Our Schedule

Res-03: Neural Dynamics, Recurrent Neural Networks and the Problem of Time

Much of the information the brain processes and stores is temporal in nature—a spoken word or a handwritten signature is defined as much by how it unfolds in time as by its spatial structure at any moment. The brain seamlessly assimilates and process temporal information, an ability that is critical to most behaviors: from reward anticipation to sensorimotor processing. We have proposed that timing on the scale of milliseconds to seconds relies on the inherent dynamics of recurrent neural networks (RNNs). And more generally, that the neural dynamics of RNNs represent a fundamental modus operandi for neural computation. Under this view information is stored and generated by dynamic attractors—locally stable neural trajectories. Thus, in contrast to the conventional view that memories are stored in static fixed-point attractors, under this view, many computations emerge from the voyage through neural state space as opposed to the destination.