Zachary P. Kilpatrick, PhD (Department of Applied Mathematics, University of Colorado Boulder, USA)
Title: Neuromechanics of working memory errors: a neural field approach
Abstract: We develop and analyze neurocomputational models to account for systematic errors in visual working memory tasks. Our framework is a neural field model with lateral inhibitory connectivity that can sustain one or more bump attractor solutions. The centroid of each bump represents the location of a remembered item. Stochastic forcing corresponding to voltage and synaptic fluctuations causes bumps to wander, describing normally distributed saccade errors observed in behavioral experiments. We also account for recent behavioral data from monkeys performing a sequence of working memory trials, demonstrating the location of the target on the previous trial produces an attractive bias of the response in the current trial. Our model proposes short term facilitation as a synaptic mechanism whereby the bias from the previous trial propagates to the current trial. Multiple timescale methods can be used to develop a low-dimensional projection for the location of the bump in the current trial, and this low-dimensional model can be linked to a suboptimal inference process where an observer assumes the previous trial predicts the next. In addition, we discuss a model of multiple item working memory in which multiple bumps are instantiated in the same neural field layer, each representing a remembered item. Attraction, repulsion, and annihilation events arising from bump interactions lead to systematic errors that can account for increases in response error as more items are added. We develop an interface method that allows for a low-dimensional projection of N bumps to a set of 2N stochastic differential equations for the bump interfaces, accurately describing the dynamics of the full system.
November 29, 10:30am, Byron “Beige” Y506
