Training of deep learning models by introducing prior knowledge

2023 Fiscal Year Final Research Report

• Project/Area Number: 21K12049
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 61040:Soft computing-related
• Research Institution: Hiroshima University
• Principal Investigator: Kurita Takio 広島大学, 先進理工系科学研究科(工), 教授 (10356941)
• Co-Investigator(Kenkyū-buntansha): 日高 章理 東京電機大学, 理工学部, 准教授 (70553519)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: 深層学習 / 事前知識 / 不変特徴抽出 / パターン認識

Outline of Final Research Achievements

We studied methods for actively incorporating task constraints into the learning results of deep learning. Specifically, we proposed methods for incorporating prior knowledge as a regularization term, removing information irrelevant to the task from the training results, and augmenting the training data using prior knowledge. We applied the proposed approaches to image identification, image region extraction, and object detection, and experimentally confirmed their effectiveness.

Free Research Field

情報科学

Academic Significance and Societal Importance of the Research Achievements

深層学習は訓練データから自動的にモデルのパラメータを推定してくれるため非常に便利であるが得られたモデルにタスクが持つ制約条件が十分に取り入れられていない．本研究では深層学習の学習結果にタスクの制約条件を積極的に取り込むための方法について研究した．これは深層学習の結果を信頼して使うためのひとつのアプローチであると考える．