Population dynamics of context-dependent computations in recurrent network models

Abstract

The mechanism underlying the way in which the brain coordinates the integration and representation of relevant and irrelevant sensory inputs based on a context has long eluded neuroscientists. In recent years, advances in cognitive modeling using artificial neural networks have allowed theoreticians to gain deeper insight into the neurological processes that give rise to complex behaviors. In this work, we demonstrate that a randomly initialized, unstructured recurrent network model can be trained to dynamically encode contextually coded stimuli, represent these stimuli in the form of experimental recordings from prefrontal cortex, and decode these representations to generate biologically realistic psychophysical behavioral responses. We then decompose the model to understand how this process unfolds, treating the model as a high-dimensional, dynamical process. We consider past proposed mechanisms underlying context-dependent feature integration, and amend and extend previous theories for how this process occurs at the population level of neurons. We find that our model integrates both relevant and irrelevant sensory information and that learned recurrent dynamics must accommodate the presence of both relevant and irrelevant stimuli in the state space of the network in order to perform relevant feature selection. Our work provides insight into a potential novel mechanism underlying context-dependent computation.