AIP AI4S (AI for Science) Seminar Series
The AIP AI4S (AI for Science) Seminar Series is organized by the RIKEN Center for Advanced Intelligence Project (AIP).
This seminar series features invited researchers presenting recent advances, emerging methodologies, and interdisciplinary applications in AI for Science.
Through this series, we aim to promote cross-disciplinary discussion, foster collaboration, and strengthen the AI for Science research community.
Details of each seminar will be announced individually.
Date & Time : 4:00pm-5:00pm (JST), May 13, 2026
Venue : Hybrid (In-person + Online)
Please note that in-person attendance is limited to AIP researchers.
Speaker : Ayaka Sakata
Affiliation : Associate Professor, Ochanomizu University / Team Director, RIKEN AIP
Title : Breaking of Replica Symmetry on the Nishimori Line and its Implications for Bayesian Inference
Abstract :
In the statistical physics of spin glasses, the Nishimori line represents a subspace where the energy is an analytic function and certain thermodynamic identities hold. It is widely understood that replica symmetry (RS) is generally stable on this line. This property is one of the fundamentals that connect Bayesian inference and statistical physics.
However, a recently proposed model, termed "Nishimori Model," has revealed phase transitions associated with the breaking of replica symmetry even on the Nishimori line. In this talk, we explain the thermodynamic behavior of the Nishimori line using mean-field theory and provide a comprehensive picture of the associated phase transitions. By going beyond the conventional paradigm of RS stability on the Nishimori line, we explore the broader implications for Bayesian inference and discuss what these new insights mean for the fundamental limits of statistical estimation.
