Discrete communication mediates effective regularization in chaotic recurrent networks

Disordered networks with discrete signaling are considered a poor substrate for computation, yet they are ubiquitous in the brain. We show that such large chaotic networks can support reliable computation, with a surprisingly long working memory. To this end, we reformulate the recurrent network’s activity in terms of an effective kernel. In particular, we find that 1) chaos in a discrete signaling network acts as an effective regularizer, 2) rich and robust computation is possible in the chaotic regime and 3) at the edge of chaos, reliable computation persists even longer.