2020 Fiscal Year Final Research Report

Study on the improvement of the forecast due to the fusion of deep learning and symbol processing

Planned Research

PDF

Project Area	Correspondence and Fusion of Artificial Intelligence and Brain Science
Project/Area Number	16H06562
Research Category	Grant-in-Aid for Scientific Research on Innovative Areas (Research in a proposed research area)
Allocation Type	Single-year Grants
Review Section	Complex systems
Research Institution	The University of Tokyo
Principal Investigator	YUTAKA MATSUO 東京大学, 大学院工学系研究科(工学部), 教授 (30358014)
Co-Investigator(Kenkyū-buntansha)	PRENDINGER HELMU 国立情報学研究所, コンテンツ科学研究系, 教授 (40390596) 中山浩太郎東京大学, 大学院工学系研究科(工学部), 学術支援専門職員 (00512097)
Project Period (FY)	2016-06-30 – 2021-03-31
Keywords	深層学習 / 深層生成モデル
Outline of Final Research Achievements	In order to realize the integration of deep learning and symbolic processing, we constructed methods for deep reinforcement learning and studied a world model to acquire the environment and interactions. In the first half of our research, we struggled with the fast pace of the deep learning domain, as our ideas were often published in papers before we could, but in the second half of our research, based on the points raised in the mid-term review, we revised our research theme and were able to lead to many paper results at top international conferences such as ICLR and ICML. Specifically, multimodal deep generative models, or deployment efficient reinforcement learning methods to utilize world models. In the final year of the project, we proposed new models for self-supervised learning of the cerebral cortex, and made significant progress in the integration of brain science and artificial intelligence.
Free Research Field	人工知能
Academic Significance and Societal Importance of the Research Achievements	世界モデルの研究は、現在の深層学習を記号処理と融合する際に基盤となるものである。そのための手法を多面的に研究し、例えば、マルチモーダルな深層生成モデルでは、複数のモーダルが与えられたときに、一部のモーダルで欠損があったときにどのように復元するかという問題を扱った。深層強化学習の分野では、モデルに基づく手法とモデルフリーな手法があるが、この両者の良いとこ取りをするデプロイ効率な手法を提案した。世界モデルの技術は、意味理解を可能とする人工知能につながり、また人工知能と脳科学の融合の土台となる可能性が高く、学術的な意義は大きい。また、今後ロボット等への活用につながれば社会的な意義も大きい。