The neural network of the basal ganglia (BG) is commonly viewed as two functionally related subsystems. These are the neuromodulators subsystem and the main-axis subsystem, in analogy with the critic-actor division of reinforcement learning agent. The neuromodulators calculate and hand out the temporal difference or the prediction error signal to the main axis subsystem which includes connections between all cortical areas, the striatum and the subthalamic nucleus and the BG output structures. The basal ganglia modify behavior through their inhibitory projections to the thalamic-frontal cortex networks and to brain-stem pre-motor nuclei.
We propose that the BG main axis is performing dimensionality reduction of the cortical input. However, this compression process is not trying to maximize the representation of the cortical inputs. Rather the BG computational goal is to find the optimal trade-off between minimization of the complexity of the representation of the current state and past memories (cost and/or generalization) while maximizing the expected (discounted) future reward (information bottleneck). This can be achieved with a biological inspired model of the BG neural network that includes value representations in the cortico-striatal axons, lateral inhibition between striatal neurons, D1 dependent triple Hebbian learning at the cortico-striatal synapses, and D1 and D2 opposite modulations of striatal excitability. The model replicates most normal and pathologic BG and dopamine related behaviors (including such as paradoxical kinesia and the immediate effects of ultra fast dopamine agonist) by the same set of equations.
In line with the information bottleneck dimensionality reduction model, BG main axis neurons maintain flat spike crosscorrelation functions, diverse responses to behavioral events, and broadly distributed values of signal (response) correlations with zero population mean. The spontaneous and the evoked activity of BG dopaminergic and cholinergic modulators (critics) are significantly more correlated than that of the BG main axis (actor) neurons. Following MPTP induced dopamine depletion and the development of Parkinsonian symptoms, the BG activity turned into an oscillatory and synchronized mode leading to a significant reduction of their information capacity and abnormal gating of the cortical state-to-action mapping. Finally, we propose that levodopa induced dyskinesias are due to both the direct modulation of striatal excitability by the dopamine replacement therapy (DRT), and the rundown of cortico-striatal synaptic plasticity due to the inappropriate pulsatile DRT.
BACK
