Theory and methods for nonlinear feature extraction

2020 Fiscal Year Final Research Report

• Project/Area Number: 18K18107
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 61030:Intelligent informatics-related
• Research Institution: Future University-Hakodate (2019-2020); Nara Institute of Science and Technology (2018)
• Principal Investigator: Sasaki Hiroaki 公立はこだて未来大学, システム情報科学部, 准教授 (80756916)
• Project Period (FY): 2018-04-01 – 2021-03-31
• Keywords: 非線形特徴抽出 / 非線形独立成分分析 / 外れ値 / 相互情報量最大化 / ニューラルネットワーク / 統計的因果推論

Outline of Final Research Achievements

This research is aimed at developing a theory and methods for nonlinear feature extraction. A rigorous theory for nonlinear independent component analysis (ICA) was established. Furthermore, a unified framework has been proposed for unsupervised nonlinear feature extraction, which includes nonlinear ICA, maximization of mutual information and nonlinear subspace estimation as special cases. Practical methods were also proposed. Especially, a robust method against outliers was developed and investigated through both theoretical analysis and numerical experiments.

Free Research Field

機械学習

Academic Significance and Societal Importance of the Research Achievements

これまでの教師なし非線形特徴抽出の実践手法は経験・発見的なアプローチに基づくことが多く，その理論的な背景を理解することが難しい状況にあった．それに対して，本研究で提案した統一的な枠組みは既存の非線形特徴抽出の理論的な基盤となる可能性があり，学術的な意義は大きい．また，提案した枠組みは密度比推定に基づくため，既存の密度比推定法を応用・拡張することによって，さらなる実践手法の発展が見込めるため，今後の成果が期待される研究内容である．