Functional roles of brain anatomical structures in the fault tolerance

• Project/Area Number: 18K11527
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 62010:Life, health and medical informatics-related
• Research Institution: Yamaguchi University
• Principal Investigator: Samura Toshikazu 山口大学, 大学院創成科学研究科, 准教授 (30566617)
• Project Period (FY): 2018-04-01 – 2021-03-31
• Project Status: Completed (Fiscal Year 2020)
• Budget Amount: ¥4,030,000 (Direct Cost: ¥3,100,000、Indirect Cost: ¥930,000); Fiscal Year 2020: ¥1,170,000 (Direct Cost: ¥900,000、Indirect Cost: ¥270,000); Fiscal Year 2019: ¥1,170,000 (Direct Cost: ¥900,000、Indirect Cost: ¥270,000); Fiscal Year 2018: ¥1,690,000 (Direct Cost: ¥1,300,000、Indirect Cost: ¥390,000)
• Keywords: 脳 / リカレントニューラルネットワーク / 故障耐性 / 過学習抑制 / 初期構造 / 初期化 / 解剖構造 / 興奮性 / 抑制性 / 対故障性

Outline of Final Research Achievements

In this study, we introduced brain anatomical structures that are ubiquitous across brain regions as initial constraint into recurrent neural network (RNNs). We evaluated their roles in computation and showed that the structural distinction between excitatory and inhibitory neurons contributes to prevent overfitting. Moreover, the partial connectivity contributes to the improvement of fault tolerance of RNNs. These structures complementarily work and improve the performance and fault tolerance of the network.

Academic Significance and Societal Importance of the Research Achievements

脳の様々な領域に共通する解剖構造によるリカレントニューラルネットワーク（RNN）の性能向上と故障に対する頑健性向上への寄与を示した．初期構造として導入するだけで学習後の性能向上につながるため，計算コストが少なくて済む．そのため，これらの知見は，計算資源が乏しく，一部が故障しても性能を維持する必要のある状況におけるRNNに適用でき，適用範囲の広いRNNの初期化手法としての応用が考えられる．

Research Products

• [Presentation] Improvement on Performance of Recurrent Neural Network through Initializing of Input and Output Structures Similar to Partial Connection (2021)
  • Author(s): Toshikazu Samura, Tomohiro Fusauchi
  • Organizer: RISP International Workshop on Nonlinear Circuits, Communications and Signal Processing 2021
• [Presentation] Application of Initialization Method Inspired by Brain Structure to Recurrent Neural Network and Long Short-Term Memory (2019)
  • Author(s): Tomohiro Fusauchi, Toshikazu Samura
  • Organizer: The 42th Annual Meeting of the Japan Neuroscience Society (NEURO2019)
• [Presentation] Suppression of Overfitting in a Recurrent Neural Network by Excitatory-Inhibitory Initializer (2019)
  • Author(s): Tomohiro Fusauchi, Toshikazu Samura
  • Organizer: The 2019 International Symposium on Nonlinear Theory and Its Applications (NOLTA2019)
  • Int'l Joint Research
• [Presentation] Evaluation of performance and robustness of recurrent neural network constrained by anatomical brain structure (2019)
  • Author(s): Fusauchi Tomonori, Toshikazu Samura
  • Organizer: The 5th International Symposium “Green and Smart Technologies for a Sustainable Society"
• [Presentation] Roles of brain-inspired initial constraint on structure of recurrent neural network for its performance and robustness (2019)
  • Author(s): Fusauchi Tomonori, Toshikazu Samura
  • Organizer: The 7th RIEC International Symposium on Brain Functions and Brain Computer
• [Presentation] Recurrent neural network initialized by brain structure improves time series prediction (2018)
  • Author(s): Tomohiro Fusauchi, Toshikazu Samura
  • Organizer: The 41th Annual Meeting of the Japan Neuroscience Society
• [Presentation] Initial Constraint on Structure of Recurrent Neural Network for Improvement of Time Series Prediction (2018)
  • Author(s): Tomohiro Fusauchi, Toshikazu Samura
  • Organizer: The 28th Annual Conference of the Japanese Neural Network Society