JSAI2025

Presentation information

Organized Session

Organized Session » OS-32

[3L6-OS-32] OS-32

Thu. May 29, 2025 5:40 PM - 7:20 PM Room L (Room 1007)

オーガナイザ:高槻 瞭大(AIアライメントネットワーク/東京大学),峰岸 剛基(東京大学),宮西 洋輔(サイバーエージェント/北陸先端科学技術大学院大学),高木 優(国立情報学研究所)

6:40 PM - 7:00 PM

[3L6-OS-32-04] Towards Reverse-engineering the Kanizsa Illusion: A Mechanistic Study of Vision Transformer

〇Ryota Takatsuki1,2,3, Sonia Joseph4,5, Ippei Fujisawa2,3, Ryota Kanai2,3 (1. The University of Tokyo, 2. Araya Inc., 3. AI Alignment Network, 4. Mila - Quebec AI Institute, 5. McGill University)

Keywords:Mechanistic Interpretability, Optical Illusion, Vision Transformer

The formation of illusory contours has been associated with predictive processing, yet its detailed mechanism remain unclear. In this study, we show that the Kanizsa illusion can also be observed in Vision Transformers, a class of feedforward neural networks, thereby challenging the conventional understanding of their formation. To elucidate the underlying mechanism, we introduce a novel mechanistic interpretability method leveraging a diffusion model to track how predictions evolve across transformer layers. Finally, We discuss the universality of mechanisms between models and biological systems and the potential of our approach to contribute to a deeper understanding of the illusory contour formation in biological systems.

Please log in with your participant account.
» Participant Log In