Towards an Algebra for Distributed Deep Neural Networks

2023 Fiscal Year Final Research Report

• Project/Area Number: 19K22865
• Research Category: Grant-in-Aid for Challenging Research (Exploratory)
• Allocation Type: Multi-year Fund
• Review Section: Medium-sized Section 61:Human informatics and related fields
• Research Institution: Tokyo Institute of Technology
• Principal Investigator: Inoue Nakamasa 東京工業大学, 情報理工学院, 准教授 (10733397)
• Project Period (FY): 2019-06-28 – 2024-03-31
• Keywords: 深層学習

Outline of Final Research Achievements

Deep learning technology for understanding images and audio has become essential in an advanced information society. In this project, results were obtained regarding a mechanism in which multiple neural networks learn cooperatively or adversarially. The main achievements include a new regularization method for generative adversarial networks called Augmented Cyclic Consistency Regularization, an adversarial sample generation method using a second-order Quasi Newton method, and a step size regularization method. The effectiveness of these methods was demonstrated using real image and audio data, and the results were presented at international conferences.

Free Research Field

パターン認識

Academic Significance and Societal Importance of the Research Achievements

本研究の成果は実社会で活用されている画像認識・画像変換、音声認識・音声話者照合システムの高度化に貢献するものである。また、学術的には新たな学習アルゴリズムが情報工学分野、特にパターン認識およびニューラルネットワークの深層学習に貢献するものである。