
Notice that the 'Dopamine' neuron affects output through the left 'To Spinal Cord' neuron. The right 'To Spinal Cord' neuron is how the cerebellum affects output. The cerebellum affects output through activation of the red nucleus and the major cerebellar neurons (DCN, Inferior Olive, Purkinje, Granule Cells).
Over time, the input output mapping will be transferred to cerebellum. This happens via learning at the parallel fibers, the connections from the granule cells to the Purkinje cell, based on the activity of the outgoing synapses from the Purkinje cell (i.e Climbing Fibers (CF)). As Peck, Streeter, and Kolsloki say: 'Parallel fiber synapses undergo LTD when co-active with climbing fiber activity and LTP when active without coincident climbing fiber activity.'
To see this in action, press the 'Train: 10 Trials' button. This causes 10 trials of dopamine-mediated learning to occur. After the 10 trials are done, press the 'Test: No Dopamine' button. This inhibits dopamine-mediated learning. But notice that the correct mapping now occurs, and is mediated by the cerebellar circuit. The cerebellum has acquired the correct mapping.
Note: The Go and No Go nodes represent sparsified outputs of the granule cells hypothetically produced by more complex afferent projections. There are other simplifications and some misleading features of this simulation. We hope to update it in a future release.