Elucidation of the Mathematical Basis and Neural Mechanisms of Multi-layer Representation Learning

2021 Fiscal Year Final Research Report

• Project Area: Correspondence and Fusion of Artificial Intelligence and Brain Science
• Project/Area Number: 16H06563
• 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: Okinawa Institute of Science and Technology Graduate University
• Principal Investigator: Doya Kenji 沖縄科学技術大学院大学, 神経計算ユニット, 教授 (80188846)
• Project Period (FY): 2016-06-30 – 2021-03-31
• Keywords: 強化学習 / 深層学習 / データ効率 / モデルベース / 大脳皮質 / 大脳基底核 / セロトニン

Outline of Final Research Achievements

In accordance with the aims of the project, we developed brain inspired algorithms for artificial intelligence　and utilized the theories and methods of artificial intelligence for advancing brain science.
We developed data-efficient model-based reinforcement algorithms and analyzed the data-efficiency of deep reinforcement learning algorithms.
We clarified representations of different variables for sensory inference and reinforcement learning in the cerebral cortex and the basal ganglia. We also revealed that the enhancement of the patience for delayed rewards by serotonin neuron stimulation is dependent on the certainty of reward delivery and the uncertainty of delivery timing, and proposed a novel Bayesian decision model to reproduce animal behaviors.

Free Research Field

計算神経科学

Academic Significance and Societal Importance of the Research Achievements

脳にならった新たな人工知能の開発と、人工知能の理論や手法による脳科学の進展という領域のねらいの具現化をめざし、限られた経験から効率よく行動学習を行う強化学習アルゴリズムの開発と、その性能を予測する理論の構築を行った。また、知覚推論と強化学習に関わる情報の大脳皮質と大脳基底核における表現を明らかにするとともに、セロトニンが将来報酬の予測に基づく行動を制御するしくみの解明を進めた。