Feature Learning from Few-shot Videos Based on Group Representation Theory

2022 Fiscal Year Final Research Report

• Project/Area Number: 19K20290
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 61010:Perceptual information processing-related
• Research Institution: The University of Tokyo
• Principal Investigator: Mukuta Yusuke 東京大学, 先端科学技術研究センター, 講師 (50830874)
• Project Period (FY): 2019-04-01 – 2023-03-31
• Keywords: Invariance / Feature Learning / Machine Learning / Computer Vision

Outline of Final Research Achievements

We proposed an extension of the model structure and the learning method for Equivariant Neural Networks, which are the Neural Networks model that considers equivariance. As an extension of the model structure, we proposed a equivariant extension of the feature coding method, in which local features in an image are summarized to form a single global feature. As a new learning method, we proposed equivariant pretext labels and invariant contrastive loss, which are equivariant losses for self-supervised learning that matches the structural restriction of Equivariant Neural Networks.

Free Research Field

Machine Learning

Academic Significance and Societal Importance of the Research Achievements

提案手法はDeep Neural Networksに対して共変性の事前知識を陽に組み込むためのEquivariant Neural Networksを活用するための枠組みになっている。提案手法により通常のDeep Neural Networksより高性能でかつ性質がより明らかな認識が行えることが期待される。理論的にも不変な特徴量の有効性の理由の一端となる性質を証明し、不変な特徴量に対する理解の促進につながった。