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

【日時】
2023年5月29日（月）午後3時00分～4時30分
【発表者】
玉井敬一／東京大学
【タイトル】
“Non-equilibrium phase transitions in artificial deep neural networks”
【概要】
Despite the rapid progress in applying the so-called deep learning in various fields, the behaviors of artificial deep neural networks are not yet fully understood from a theoretical perspective, even for the classical ones (such as the fully-connected feedforward multilayer perceptrons and the convolutional neural networks). Improvement of the theoretical understanding is desirable for a more principled design of efficient networks. In particular, given the energy efficiency of human brains in learning and generalizing, it is natural to ask what is different between artificial and biological neural networks.
In this talk, I will demonstrate that the theoretical framework of non-equilibrium phase transitions in statistical mechanics is a promising tool for describing the behavior of classical artificial deep neural networks (as well as biological ones). After briefly recalling the recent progress in understanding biological/artificial neural networks, I will combine theoretical and numerical approaches to uncover the universal scaling properties in the initialized artificial networks. The universality of the critical phenomena in non-equilibrium phase transitions allows for an intuitive understanding of the behaviors of the artificial networks across the different architectures.
【会場】
理学部１号館５階５１２号室