Study of dynamic attentional selection mechanisms in deep neural networks

2018 Fiscal Year Final Research Report

• Project/Area Number: 16K16090
• Research Category: Grant-in-Aid for Young Scientists (B)
• Allocation Type: Multi-year Fund
• Research Field: Perceptual information processing
• Research Institution: Tokyo Denki University
• Principal Investigator: HIDAKA Akinori 東京電機大学, 理工学部, 准教授 (70553519)
• Research Collaborator: KURITA Takio
• Project Period (FY): 2016-04-01 – 2019-03-31
• Keywords: 深層学習 / 選択的注意 / 畳み込みニューラルネットワーク

Outline of Final Research Achievements

For analyzing the mechanism of consecutive visual information processing which is observed in both of visual cortex of humans and deep convolutional neural networks (CNN), we proposed interlayer correlations analysis based on canonical correlation analysis. It can realize quantitative evaluation and visualization of similarity or difference between an arbitrary pair of internal layers in CNN.
By using CNN trained for predicting the location of attentional selection and human gaze, we conducted the experiments for quantitative comparison between the neural system in visual cortex and CNN. We compared internal responses of CNN which is trained for saliency map prediction and the responses of V1, V4, IT area when they see (or process) several natural images. As a result, we obtained several novel findings about similarity and difference between CNN and the neural system in visual cortex.

Free Research Field

深層学習

Academic Significance and Societal Importance of the Research Achievements

畳み込みニューラルネットワーク（CNN）の極めて強力な画像認識性能は，現在の人工知能ブームを巻き起こす直接的なきっかけであり，今なお中心的な原動力であり続けている．本研究では，CNNの従来研究では長らく考慮されてこなかった大脳視覚野の認知処理過程における「注意（Attention）」の機構に焦点を当て，CNNの内部反応と大脳視覚皮質との類似性や相違性を定量的に分析し，新たな知見を得た．これらは，強力な認識性能を持つが静的な振る舞いしかできないCNNに対し，ヒトの視覚認知機構における「注意」に基づく動的なフィードバック機構等を組み込んで柔軟な認知能力を実現しようとする際に，重要な知見となる．