Abstract
Early fMRI investigations of visual working memory identified a spatially localized, load-dependent increase in BOLD activity in the intraparietal sulcus, and this region has been hypothesized to play a particularly important role in working memory maintenance. However, emerging evidence suggests that working memory codes are widely distributed across cortex. For example, the identity of a single remembered item can also be decoded from activity patterns in early visual cortex. Here, we looked for evidence of working memory load signals across the visual stream, while ruling out potential sensory confounds that had not been previously controlled. In a pre-registered fMRI experiment, n = 12 human observers (t = 12 scanner hours per observer) performed a visual working memory task with set sizes 0, 1, 2, and 4. On each trial, observers encoded 4 items (colored squares). After a short interval (.8 sec), observers were given a retro-cue indicating which items should be prioritized, remembered, or dropped from memory. We replicated a classic load-dependent increase in BOLD activity with set size in the intraparietal sulcus. However, we found evidence against the notion that working memory load signals are spatially localized within IPS. Rather, we observed diverse univariate and multivariate load signals across the entire visual stream. For example, we observed a load-dependent decrease in univariate activity in visual cortex (V2, V3, and V3AB). Despite opposing univariate effects, we were able to decode working memory load from multivariate patterns of activity in V1 through IPS in a sustained fashion, and the dimensionality of population activity increased with set size in V1-IPS. Our results highlight the importance of looking beyond monotonic increases in univariate activity and point to future avenues for understanding the distributed nature of working memory codes.