Information content of visual representations depends on attentional priority and working memory load

Though our experience of the world often appears rich and detailed, less information is immediately accessible than our intuition would suggest. When viewing a complex scene – such as a crowded city street – information about irrelevant features of the environment is lost due to noisy neural processing, while information about relevant features (those selected by visual attention) is spared. Similarly, behavioral experiments demonstrate that when even modest amounts of information must be held briefly in mind (in visual working memory), the amount of available information about each item is diminished, and this available information decreases with increasing information load. In what manner do visual representations across large-scale neural activity patterns support these behavioral information processing limits? In three studies, we examined the fidelity with which human cortical neural activation patterns measured with functional magnetic resonance imaging represent visual information. To this end, we developed a novel analysis technique whereby we reconstruct images of visual stimuli using neural activation patterns measured over entire brain regions. Using this technique, we established that the neural representation of a relevant visual stimulus is enhanced in its amplitude over a noisy baseline in several visual and parietal cortical regions, suggestive of an increase in the representation’s information content. Subsequently, we demonstrated that under conditions where no information is available in a display, the maintenance of a larger number of items in visual working memory is accompanied by a degradation in each item’s representation amplitude, indicative of lower population-level information content. Finally, we evaluated the relationship between these two findings by directing participants to attend to one of several items held in visual working memory. Surprisingly, we discovered that degraded representations can recover with visual attention, and the degree of recovery was related to behavioral task performance. Such recovery of degraded information suggests that additional information must be available to the system but invisible to our measurements before attention is allocated. Together, these results demonstrate that behavioral limits on information processing are related to the fidelity with which visual information is represented in large-scale neural codes