Presentation Information
[9a-N106-11]Modeling learning dynamics of generative models via magnetization ordering induced by spin dynamics
〇(M2)Gun Yoon1, Tomosato Hioki1,3,4, Koujiro Hoshi1,2,4, Naoto Yokoi1,2,4, Eiji Saitoh1,2,3,4 (1.Univ. Tokyo, 2.Inst. AI and Beyond, 3.AIMR, Tohoku Univ., 4.CEMS, RIKEN)
Keywords:
spintronics,Magnetism,Deep learning
This study reinterprets machine learning as an open thermodynamic system of magnetic matter, proposing a framework that links learning dynamics with physical principles. Focusing on the equivalence between regularization and magnetization order formation, we aim to unify perspectives from information theory and condensed matter physics. We construct a generative model—termed the spin autoencoder—comprising an encoder, a spin relaxation layer based on the stochastic Landau-Lifshitz-Gilbert equation, and a decoder. Learning is achieved by minimizing energy defined via reconstruction error and effective magnetic fields, with temperature controlling regularization strength.