Abstract
Models of working memory (WM) suggest that individual memories are maintained as complex distributions over feature space that encode both the memory and the memory uncertainty. However, there is little empirical evidence to support this idea, and contemporary research (either implicitly or explicitly) assumes that WM representations encode discrete points in feature space. To investigate whether WM representations are probabilistic, we asked whether memory uncertainty measured using a betting paradigm was reflected in neural measures of WM representations measured with fMRI. Subjects memorized the direction of a dot-motion stimulus. After a delay, instead of making a single report about the direction of motion that they were maintaining, subjects placed 6 “bets” about the memorized direction, resulting in a distribution over 360° direction space that reflected subjects’ memory uncertainty. We used a linear classifier to identify WM representations in visual cortex. Instead of simply decoding the maintained direction of motion (i.e., the direction indicated by subjects’ first bet), we examined whether the classifier evidence captured the complexity of the memory uncertainty. Subjects’ distributions were often shifted either clockwise or counter-clockwise from the first response. Does this aspect of the uncertainty distribution reflect meaningful information above and beyond the initial response? If so, classifier evidence should be greater on trials where the bet distribution was shifted towards the nearest class than on trials where the bet distribution was shifted away from the nearest class. Indeed, classifier evidence was significantly greater when the distribution was shifted towards the nearest class, despite these trials having first bets that were further away from the nearest direction class. These results demonstrate that visual cortex stores rich and detailed information about WM memoranda that extend beyond a simple point estimate, thus lending empirical support for the notion that WM representations are stored as probability distributions.