大脳皮質局所回路に学ぶ新しいアーキテクチャと学習モデルの構築

• Project Area: Correspondence and Fusion of Artificial Intelligence and Brain Science
• Project/Area Number: 17H06036
• 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: Institute of Physical and Chemical Research
• Principal Investigator: 深井 朋樹 国立研究開発法人理化学研究所, 脳神経科学研究センター, チームリーダー (40218871)
• Project Period (FY): 2017-04-01 – 2019-03-31
• Project Status: Completed (Fiscal Year 2018)
• Budget Amount: ¥24,440,000 (Direct Cost: ¥18,800,000、Indirect Cost: ¥5,640,000); Fiscal Year 2018: ¥12,480,000 (Direct Cost: ¥9,600,000、Indirect Cost: ¥2,880,000); Fiscal Year 2017: ¥11,960,000 (Direct Cost: ¥9,200,000、Indirect Cost: ¥2,760,000)
• Keywords: 樹状突起計算 / 学習と記憶 / 自発発火 / ベイズ推定 / 合目的行動 / リザーバ計算 / スパイク時間 / 構造的シナプス可塑性 / 領野間連絡 / チャンク / 記憶学習 / 予測符号化 / 時間的神経情報 / チャンク学習 / 大脳皮質局所回路

Outline of Annual Research Achievements

記憶と脳の階層計算（予測符号化）の統合に向けて以下の研究を行った。
1.時系列入力の特徴を学習するミニマムな神経回路機構として、樹状突起をもつニューロンモデルを開発した。樹状突起から細胞体への「逆伝搬信号」を受けながら、両者の活動のミスマッチを最小化するように確率的学習を行うと、時系列入力に繰り返し出現するパターンを検出できること、またこれによりチャンクが検出できることを示した（Asabuki and Fukai, bioRxiv 517888）。
2.上記モデルにより、ICAなど従来手法では分離できない、相関を持つ混合信号を分離することができることを示した。さらに、多様な時間発展パターンをもつ信号の分離に於いては、短期シナプス可塑性が重要であることを明らかにした。
3.脳は多様な時間スケールで起こる事象を相互に関連付けながら、まとまりのあるエピソードとして記憶する。この機能が生じるメカニズムを理解するために、新しい連想記憶モデルの枠組みを構築した（Haga and Fukai, arXiv 1809.05254）。このモデルでは、興奮性シナプス入力と抑制性シナプス入力のバランス、さらに局所的抑制と大域的抑制の相対バランスを調節することにより、時間的に連続した事象を記銘するアトラクター状態の、空間的な相関長を自在に制御できる。またモデルの学習則は海馬CA3で発見された対称STDPと整合性が高い。
4.思決定における動物の個体差を、同様の課題をリザーバ計算機に訓練することにより、再現してみせた（Kurikawa et al., Nat Neurosci, 2018）。この結果は個体差の表出に神経回路ダイナミクスが影響する可能性を示すとともに、同一ルールで訓練された人工知能でも、高次機能に於いて個体差が現れ得ることを示唆している。

Research Progress Status

平成30年度が最終年度であるため、記入しない。

Strategy for Future Research Activity

平成30年度が最終年度であるため、記入しない。

Research Products

• [Journal Article] Hippocampal Neurogenesis Reduces the Dimensionality of Sparsely Coded Representations to Enhance Memory Encoding (2019)
  • Author(s): DeCostanzo Anthony J.、Fung Chi Chung Alan、Fukai Tomoki
  • Journal Title: Frontiers in Computational Neuroscience Volume: 12 Pages: 1-21
  • DOI: 10.3389/fncom.2018.00099
  • Peer Reviewed / Open Access
• [Journal Article] Discrete-Attractor-like Tracking in Continuous Attractor Neural Networks (2019)
  • Author(s): Fung Chi Chung Alan、Fukai Tomoki
  • Journal Title: Physical Review Letters Volume: 122 Issue: 1 Pages: 1-5
  • DOI: 10.1103/physrevlett.122.018102
  • Peer Reviewed
• [Journal Article] Learning Task-Related Activities From Independent Local-Field-Potential Components Across Motor Cortex Layers (2018)
  • Author(s): Martin-Vazquez Gonzalo、Asabuki Toshitake、Isomura Yoshikazu、Fukai Tomoki
  • Journal Title: Frontiers in Neuroscience Volume: 12 Pages: 429-429
  • DOI: 10.3389/fnins.2018.00429
  • Peer Reviewed / Open Access / Int'l Joint Research
• [Journal Article] Recurrent network model for learning goal-directed sequences through reverse replay (2018)
  • Author(s): Haga Tatsuya、Fukai Tomoki
  • Journal Title: eLife Volume: 7 Pages: 34171-34171
  • DOI: 10.7554/elife.34171
  • Peer Reviewed / Open Access
• [Journal Article] Transient and Persistent UP States during Slow-wave Oscillation and their Implications for Cell-Assembly Dynamics (2018)
  • Author(s): Fung Chi Chung Alan、Fukai Tomoki
  • Journal Title: Scientific Reports Volume: 8 Issue: 1 Pages: 10680-10680
  • DOI: 10.1038/s41598-018-28973-y
  • Peer Reviewed / Open Access
• [Journal Article] Redundancy in synaptic connections enables neurons to learn optimally (2018)
  • Author(s): Hiratani Naoki、Fukai Tomoki
  • Journal Title: Proceedings of the National Academy of Sciences Volume: 115 Issue: 29
  • DOI: 10.1073/pnas.1803274115
  • Peer Reviewed / Int'l Joint Research
• [Journal Article] Interactive reservoir computing for chunking information streams (2018)
  • Author(s): Asabuki Toshitake、Hiratani Naoki、Fukai Tomoki
  • Journal Title: PLOS Computational Biology Volume: 14 Issue: 10 Pages: e1006400-e1006400
  • DOI: 10.1371/journal.pcbi.1006400
  • Peer Reviewed / Open Access / Int'l Joint Research
• [Journal Article] Dendritic processing of spontaneous neuronal sequences for single-trial learning (2018)
  • Author(s): Haga Tatsuya、Fukai Tomoki
  • Journal Title: Scientific Reports Volume: 8 Issue: 1 Pages: 15166-15166
  • DOI: 10.1038/s41598-018-33513-9
  • Peer Reviewed / Open Access
• [Journal Article] Neuronal stability in medial frontal cortex sets individual variability in decision-making (2018)
  • Author(s): Kurikawa Tomoki、Haga Tatsuya、Handa Takashi、Harukuni Rie、Fukai Tomoki
  • Journal Title: Nature Neuroscience Volume: 21 Issue: 12 Pages: 1764-1773
  • DOI: 10.1038/s41593-018-0263-5
  • Peer Reviewed / Int'l Joint Research
• [Presentation] Implications of symmetric STDP, replay and inhibitory engrams in sequence learning (2019)
  • Author(s): Tomoki Fukai
  • Organizer: Computational and Systems Neuroscience (Cosyne) 2019 Workshop
  • Int'l Joint Research / Invited
• [Presentation] 自己条件付きエントロピー最小化による時系列情報の特徴学習 (2019)
  • Author(s): 朝吹俊丈、深井朋樹
  • Organizer: 日本物理学会第74回年次大会
• [Presentation] 動的均衡にある神経回路の理論解析とレザボワ計算 (2019)
  • Author(s): 早川隆、深井朋樹
  • Organizer: 日本物理学会第74回年次大会
• [Presentation] Learning information streams in brain-inspired neural network models (2019)
  • Author(s): Tomoki Fukai
  • Organizer: HangZhou Forum on Cognitive Neuroscience and Intelligent Application 2019
  • Int'l Joint Research / Invited
• [Presentation] A neural circuit model of pattern classification with dendritic activities (2018)
  • Author(s): Kento Suzuki, Tatsuya Haga, Tomoki Fukai
  • Organizer: 第41回日本神経科学大会
• [Presentation] Chunking information streams by dendritic minimization of surprise signals (2018)
  • Author(s): Toshitake Asabuki, Tomoki Fukai
  • Organizer: 第41回日本神経科学大会
• [Presentation] Spontaneous and stimulus-induced coherent states of dynamically balanced neuronal networks (2018)
  • Author(s): Takashi Hayakawa, Tomoki Fukai
  • Organizer: 第41回日本神経科学大会
• [Presentation] Sequence learning by symmetric STDP and inhibitory engram (2018)
  • Author(s): Tomoki Fukai
  • Organizer: 18th China-Japan-Korea Joint Workshop on Neurobiology and Neuroinformatics (NBNI2018)
  • Int'l Joint Research / Invited
• [Presentation] Unsupervised learning of information streams in networks of dendritic neurons (2018)
  • Author(s): Tomoki Fukai
  • Organizer: Champalimaud Research Symposium 2018
  • Int'l Joint Research / Invited
• [Presentation] Analysis of complex temporal information streams by dendritic neuron model (2018)
  • Author(s): Toshitake Asabuki, Tomoki Fukai
  • Organizer: Neuroscience 2018 (SfN's 48th annual meeting)
  • Int'l Joint Research
• [Presentation] Learning efficient search for reward by CA3 recurrent network model (2018)
  • Author(s): Tatuya Haga, Tomoki Fukai
  • Organizer: Neuroscience 2018 (SfN's 48th annual meeting)
  • Int'l Joint Research
• [Presentation] Exploring a bridge between the brain and AI (2018)
  • Author(s): Tomoki Fukai
  • Organizer: Global Leaders Forum 2018
  • Int'l Joint Research / Invited
• [Presentation] Unsupervised learning of temporal information streams in networks of dendritic neurons (2018)
  • Author(s): Toshitake Asabuki, Tomoki Fukai
  • Organizer: 新学術領域研究 「人工知能と脳科学」第５回領域会議
• [Presentation] Cortical network models from small scales to large scales for cognition (2018)
  • Author(s): Tomoki Fukai
  • Organizer: Orchestrating Networks
  • Int'l Joint Research / Invited
• [Presentation] Dendritic canonical correlation analysis for memory and sensory information processing (2018)
  • Author(s): Tomoki Fukai
  • Organizer: Dendritic integration and computation with active dendrites
  • Int'l Joint Research / Invited
• [Presentation] Sequence learning through reverse replay and preplay in hippocampal circuit models (2018)
  • Author(s): Tomoki Fukai
  • Organizer: Institut d’Investigacions Biom?diques August Pi i Sunyer (IDIBAPS)
  • Int'l Joint Research / Invited
• [Presentation] Reservoir computing for chunking and decision-making (2018)
  • Author(s): Tomoki Fukai
  • Organizer: Colloquium at Laboratoire de Neurosciences Cognitives, Ecole Normale Superieure, Paris
  • Int'l Joint Research / Invited
• [Presentation] 相互に教え合うリザーバ計算モジュールによるチャンク学習 (2018)
  • Author(s): 朝吹俊丈, 平谷直輝, 深井朋樹
  • Organizer: 日本物理学会 第73回年次大会（2018年）
• [Presentation] Dendritic computing gives a meso-scopic level framework of brain's leaning rule (2017)
  • Author(s): Tatsuya Haga and Tomoki Fuka
  • Organizer: Forum on Cognitive Neuroscience in Hangzhou
  • Int'l Joint Research / Invited
• [Presentation] ２コンパートメントニューラルネットワークモデルにおける頑健なシーケンス記憶の形成 Robust formation of sequential memory in two-compartment neural network model (2017)
  • Author(s): 芳賀 達也 / Tatsuya Haga, 深井 朋樹 / Tomoki Fukai
  • Organizer: Neuroscience2017 第40回日本神経科学大会
  • Int'l Joint Research
• [Presentation] リカレント神経回路による時系列からの教師無しチャンク学習 Unsupervised chunking learning from sequences by recurrent neural networks (2017)
  • Author(s): 朝吹 俊丈 / Toshitake Asabuki, 深井 朋樹 / Tomoki Fukai
  • Organizer: Neuroscience2017 第40回日本神経科学大会
  • Int'l Joint Research
• [Presentation] 順伝播型深層神経回路網における樹状突起の計算 Dendritic computing in deep feedforward neural networks (2017)
  • Author(s): 鈴木 堅斗 / Kento Suzuk 芳賀 達也 / Tatsuya Haga 深井 朋樹 / Tomoki Fukai
  • Organizer: Neuroscience2017 第40回日本神経科学大会
  • Int'l Joint Research
• [Presentation] Chunking by mutual supervision in reservoir computing (2017)
  • Author(s): Toshitake Asabuki, Naoki Hiratani, Tomoki Fukai
  • Organizer: Topical Problems of Nonlinear Wave Physics (NWP-2017) - Nonlinear Dynamics and Complexity (NWP-1)
  • Int'l Joint Research / Invited
• [Presentation] 外界をモデル化する脳の回路メカニズム ―海馬と大脳皮質 (2017)
  • Author(s): 深井朋樹
  • Organizer: 第54回 脳科学ライフサポート研究センターセミナー
  • Invited
• [Presentation] 大脳皮質と海馬の神経集団活動から読み解く回路機能 (2017)
  • Author(s): 深井朋樹
  • Organizer: 平成29年度生理学研究所研究会「大脳皮質回路の機能原理を探る」
  • Invited
• [Presentation] Interplay of microscopic and macroscopic dynamics in randomly connected neuronal networks (2017)
  • Author(s): Takashi Hayakawa, Tomoki Fukai
  • Organizer: Bernstein Conference 2017
  • Int'l Joint Research
• [Presentation] Chunk learning from complex sequences by mutually supervising recurrent neural networks (2017)
  • Author(s): Toshitake Asabuki, Tomoki Fukai
  • Organizer: Neuroscience 2017 (SfN's 47th annual meeting)
  • Int'l Joint Research
• [Presentation] A local supervised learning rule protects memories from catastrophic interference during subsequent unsupervised learning (2017)
  • Author(s): Anthony DeCostanzo, Tomoki Fukai
  • Organizer: Neuroscience 2017 (SfN's 47th annual meeting)
  • Int'l Joint Research
• [Presentation] Reverse replay strengthens forward pathways to reward through Hebbian learning and short-term depression (2017)
  • Author(s): Tatsuya Haga, Tomoki Fukai
  • Organizer: Neuroscience 2017 (SfN's 47th annual meeting)
  • Int'l Joint Research
• [Presentation] Unsupervised reservoir computing for chunking sequence information (2017)
  • Author(s): Tomoki Fukai
  • Organizer: NBNI 2017 (17th China-Japan-Korea Joint Workshop on Neurobiology and Neuroinformatics)
  • Int'l Joint Research