Learning Algorithm for Advanced Generative Models

2020 Fiscal Year Final Research Report

• Project/Area Number: 17H04693
• Research Category: Grant-in-Aid for Young Scientists (A)
• Allocation Type: Single-year Grants
• Research Field: Intelligent informatics
• Research Institution: The University of Tokyo
• Principal Investigator: Sato Issei 東京大学, 大学院情報理工学系研究科, 准教授 (90610155)
• Project Period (FY): 2017-04-01 – 2021-03-31
• Keywords: 機械学習 / 生成モデル / 学習アルゴリズム

Outline of Final Research Achievements

In recent years, the development of generative models in the field of machine learning has been remarkable, reaching a level where images and videos of people who do not exist in reality are indistinguishable to the human eye. However, learning such generative models differs from learning ordinary discriminative models in that it often requires many heuristics in the learning method due to the instability of learning caused by the difficulty of optimizing the objective function and the non-triviality of formulation caused by the difficulty of evaluating the products. In this research, we have worked on elucidating and mitigating the learning instability, developing a generative model for tabular data, which has not been dealt with in existing research, and applying it to minority data analysis.

Free Research Field

機械学習

Academic Significance and Societal Importance of the Research Achievements

生成モデルは通常の機械学習における識別モデルによる予測とはことなり，現実には存在しないデータを生成する．現在では，画像や動画などの生成において人が認識できないレベルまで到達しており，人が数時間数日かけて作成するようなものでも数秒で生成してしまう．これは人の創作活動において大きな変革をもたらす可能性を秘めている．しかし，その学習アルゴリズムは未だに解明できていないことも多く，特に学習の不安定さは問題で，学習アルゴリズムに対して多くのヒューリスティックスと試行錯誤が必要となる．本研究では，学習の安定性に関する基礎研究を行い，また生成モデルの適用範囲を表形式データに拡張し，少数データ解析へ応用した．