Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster E in Poster Session E: Thursday, August 6, 10:30 am – 12:15 pm, Kimmel Center, Shorin & Rosenthal Rooms

Towards Modeling Recurrence in Biological Vision with Recurrent Vision Transformers

Brian S Robinson¹, Michael Bonner¹; ¹Johns Hopkins University

Presenter: Brian S Robinson

Creating performant computational models of biological visual perception that capture properties of recurrence is a long-standing challenge and central to investigating observed experimental phenomena. Vision transformers offer a flexible framework in which a fixed embedding dimension naturally supports recurrence, enabling recurrent application of a single layer with fixed weights rather than the typical feedforward stack of distinct layers. However, it is not well understood how such recurrent models perform for computer vision tasks and their relative brain similarity. In this work, we find that recurrent vision transformers can match performance of non-recurrent counterparts, despite having an order of magnitude fewer parameters. Furthermore, we find that recurrent vision transformers have qualitative differences in their brain similarity as compared to non-recurrent versions, which suggests differences in the functional computation of these networks.

Topic Area: Computational Models of Vision & Visual Cortex