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

Unifying Gestalt Principles of Perceptual Organization Through Feedback-Mediated Learned Prior Integration

Tahereh Toosi¹, Kenneth D. Miller¹; ¹Columbia University

Presenter: Tahereh Toosi

Gestalt principles of perceptual organization (such as closure, similarity, continuation, and figure-ground segregation) have long been characterized as distinct rules guiding visual perception. We demonstrate that these diverse phenomena emerge from a single mechanism: inference-time prior integration through cortical feedback pathways. We introduce Generative Inference, implemented via Prior-Guided Drift Diffusion (PGDD), which repurposes feedback connections used during learning to refine neural activations at inference time. Applied to a pre-trained robust network without additional training, PGDD produces induced activity (feedback-driven responses at stimulus-absent receptive field locations), capturing the neural and perceptual signatures of all four Gestalt phenomena tested in neurophysiological experiments. These effects depend on dynamic integration of learned priors and are absent in large-scale recurrent and predictive coding models despite their feedback connections, demonstrating that what matters is not recurrence per se, but whether feedback carries learned prior information. These results suggest Gestalt principles are emergent consequences of feedback-mediated prior integration.

Topic Area: Computational Models of Vision & Visual Cortex