知の物理学研究センター / Institute for Physics of Intelligence (iπ)

【日時/Date】
１１月２８日（火）１０：００－１１：３０

【場所/Venue】
Zoom

【発表者/Speaker】
平谷　直輝 先生（Washington University in St Loui）

【タイトル/Title】
"Principles of Online Continual Learning in Brains and Machines"

【概要/Abstract】
Modern AI systems now surpass human performance in many cognitive tasks, yet they remain notoriously brittle when it comes to learning continuously in an online setting. By contrast, the human brain integrates new knowledge flexibly and efficiently, without erasing what was learned before. In this talk, I will explore why online continual learning poses such challenges for artificial neural networks and how to overcome them. I begin by showing that online kernel learning can be understood as equivalent to offline kernel regression with biased labels, and discuss how this perspective informs curriculum learning strategies. I then turn to the problem of continual learning across multiple tasks, examining how both task similarity and task order shape learning outcomes. From this analysis, I derive two principles for optimizing task order and demonstrate their applicability. Finally, I will present preliminary results comparing artificial neural networks with human participants in continual learning tasks, focusing on similarities and divergences between biological and artificial learners.

これらの講演に関するZoomのリンク等の案内を受け取ることを希望されるかたは、
下記のgoogle formからメールアドレスをご記入ください。こちらに登録頂いた情報は、案内の配信のみに利用いたします。

登録フォーム：https://forms.gle/xnLmd9Kc1BaaNPgq8

過去の発表リスト：https://www.phys.s.u-tokyo.ac.jp/about/17106/

　　　　　　　　　　世話人：知の物理学研究センター　髙橋昂

⇦Top page