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

Poster B in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Emergent Demand-Driven Precision in Neural Networks Optimized for Multiple Object Tracking

Sun Minni¹, Eivinas Butkus¹, Nikolaus Kriegeskorte¹, Benjamin Peters²; ¹Columbia University, ²University of Edinburgh

Presenter: Sun Minni

Humans can reliably track multiple objects moving among visually identical distractors in a task known as multiple object tracking (MOT). This ability suggests that the brain continually updates location estimates of tracked objects. However, success in MOT, as in real-world tracking, requires only identifying the targets at the end rather than maintaining precise location estimates throughout. Previous work has found that tracking precision in the MOT task is dynamically modulated by task demands, such as the distance between a target and its nearest distractor. To better understand the underlying computations, we trained neural networks with two distinct objectives: a classification objective, which mirrors the actual requirement of MOT to identify the targets correctly, and a regression objective, which drives the model to precisely predict the target locations. We found that classification models generalized more effectively to larger target sets than regression models. Furthermore, only classification models reproduced the demand-driven tracking precision observed in human behavior. These results suggest that demand-driven precision emerges when neural networks learn to distinguish targets from distractors rather than track their precise locations.

Topic Area: Computational Models of Vision & Visual Cortex