Recurrent computations for pattern completion

Summary

Date: 
February 10, 2016 - 1:00pm
Location: 
Northwest 243
About the Speaker
Name: 
Gabriel Kreiman (Children's/HMS)

Joint work with Hanlin Tang, Bill Lotter and David Cox.
Making inferences from partial information constitutes a critical and ubiquitous aspect of cognition. In the context of visual recognition, pattern completion leads to sophisticated interpolations from images containing large amounts of noise or heavily occluded objects. We hypothesized that the ability to perform visual pattern completion is implemented in neural circuits by recurrent computations. We tested this hypothesis at the behavioral, physiological and computational levels, the results from these tests are consistent with a mechanistic explanation of pattern completion instantiated by recurrent computations. First, pattern completion behavior was robust even for images that contained less than 10% of visible pixels. Performance was significantly deteriorated by the presentation of a rapid backward mask, which is presumed to interrupt feedback and recurrent computations. Second, invasive physiological responses along the human ventral cortex also showed pattern completion and were delayed with respect to the responses to whole objects; this physiological delay was correlated with the behavioral effects of the backward mask. Third, the representation in state-of-the-art purely bottom-up computational architectures were not able to categorize objects from partial information but a computational model implementing all-to-all recurrent connectivity at the top layer was able to approximate human performance levels overall and at the level of individual images. Taken together, these results provide strong arguments of plausibility for the role of recurrent computations in visual inferences from partial information.