Computational and Theoretical Neuroscience Seminar

Kohitij Kar (postdoc,MIT,DiCarlo Lab) will present "Recurrent computations during visual object perception-investigating within and beyond the primate ventral stream."
Email d.shipman@duke.edu for Zoom link.
Partial Abstract: Recurrent circuits are ubiquitous in the primate ventral stream that supports core object recognition - primates' ability to categorize objects rapidly. While recurrence has long been thought to be functionally relevant to visual processing, this has remained mostly a motivating idea and challenging to probe in the visual system mechanistically; especially at the shorter time scales (<200 ms) of core object recognition. Our work has achieved three advances that help demonstrate and localize the functional importance of recurrent computations during object recognition. Advance 1: We compared primate behavior with that of the feedforward deep convolutional neural network models of object recognition. This led to the discovery of many "challenge" images that are easily categorized by primates, but not by the current models. A large-scale electrophysiological investigation across the macaque ventral visual pathway revealed that putative feedback-related neural signals are critical for solving these "challenge" images by the primates, one that is strikingly missing from most feedforward artificial neural network models.