Reservoir computing for chunking and decision-making

Abstract:

Chunking refers to a process of dividing a sequence into discrete sections that are easier to analyze. Chunking is crucial for many higher cognitive functions, but how the brain acquires such discretization in an unsupervised fashion remains unknown. A widely-believed hypothesis assumes that the brain learns to predict recurring patterns in input based on its statistical structure. However, this hypothesis was challenged by recent experiments. We propose a novel mechanism in which neural systems learn to predict their responses to input, but not the input per se. We formulate this mechanism in a system of mutual supervising modules of reservoir computing, and show an interesting similarity to “stop cells” that emerge in the basal ganglia after motor habituation. Next, if time allows, I will introduce our recent attempts to unravel the cortical mechanisms underlying ambiguous decision-making behavior. The rats were trained on an alternative choice task under unfamiliar sensory conditions. Both behavioral responses and neural responses in the medial frontal cortex exhibited considerable variances across animals. We trained a reservoir computing system with reinforcement learning and replicated the observed spectrum of individual differences.