A Study on Controllable Representation Learning using Adversarial Training

2022 Fiscal Year Final Research Report

• Project/Area Number: 18K18101
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 61030:Intelligent informatics-related
• Research Institution: The University of Tokyo
• Principal Investigator: Iwasawa Yusuke 東京大学, 大学院工学系研究科(工学部), 講師 (70808336)
• Project Period (FY): 2018-04-01 – 2023-03-31
• Keywords: Deep Learning / Firaness / Privacy / Transfer / Adversarial Training

Outline of Final Research Achievements

Throughout the research period, I achieved the following technical accomplishments:
(1) Analyzed the instability of the existing method, Adversarial Feature Learning, and proposed a solution (accepted in IJCAI2020 and other conferences). (2) Proposed a new criterion for invariance, called Sufficient Invariance, which maximizes invariance with respect to a factor of interest in an informationally novel range, and suggested methods to achieve Sufficient Invariance (accepted in ECML2019 and other conferences). (3) Proposed a framework based on graphical models to remove information from data without providing detailed information about the specific aspects the user wants to eliminate, and presented the corresponding methodology (accepted in ECML2021 and other conferences).

Free Research Field

Deep Learning

Academic Significance and Societal Importance of the Research Achievements

大規模言語モデルの登場などにより深層学習の実世界での活用は本格化しているが、通常の学習アルゴリズムは内部にあるバイアスを増長してしまう可能性がある。また、意図しない状況で不安定な挙動をすることがある。本研究の目的は、深層NNの表現が特定の情報を持たないように制御する要素技術の開発である。本研究成果により、未知ユーザの行動を高精度に認識したり、深層NNの判断基準が特定の因子によらないことを保証（プライバシー保護、公平性配慮）できると考えられる。