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

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Contributions of Visual Pathway Segregation to Learning Multiple Behaviors with Divergent Visual Demands

Aida Mirebrahimi¹, David C. Plaut¹, Marlene Behrmann²; ¹Carnegie Mellon University, ²University of Pittsburgh

Presenter: Aida Mirebrahimi

The primate visual system can support both object recognition and geometry-based grasping from the same input, while they impose different learning demands on the system. E.g. recognition benefits from invariance to changes in viewpoint and scale, whereas grasping requires sensitivity to those same features. In the brain, these behaviors are supported by two partially-segregated dorsal and ventral pathways. Although these pathways have been shown to differ in their representational properties, it remains unclear whether this anatomical segregation plays a role in resolving the multi-task learning tension between these behaviors. Here we test this using computational models, comparing a fully shared single-pathway (SP) architecture with an early-shared, late-segregated dual-pathway (DP) one, alongside single-task baselines. DP matched SP on recognition while learning grasping faster and generalizing better as distribution shifts grew. The analysis of task gradients showed that segregation reduced recognition-dominated interference in shared representations while preserving early feature sharing. Partial pathway segregation thus acts as an inductive bias for efficient simultaneous learning under competing visual demands.

Topic Area: Computational Models of Vision & Visual Cortex