Cortical control of working memory prioritization

Humans distribute working memory (WM) resources across items according to their behavioral relevance. Prioritized items are recalled with better precision and less uncertainty. Here, we test the hypothesis that human cortex represents the mnemonic uncertainty of items using a probabilistic neural code whose gain is modulated according to priority. Using fMRI, we scanned participants while they remembered the locations of two targets whose priorities were precued. Priority was operationalized as the probability with which the target would be the goal of a memory-guided saccade generated after a long 12 second delay. Using Bayesian decoding, we estimated the location and uncertainty of each item in WM simultaneously by modifying an existing model of neural uncertainty (van Bergen et al., 2015; Li, Sprague et al., 2021). To do so, we assumed that activity evoked by the two targets was a weighted sum of the activity to each presented alone, and the weights were gain factors based on each target’s priority. Supporting our hypothesis, in visual and parietal cortex, we found that low-priority targets were associated with lower gain factors, and the high-priority targets were decoded with smaller errors and lower uncertainty. Moreover, the difference between the decoded uncertainty of the high- and low-priority targets predicted the degree to which participants prioritized the targets behaviorally. To identify the brain regions that control how WM resources are allocated, we conducted a whole-brain GLM with trial-by-trial decoded uncertainty as regressors. Remarkably, we found that neural activity in multiple areas across temporal, parietal and frontal cortex predicted decoded memory uncertainty in higher-level visual cortex. These results support a model in which activity in association cortex is the source of feedback signals that sculpt the gain of WM representations in visual cortex according to behavioral relevance.