Performing decision-making tasks through dynamics of recurrent neural networks trained with reinforcement learning

In this work, recurrent neural networks are considered as functional models that perform target tasks inspired by cognitive neuroscience experiments. These networks are trained with reinforcement learning methods, after that structure and dynamics mechanisms underlying their behaviors are studied. We use two versions—with and without a context signal—of the cognitive perceptual decision-making task. Population and single-neuron dynamics are studied resulting in stimulus processing as well as successful completing the target tasks. Functionally specialized neurons as well as cluster structure of the trained networks are found and investigated. Similarities and differences between the model neural networks and biological prototypes are discussed.