Abstract
Every day, we encounter multiple visual scenes, which contain vast amounts of information that must be selectively attended or inhibited to avoid sensory overload. Here, we sought to understand the selective information contents that underlie successful categorization of faces, objects and scenes within the same images. This approach confers the advantage of isolating the effects to the active observer resolving a categorization task because the input to the visual system is constant across tasks. Five observers each performed five categorization tasks across 4,482 trials (facial expression, identity, general scene, specific scene, and object) on the same set of complex naturalistic selfie images. We decomposed each selfie image with Gabor features at 6 orientations, 7 spatial scales, and 3,108 spatial locations. We kept the top 35% of the Gabor features as specified by power ranking range averaged across all selfies. Then, we randomly sampled 5% of the Gabor features (i.e. ~2,400 Gabor features) to produce a sparse stimulus shown on an individual trial (see Supplementary Figures). Following the experiment, independently for each observer we used binary linear regressions to reverse correlate the single trial relationship between random sampling of Gabor coefficients and the face, object or scene categorization responses in the task. We demonstrate selective task-dependent information compression. On average, observers used 10.38% of available Gabor features for categorizing expressions, 16.94% for identity, 39.62% for objects, 90.22% for general scenes, and 94.56% for specific scenes. Task-dependent information compression reveals the specific face, object and scene features that the brain must differentially represent from the same images to achieve successful categorization behavior. Therefore, as categorization tasks change the information content that the brain must process, they should play a prominent role in explanations of information processing mechanisms in brain and artificial networks.
Meeting abstract presented at VSS 2018