[1G5-OS-21b] 世界モデルと知能

[1G5-OS-21b-01] Intrinsic Motivation for Learning Object-Centric Representations

〇Akihiro Nakano Nakano¹, Masahiro Suzuki¹, Yutaka Matsuo¹ (1. the University of Tokyo)

Keywords:Representation learning, World models, Curiosity, Object-centric learning

Curiosity is a powerful ability that humans possess which motivates them to interact with its surroundings to construct a compositional understanding of the world. Recent object-centric learning methods have enabled models to learn similar, compositional representations from videos. However, whether learning object-centric representations and exploring to interact with the environment in tandem by intrinsically motivating itself is possible remains unsolved. In this paper, we focus on investigating whether the model can use its own prediction to prioritize uncertain experiences and construct object-centric representations more efficiently, purely from images without supervision.