講演情報
[P1-11]Infusing Theories into Deep Learning for Human-Aligned Interpretable Boundary Anticipation
*江 恩傑1、日高 昇平1 (1. 北陸先端科学技術大学院大学)
キーワード：
事象分節化（Event Segmentation）、境界予測(Boundary Anticipation)、深層学習(Deep Learning)、グラフニューラルネットワーク(Graph Neural Networks)、ベルンシュタインの協応構造(Bernstein’s Synergies)
This study aims to construct a computational model that explains the general tendency of human event boundary anticipation (Zacks et al., 2007). Existing change-based (error-driven) hypotheses, primarily rooted in predictive processing (Friston, 2010), exhibit inherent predictive limitations; they identify boundaries only after physical transitions occur, failing to account for "pre-change" anticipation. Drawing on Bernstein's (Bernstein., 1967) theory of motor synergies, we suggest that boundary anticipation relies on the continuous evaluation of inner-event structural progression rather than discrete inter-event prediction errors. This perspective challenges the fundamental premise of purely error-driven approaches in previous computer vision studies. Moving forward, we will implement a dual-mechanism framework using Graph Neural Networks (GNNs) to reformulate boundary detection.