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

【日時/Date】
2022年11月24日（木）10時30分～12時00分

【講演者/Speaker】
原田 健自 Kenji Harada /京都大学 Kyoto University

【講演タイトル/Title】
"Compressing neural networks by tensor networks"

【概要/Abstract】
Neural networks in artificial intelligence are computational models to simulate living neural networks that perform various learning tasks. They are highly important in recent breakthroughs in artificial intelligence and have been the subject of various studies. In general, recent neural networks tend to be large, with many parameters, and their compression is desired. In this talk, I will introduce studies in which tensor networks are applied to compress neural networks[1,2]. Tensor networks can compress high degrees of freedom with strong correlations, such as quantum states. I will also report on recent research on various aspects of tensor networks in neural networks.
Reference:
[1] A. Novikov, D. Podoprikhin, A. Osokin and D. P. Vetrov, ``Tensorizing Neural Networks’‘, NIPS (2015).
[2] Z.-F. Gao, S. Cheng, R.-Q. He, Z. Y. Xie, H.-H. Zhao, Z.-Y. Lu and T. Xiang, ``Compressing deep neural networks by matrix product operators’’, Phys. Rev. Res. 2, 023300 (2020).

資料PDF