Abstract
How is working memory (WM) information represented in the brain? Neural and computational models have used data aggregated over hundreds of trials to argue for different perspectives on how neural activity encodes individual memories. The two main perspectives are information rich representations such as in probabilistic coding models (a probability distribution over the whole feature space), and information sparse representations, such as in high-threshold ( a precise feature value) or drift models (a value with a confidence interval unrelated to the direction of drift). The use of aggregate data represents a key inferential bottleneck that critically limits the ability to adjudicate between different formats of individual memory coding in WM. This study used a powerful method to link behavioral and neural estimates of WM representation on individual trials. We asked participants (n = 12) to memorize a motion direction over a brief delay. After the delay, instead of making a single report about the memorized direction, they indicated their memory by placing 6 “bets”, resulting in a distribution over the 360° direction space that reflected their probabilistic memory representation on individual trials. Additionally, we used a Bayesian decoder to estimate the posterior of the memorized direction given the fMRI signal during memory maintenance on individual trials. Comparing the shape of the behavioral and neural estimates on individual trials, we found a significant correspondence in their width in occipital, parietal and frontal regions (ps < .007; Cohen’s ds > .767), and critically, a significant correspondence in their asymmetry in early visual cortex (p < .001; Cohen’s d = .779). These results indicate (1) individual WM representations are complex probability distributions that contain more information than that can be deduced from aggregate data; (2) early visual cortex contains richer information about WM than other brain regions, with meaningful asymmetry information influencing behavior.