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

Perception as Phase-Based Dynamics over Competing Interpretations

Sungryong Koh¹, Dongjun Kim¹, Seoyoung Ahn²; ¹Seoul National University, ²Hankuk University of Foreign Studies

Presenter: Seoyoung Ahn

Conventional deep learning models do not explicitly represent competing interpretations and their interactions, typically collapsing representations into a single dominant hypothesis early in processing. We argue that structured dynamic interactions provide a principled alternative. This study proposes a phase-based neural architecture in which multiple candidate interpretations are maintained at each spatial location and resolved through Kuramoto-style synchronization. The model consists of two parallel pathways (form and luminance) organized into stacked Kuramoto layers with residual connections, and integrates cross-pathway information via Possible World Attention (PWA). Evaluated on grayscale CIFAR-10, the model achieves 80.41% accuracy with fixed random filters, demonstrating that dynamic interactions alone, without extensive feature learning, can drive strong performance. Ablation studies confirm that PWA is essential. Replacing it with simple concatenation drops performance to approximately 50%, regardless of channel count. This study suggests that a dynamics-based perspective enables both competitive performance and more straightforward interpretation of the model's internal processes.

Topic Area: Computational Models of Vision & Visual Cortex