Abstract
Working memory is imperfect and dynamic – memories becomes less accurate with time. We combine psychophysical methods and computational modeling to precisely describe the dynamic forces governing working memory. We show that the accumulation of error in working memory over time is not random but reflects underlying attractor dynamics. Furthermore, these dynamics are modulated by memory load and they are modified by experience: memories drift towards common stimuli. To identify these attractor dynamics, we trained both monkeys and humans to perform a continuous-report working memory task. In brief, subjects were asked to remember 1 to 3 colored squares. After a variable memory delay, subjects had to report the color of a cued stimulus on a continuous color wheel. Dynamic models fit to the behavioral responses of both monkeys and humans showed random diffusion dominated the dynamics at low memory load but strong attractors dominated dynamics at high memory load. In other words, memory dynamics are load-dependent: attractor depth increased with the number of items in working memory. These results provide a mechanistic explanation for why increasing the number of items in memory impairs memory accuracy and provide a bridge between descriptive models of memory reports and neural network models of working memory.
Meeting abstract presented at VSS 2018