Recorded Seminar Link: https://wse.zoom.us/rec/play/u4lqnM8DZSTXv5299fmtZNUitejWRcwNy3tDL2lA90SUJEgSbw3ZSHkQ8nbCCmSizPfkWTbnFPQ9ttuh.SpUYfG63jZAUVURm?startTime=1631635165000&_x_zm_rtaid=PP4fIybxTayNnZ6b19hpsA.1632842736978.4df5c32815f94f45bb3f09ff8fad339d&_x_zm_rhtaid=584
?Population Dynamics in Neural Systems?
Sara A. Solla, PhD
Abstract: The ability to simultaneously record theactivity from tens to thousands and maybe even tens of thousands of neurons hasallowed us to analyze the computational role of neural population activity asopposed to single neuron activity. Recent work on a variety of corticalareas suggests that neural function may be built on the activation ofpopulation-wide activity patterns, the neural modes, rather than on theindependent modulation of individual neural activity. These neural modes, thedominant covariation patterns within the neural population, define a lowdimensional neural manifold that captures most of the variance in the recordedneural activity. We refer to the time-dependent activation of the neural modesas their latent dynamics, and argue that latent cortical dynamics within themanifold are the fundamental and stable building blocks of neural populationactivity.
Bio: Sara Solla’s interest is in the brain asa device for the acquisition, storage, transmission, and processing ofinformation. Her work is theoretical; it combines numerical modeling withanalytic and conceptual tools from statistical physics, information theory, andnonlinear dynamics.
At the systems level, they work with neural network models consisting of arraysof highly interconnected nonlinear units that incorporate salient features ofbiological neurons. One of their projects has led to the development of amodular neural network that iscapable of executing an oculomotor delayedresponse task. Damage experiments in this simulated network attempt toreproduce the deficient performance of schizophrenic patients in this task, asmeasured by their colleague, Sohee Park, at the Department of Psychology.
In addition totheir ability to model neuronal processing at the systems level, neural networkmodels provide powerful computational tools for pattern recognition andnonlinear control. Current applications include linkage analysis of geneticdata (in collaboration with the Laboratory of Statistical Genetics, headed byJurg Ott, at Rockefeller University) and the control of springback angle andmaximum strain in the manufacture of sheet metal parts (in collaboration withJianCao from the Department of Mechanical Engineering).
Neural networkmodels provide a prototype for the investigation of systems that interact withthe environment through the execution of an action in response to a stimulus.If the action generates an error signal that is used by the system to modifyits internal state, the system exhibits learning and adaptation capabilities.Much of Dr. Solla’s work in recent years has focused on learning andadaptation; she is currently involved in studies of the dynamical properties ofonline algorithms for supervised and unsupervised learning.