Center for Advanced Intelligence Project

Date & Time: 4:15pm-5:15pm (JST), March 18, 2026
Venue: Hybrid (In-person + Online)
Speaker: Masaaki Imaizumi
Affiliation: Associate Professor, University of Tokyo / Team Director, High-Dimensional Structure Theory Team, RIKEN AIP
Title: Physics for AI: Dynamics of neural network learning and transformer inference
Abstract:
This talk introduces several analysis on dynamics of AI architecture: the learning dynamics of neural networks and the inference dynamics of transformers. These analysis leverages the recent development of physics-oriented theory for neural networks. While neural networks exhibit complex dynamics in many aspects, employing the high-dimensional limit to reduce it to the dynamics of element distributions enables effective analysis. The first topic describes several approaches for analyzing the training dynamics of deep neural networks, followed by an estimation of generalization error estimation and multi time-scales for feature unlearning. The second topic explains the research background of representing transformer inference using nonlinear dynamical models with coupled oscillators, and analyzes the mechanism by which this model induces degeneracy. By extending these studies, we discuss the prospect of advancing the fundamental understanding of deep learning and artificial intelligence by applying knowledge from physics.