Neural Circuits for Visual Working Memory

Latent representations are critical for disambiguating the sensory world1 and guiding perceptual decisions. Visual working memory is often used to study these latent representations, but the associated neural activity patterns, their maintenance, and their distribution across the brain, remain contentious. One difficulty has come in disambiguating the neural representations underlying working memory from confounding variables introduced by the task environment. We therefore investigated visual working memory in mice alternating between performing a delayed (non)match-to-sample working memory task and a simple Pavlovian discrimination task. This experimental design isolated visual working memory engagement as the only independent variable, separable from activity associated with sensory input, movement, and reward. Transient optogenetic silencing of different cortical areas revealed a selective role of highly distributed areas of the neocortex for working memory maintenance. Neural population activity in some of these areas, namely higher visual area AM and premotor area M2, during the inter-stimulus delay period was dominated by orderly low-dimensional dynamics, which we found to be completely independent of working memory engagement. In contrast, by taking advantage of our alternating task design, we were able to decode a high-dimensional population representation of visual working memory, which was (1) present in distributed cortical areas, (2) persisted throughout the inter-stimulus delay period, and (3) predicted correct responses to the subsequent stimulus during the working memory task. Given the recruitment of such distributed neocortical representations during working memory engagement, and having observed that silencing any single area disrupted working memory, we hypothesized that these representations were instantaneously interdependent (‘bound’) by cortical feedback loops. We tested this hypothesis directly by silencing a source cortical area while recording the feedback it received from a reciprocally connected target area. We found that transiently breaking the cortical feedback loop at the onset of the working memory delay had little effect on the low-dimensional dynamics, but selectively abolished representations of visual working memory. Our findings identify reciprocal inter-areal cortical feedback loops as key circuit motifs underlying the maintenance of distributed and high-dimensional latent representations of visual working memory.