Dynamics of reward-based training of piece-wise linear recurrent neural networks for context-dependent decision making

In this work, we consider an ensemble of recurrent neural networks which are reinforcement-trained for the cognitive-like task of context-dependent perceptual decision making. The systems are characterized by the actor-critic architecture and consist of piece-wise linear neurons in the form of rectified linear units. We study dynamic mechanisms underlying the task completing during trial implementation as well as the dynamics of training, i.e., the process of how particular dynamical and structural aspects emerge in the course of reinforcement learning.