Kohitij Kar
Title: Recurrent computations during visual object perception—investigating within and beyond the primate ventral stream
Abstract
Recurrent circuits are ubiquitous in the primate ventral stream, that supports core object recognition — primate’s ability to rapidly categorize objects. While recurrence/feedback has long been thought to be functionally important to visual processing, this has remained mostly a motivating idea and very difficult to mechanistically demonstrate and probe in the visual system; 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 tasks.
Advance 1: Using extensive behavioral testing, we have recently discovered a set of putative recurrent computation dependent challenge images where object categorization is specifically difficult for nonrecurrent (exclusively feed-forward) deep artificial neural networks (ANNs), but are nevertheless easy for primates (both humans and macaques). Interestingly, simultaneous large-scale measurements of image-by-image neural population response dynamics in the macaque inferior temporal (IT) cortex revealed that, relative to primate-behavior matched control images, a linearly-available solution to these challenge images develop ~30 ms later in IT. This finding, along with a series of control experiments, suggested that the delay is due to recurrent processing — not yet captured by deep ANN models of the ventral stream. However, we do not yet know which brain circuits are most responsible for these additional, recurrent computations: circuits within the ventral stream? within IT? outside the ventral stream? all of the above?Advance 2: Based on the presence of object-category selective neurons, and anatomical projections to IT, we hypothesized that the ventral prefrontal cortex (PFC) might be a critical recurrence node in the development of late phase object representations for challenge images in IT. To test this hypothesis, we silenced (via muscimol) 10mm2 of ventral PFC and simultaneously measured IT population activity (with chronically-implanted Utah arrays) while monkeys performed a battery of core object recognition tasks. Our results show that muscimol injections (10 ul) in ventral PFC produced a significant overall behavioral deficit. Interestingly, this deficit was significantly more pronounced for challenge compared to the control images. Consistent with this, ventral PFC inactivation also reduced the quality of the late-phase IT population code. Taken together, these results imply that PFC is a critical part (possibly one of many) of the recurrent circuitry underlying the production of explicit object representations in IT. Advance 3: In addition to pharmacological perturbation strategies, we have began testing an array of chemogenetic neuronal silencing strategies (via viral delivery of DREADDs) that will allow suppression of targeted recurrent circuits in the primate brain. This technique will allow us to probe the functional role of the recurrent circuits within the ventral stream (e.g. IT to V4, V4 to V2 etc). Although at its preliminary stages, the preliminary results demonstrate the effects of DREADDs to reversibly perturb 20mm2 of macaque V4 cortex to produce reduced neuronal firing rates and significant behavioral deficits during core object recognition.
In sum, our approach provides key architectural as well as image-level behavioral and neural constraints that will guide the next-generation recurrent models of the visual system.
Details
43 Vassar St, MIT Bldg 46-5165, Cambridge MA 02139