Sample selection method for large scale datasets to improve robustness in recognition tasks

2022 Fiscal Year Final Research Report

• Project/Area Number: 19K12034
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 61010:Perceptual information processing-related
• Research Institution: National Institute of Advanced Industrial Science and Technology
• Principal Investigator: Kenji Watanabe 国立研究開発法人産業技術総合研究所, 情報・人間工学領域, 研究員 (50571064)
• Project Period (FY): 2019-04-01 – 2023-03-31
• Keywords: 多変量解析 / 因子分解

Outline of Final Research Achievements

Machine learning methods have been applied to solve recognition tasks in many academic and commercial fields, and the methods are demanded for the improvement of robustness to solve the tasks. Overcoming this problem, training datasets should be re-constructed only using favorable samples which are subtracted to outliers.
In this research, we studied a matrix factorization which is applied in a sample selection framework for large scale and unknown dataset. Because we may be able to subtract the outliers from the datasets by measuring distances and/or simple criteria in the feature space for the input (original) samples and obtained samples from the factorization.

Free Research Field

パターン認識　多変量解析

Academic Significance and Societal Importance of the Research Achievements

本研究で着目した因子分解手法は古典的な多変量解析手法の一つであり、昨今の隆盛を極める深層学習手法を検討対象とすることをあえて避けたのは、一定の理論的基準と確信を持って、汎化性能の向上に臨めるからである。これは、現在の学術・商用を問わず一定の性能が望めるという一点のみで、「なぜ、所望の性能を達成できたのか？」という理論的解析が困難な深層学習手法を軽々と利用する風潮に一石を投じる意味で学術的・社会的意義のある研究であるものと考える。