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

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Predictive learning explains emergence of objects in infants

Alekh Karkada Ashok¹, Drew Linsley¹, Thomas Serre¹; ¹Brown University

Presenter: Alekh Karkada Ashok

How do young animals learn to parse visual scenes into objects with so little experience? Human infants initially rely on motion cues to segregate objects, while sensitivity to static shape cues emerges only later --- mirroring the earlier maturation of the dorsal visual stream relative to the ventral stream. Here we introduce a two-stream neural network in which a shape pathway and a motion pathway cooperate to predict future visual input. Trained on videos of moving objects, the model first discovers objects through coherent motion, then develops sensitivity to static shape cues, recapitulating the developmental trajectory observed in infants. These results demonstrate that cross-stream predictive learning is sufficient to produce object representations from a realistic amount of visual experience, without requiring innate perceptual priors.

Topic Area: Computational Models of Vision & Visual Cortex