Statistical Physics of Visual Information

2022 Fiscal Year Final Research Report

• Project/Area Number: 19K03657
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 13010:Mathematical physics and fundamental theory of condensed matter physics-related
• Research Institution: Shibaura Institute of Technology
• Principal Investigator: Tomita Yusuke 芝浦工業大学, 工学部, 教授 (50361663)
• Project Period (FY): 2019-04-01 – 2023-03-31
• Keywords: 統計物理学 / 視覚情報 / 特徴抽出

Outline of Final Research Achievements

The differences between spin configurations and spin correlations in machine learning are investigated using spin models. Recent progress in machine learning stimulates applications to physics and analyzing mechanisms of machine learning from the viewpoint of statistical physics. We paid attention to the different effects on learning efficiency between apparent spin configurations and graph representations which contain information of correlations. In the case of the application to the classical and the quantum XY model, we have shown that the neural network trained by the classical XY model can discriminate phases in the quantum XY model. Through the application to the inverse renormalization group, we have confirmed that feeding information of correlations between spins by the graph representation has improved the precision of the inverse renormalization.

Free Research Field

統計物理学

Academic Significance and Societal Importance of the Research Achievements

スピン模型のグラフ表現は元のスピン変数を用いた記述に比べ, 様々な場面で有用であることが知られている. 本研究では機械学習の学習データとしてグラフ表現がスピン変数より常に同等もしくは優位であることが示され, 視覚情報(スピン変数)が陰に持っているスピン相関(グラフ表現)が重要となることと, 学習データから真に重要な情報がいつでも取得されるわけではないことが明らかになった. 視覚情報のみの場合とスピン相関を含めた場合とで学習に有意な差が見られたことは今後の研究にも生かされる重要な知見が得られたと考えている. 本研究で得られた結果は今後の人工知能技術など社会への波及効果も期待される.