Object Recognition by Deep Neural Network with Knowledge Graph Embedding - Proposal for Semantic Object Projection -

2019 Fiscal Year Final Research Report

• Project/Area Number: 17K00236
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Perceptual information processing
• Research Institution: Kobe University
• Principal Investigator: ARIKI Yasuo 神戸大学, 都市安全研究センター, 名誉教授 (10135519)
• Project Period (FY): 2017-04-01 – 2020-03-31
• Keywords: ゼロショット学習 / 意味的特徴 / 画像特徴 / 知識グラフ / 深層学習 / 意味ベクトル / 対話システム

Outline of Final Research Achievements

In this study, zero-shot deep learning was investigated, which can learn both of image features and the related knowledges simultaneously, and then improve the recognition accuracy of known objects as well as unknown objects. As the results, knowledge graph was proven to be superior to words and texts in terms of the description ability and the knowledge graph of the complete Wordnet improved the recognition rate by 45% compared with state of the arts. In addition, as the application of zero-shot deep learning, we studied about dialogue systems which can incorporate external knowledge.

Free Research Field

知能情報学

Academic Significance and Societal Importance of the Research Achievements

物体画像だけから特徴を抽出し認識するという方法には限界があり、「物体に関する知識も同時に学習させることにより、物体認識精度の向上、未知物体の識別と学習（ゼロショット学習）が可能になる」という考えに基づき、シンボル表現された知識と画像信号とを融合して学習を進める方法を明らかにした。また、この考えに基づき、発話文と知識を融合することで、知的な対話を行う方法を明らかにした。