Alessia Cavallo: No need for noise - Chaos-driven exploration in a computational cortico-striatal circuit model

BCCN Berlin / Technische Universität Berlin

Abstract

 

Computational models from recent decades have been crucial in the modern interpretation of learning supported by the cortex-basal ganglia circuit. However, traditional models are often limited in their informative value in several ways. Firstly, most models cannot explain the acquisition of new behaviors but only the correct choice from a pre-existing repertoire of behaviors. Secondly, the exploration needed for learning originates from stochastic activity of neural units which is at odds with empirical evidence indicating the deterministic nature of neurons. We propose a neural network model of the cortico-striatal circuit that allows the acquisition of actions that are not hard-wired into the model. Moreover, it uses deterministic chaos instead of stochastic noise as source of exploration. The model consists of a chaotic recurrent neural network (RNN) representing the cortex and an inhibitory RNN representing the striatum divided into a direct and indirect pathway population. We show that the direct and indirect pathway have opposing roles in controlling cortical chaoticity and thereby adaptively balance states of exploitation and exploration. Our study opens new horizons to modeling cortex-basal ganglia circuits, paving the way for unprecedented opportunities to study neural and artificial learning. In the future, the model can be extended to investigate functional consequences of pathological circuit alterations from neurological disorders like Parkinson’s disease.

 

Additional Information

Master Thesis Defense

 

Organized by

Prof. Dr. Wolf-Julian Neumann   & Dr. Andrea Mattera   / Lisa Velenosi