Speaker: Marcelo Magnasco
Professor and Head of Lab of Integrative Neuroscience, Rockefeller University
Title: Signal propagation in recurrent networks
When the linear component of the dynamics of a system is marginally stable, the first nonlinear term controls stability. As a result, small inputs may cause large effects, and can directly affect characteristics such as relaxation times and decay lengthscales. In a spatially-extended system, this raises the possibility that one may use the input itself to pattern signal propagation paths, thereby controlling dynamically the flow of information and setting up ephemeral “circuits”. We illustrate this idea by using a single layer convolutional recurrent network, whose connectivity matrix is unitary, thereby having all of its eigenvalues in the unit circle; we use an input with two components, a steady state input creating “walls”, and an oscillatory input applied to one single site, which then propagates through the network. The ability of critical recurrent networks to generate input-dependent reconfigurable circuitry may shed light on the dynamical gating of signal propagation in the brain and other complex biological networks.