Group theoretic approach to deep learning on graphs【Cancel】

Reference No.	20200009
Type/Category	Grant for General Research-Short-term Joint Research
Title of Research Project	Group theoretic approach to deep learning on graphs【Cancel】
Principal Investigator	Akiyoshi Sannai（RIKEN・research scientist）
Research Period	January 31,2022. - February 4,2022.
Keyword(s) of Research Fields	Graph theory, Deep learning, invariant theory
Abstract for Research Report	Graph data appears in various aspects of society, such as the molecular structure of compounds, social networks, and transmission routes of infectious diseases. Analyzing those structures has become an important issue for humankind. On the other hand, learning by deep neural networks, such as image recognition, has achieved SOTA in many tasks. Graphs are no exception, and several graph deep neural nets that handle graphs have been defined, and have shown good results experimentally. However, there are also problems with these models, and for example, the limitations of expressiveness have been pointed out. The purpose of this joint research is to construct better graph neural nets (which can approximate any function under ideal conditions) by focusing on group actions naturally derived from graphs and considering the invariant rings by them. The construction of an excellent graph neural network not only contributes to drug discovery, compound property prediction, SNS analysis, etc. through actual learning, but can also be applied to academically important graph isomorphism problems This is a significant issue. In addition, the construction using invariant rings is mathematically advanced, and is considered to be meaningful for both machine learning and mathematics from the viewpoint of application of mathematics to machine learning. Deep sets (NeurIps 2017) constructed a deep neural network that is good for approximating functions on sets, with the idea of ??looking at generators of invariant rings. This can be viewed as a neural network that takes as input a graph without edges, or as a generalization of their composition. Another advantage is that it is clear which functions are easy to approximate compared to the existing methods.
Organizing Committee Members (Workshop) Participants (Short-term Joint Usage)	Kenta Oono（Preferred Networks・Engineer） Yuuki Takai（RIKEN AIP・Postdoctoral Researcher） Takanori Maehara（RIKEN AIP・Unit reader） Akiyoshi Sannai（RIKEN AIP・Researcher）