Presentation 2022-03-08
[Invited Talk] ---
Takashi Matsubara,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Deep learning is being considered as the most promising approach to building an artificial intelligence (AI) system; it sometimes recognizes and edits images and natural languages at a superhuman level. Given a sufficient amount of data and computational resources, deep learning can approximate an arbitrary function. However, deep learning makes decisions difficult to understand and control, and it is often described as ``unreliable''. The phrase ``AI is unreliable'' implies that ``non-AI approaches are reliable.'' In contrast to regular deep learning, mathematical models are designed to guarantee properties of targets, such as a dependency between factors, a geometric symmetry, and laws of physics. If deep learning guaranteed these properties, it would provide the same level of reliability as mathematical models. When replacing operations that compose deep learning appropriately, the function space to be approximated is restricted to a certain subset with desired properties, and the deep learning after training is guaranteed to have those properties. In fact, convolutional and graph neural networks have the translation and permutation equivariance, respectively. Geometric deep learning is a generalization these approaches, which guarantees various properties described using geometric concepts. Conservation laws of physical systems are associated with certain geometric symmetries and are included as objects of geometric deep learning. In this talk, the author will introduce geometric deep learning that guarantees various properties of dynamical systems, with a focus on recent publications by the author or his collaborators.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) geometric deep learning / geometric symmetry / dynamical system / conservation law
Paper # IBISML2021-34
Date of Issue 2022-03-01 (IBISML)

Conference Information
Committee IBISML
Conference Date 2022/3/8(2days)
Place (in Japanese) (See Japanese page)
Place (in English) Online
Topics (in Japanese) (See Japanese page)
Topics (in English) Machine Learning, etc.
Chair Ichiro Takeuchi(Nagoya Inst. of Tech.)
Vice Chair Masashi Sugiyama(Univ. of Tokyo)
Secretary Masashi Sugiyama(Univ. of Tokyo)
Assistant Tomoharu Iwata(NTT) / Atsuyoshi Nakamura(Hokkaido Univ.)

Paper Information
Registration To Technical Committee on Infomation-Based Induction Sciences and Machine Learning
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) [Invited Talk] ---
Sub Title (in English)
Keyword(1) geometric deep learning
Keyword(2) geometric symmetry
Keyword(3) dynamical system
Keyword(4) conservation law
1st Author's Name Takashi Matsubara
1st Author's Affiliation Osaka University(Osaka Univ.)
Date 2022-03-08
Paper # IBISML2021-34
Volume (vol) vol.121
Number (no) IBISML-419
Page pp.pp.27-27(IBISML),
#Pages 1
Date of Issue 2022-03-01 (IBISML)