Abstract
Several recent findings have indicated that the core object recognition is primarily solved through the feedforward sweep of visual information processing. On the other hand, while recurrent connections are ubiquitous in our visual system, their role in object-processing is not yet fully understood. Here, we investigated the contribution of recurrent processes in object recognition under a prevalent challenging condition, that is when objects are occluded by other natural or artificial occluders in the environment. To characterize neural dynamics of object recognition under occlusion, we acquired magnetoencephalography(MEG) data (N=15 subjects), while subjects were presented with images of objects with 0%(no-occlusion),60% and 80%occlusion –with and without backward-masking. We provide evidence from multivariate analysis of MEG data, behavioral data, and computational modelling, demonstrating an essential role for recurrent processes in object recognition under occlusion. First, multivariate analysis of MEG data showed that object discrimination is significantly delayed (by~60ms) under occlusion compared to the no occlusion condition(p< 10^-4, two-sided signrank-test), likely due to the additional time needed for recurrent processes. Second, temporal generalization analysis (King & Dehaene,2014), which provides information about temporal organization of information processing stages, showed that initial sensory signals undergo a relatively long sequence of processing stages that involve recurrent interactions to establish a discriminative representation of occluded objects. Third, backward-masking which is thought to disrupt recurrent processes, impaired MEG object-discrimination time-courses, and subjects' behavioral performances only under occlusion. Fourth, a feedforward CNN failed to explain the MEG data and the behavioral data when objects were occluded; however, a CNN with local recurrent connections reached the human-level performance under occlusion, and partially explained the MEG data for occluded objects. Taken together, our empirical results suggest an essential role for recurrent processing when objects are occluded, and our computational model with local recurrent connections explains how our brain might be solving this problem.
Meeting abstract presented at VSS 2018