Knowledge-embedded Bayesian deep learning and its application to small data analysis

2023 Fiscal Year Final Research Report

• Project/Area Number: 21H03511
• Research Category: Grant-in-Aid for Scientific Research (B)
• Allocation Type: Single-year Grants
• Section: 一般
• Review Section: Basic Section 61040:Soft computing-related
• Research Institution: Osaka University (2022-2023); Kyushu University (2021)
• Principal Investigator: Hayashi Hideaki 大阪大学, データビリティフロンティア機構, 准教授 (00790015)
• Co-Investigator(Kenkyū-buntansha): 古居 彬 広島大学, 先進理工系科学研究科(工), 助教 (30868237)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: ニューラルネットワーク / 深層学習 / ベイズ推定 / 希少データ

Outline of Final Research Achievements

Applying existing deep learning algorithms is often challenging for small data, such as medical data, where data collection is costly and annotation requires specialized knowledge. In this study, we proposed a framework for appropriately learning small data, called knowledge-embedded Bayesian deep learning. During the study, we developed the foundational techniques and applied them to real-world data analysis. Specifically, we proposed methods to reduce the amount of labeled data by estimating the distribution of input data while training the classifier, and methods to efficiently select training data through collaboration between the classifier and humans. Additionally, we performed real-world data analysis on various datasets such as infant motion images and electromyograms.

Free Research Field

パターン認識

Academic Significance and Societal Importance of the Research Achievements

社会的意義としては，機械学習アルゴリズムを実データ解析に応用する際，データ収集やラベル付けに必要なコストを削減する技術を提案した点である．特に，医用データ解析では大規模データの収集そのものが難しく，ラベル付けにも医師の専門知識が必要となるため，有効な手段となる．学術的意義としては，単一の深層学習モデルで識別モデルと生成モデルを同時学習できる半教師あり学習や信頼度較正に有効な手法の提案や，相対ラベルを用いて学習できるベイズ深層学習モデルの提案とその妥当性の理論的証明が主な貢献となる．