Development of Machine Learning Framework Based on Structure Optimization of Computational Graph

2022 Fiscal Year Final Research Report

• Project/Area Number: 20H04240
• Research Category: Grant-in-Aid for Scientific Research (B)
• Allocation Type: Single-year Grants
• Section: 一般
• Review Section: Basic Section 61030:Intelligent informatics-related
• Research Institution: Yokohama National University
• Principal Investigator: Shirakawa Shinichi 横浜国立大学, 大学院環境情報研究院, 准教授 (90633272)
• Project Period (FY): 2020-04-01 – 2023-03-31
• Keywords: 機械学習 / 最適化 / 計算グラフ / 進化計算

Outline of Final Research Achievements

In this research, we developed machine learning methods for learning knowledge representation models from data by optimizing the structure of computational graphs. We showed that our methods could obtain interpretable and resource-efficient models by optimizing the computational graph consisting of interpretable operation units. In addition, to improve the scalability and efficiency of structure optimization, we developed and improved optimization algorithms based on the gradient methods using relaxation schemes called stochastic relaxation and continuous relaxation.

Free Research Field

知能情報学、人工知能

Academic Significance and Societal Importance of the Research Achievements

本研究で開発した計算グラフの構造最適化に基づく機械学習方式は、モデルの構造自体を効率的に学習できるという利点がある。これにより、解釈可能な演算ユニットからなる計算グラフの学習や、コンパクトな構造の学習が可能となるため、解釈性や計算効率の良いモデルが求められる応用で活用できる。さらに、本研究で開発した最適化方式は、計算グラフの構造最適化以外の問題にも応用できる可能性がある。