• Search Research Projects
  • Search Researchers
  • How to Use
  1. Back to previous page

Mathematical Foundations of Random Deep Neural Networks and their applications to machine-learning problems

Research Project

Project/Area Number 19K20366
Research Category

Grant-in-Aid for Early-Career Scientists

Allocation TypeMulti-year Fund
Review Section Basic Section 61040:Soft computing-related
Research InstitutionNational Institute of Advanced Industrial Science and Technology

Principal Investigator

Karakida Ryo  国立研究開発法人産業技術総合研究所, 情報・人間工学領域, 主任研究員 (30803902)

Project Period (FY) 2019-04-01 – 2023-03-31
Project Status Completed (Fiscal Year 2022)
Budget Amount *help
¥4,160,000 (Direct Cost: ¥3,200,000、Indirect Cost: ¥960,000)
Fiscal Year 2022: ¥1,040,000 (Direct Cost: ¥800,000、Indirect Cost: ¥240,000)
Fiscal Year 2021: ¥1,040,000 (Direct Cost: ¥800,000、Indirect Cost: ¥240,000)
Fiscal Year 2020: ¥1,040,000 (Direct Cost: ¥800,000、Indirect Cost: ¥240,000)
Fiscal Year 2019: ¥1,040,000 (Direct Cost: ¥800,000、Indirect Cost: ¥240,000)
Keywords深層学習 / 機械学習 / ニューラルネットワーク / 統計力学 / ランダム行列 / 数理工学 / 統計力学的解析 / レプリカ法 / 継続学習
Outline of Research at the Start

深層モデルは高次元の非線形変換を繰り返すため, そのままでは数学的な取り扱いが困難で, 動作原理はブラックボックスである. この問題に対し, ランダム結合パラメータを持つモデルでは複雑な動作を粗視化し, 少数次元の理論式に縮約できるため, この困難を克服できると期待される. また, 粗視化によって, モデルや学習の設定の詳細に依存しない普遍的な数理的基礎付けが実現できると考えられる. 本研究では, まず, モデルの学習のしやすさや汎化性能に関係した幾何構造を解析し, 深層モデルの情報処理原理の解明を目指す. 次に, 構築された理論に基づき, 学習手法への応用を行う.

Outline of Final Research Achievements

The purpose of this research is to gain mathematical insights for deep learning, based on the analysis of random neural networks. Toward this goal, we first analyzed the eigenvalues of their Fisher information matrix, which determine the geometric structure of the parameter space. This allowed us to provide a quantitative explanation of the effects of normalization layers and appropriate settings for learning rates. In the NTK regime, characterized by learning within the range of perturbation around initial random weights, we clarified the appropriate designs of approximated natural gradient methods. Related to associative memory models, we elucidated Boltzmann machines corresponding to Modern Hopfield networks and the memory recall process in VAEs.

Academic Significance and Societal Importance of the Research Achievements

ランダム神経回路は古典的に理論的神経科学の枠組みで発展してきたが, 近年は深層学習にその枠組みを拡張し, たとえば逆誤差伝播における解析が進みつつある. 本研究課題もこの流れに沿うもので, 特に, 学習のプロセスに大きく影響を与えるFisher情報行列やNTK行列に着目し, 各種モデルや学習手法の性質を明らかにした点に独自性があり学術的意義がある. 本成果は様々な応用を支える基礎技術に理解を与えており, 今後の深層学習技術の研究開発を進めるうえで有用となることが期待でき, その点で社会的意義もあるといえるだろう.

Report

(5 results)
  • 2022 Annual Research Report   Final Research Report ( PDF )
  • 2021 Research-status Report
  • 2020 Research-status Report
  • 2019 Research-status Report
  • Research Products

    (35 results)

All 2023 2022 2021 2020 2019 Other

All Int'l Joint Research (1 results) Journal Article (13 results) (of which Peer Reviewed: 12 results,  Open Access: 10 results) Presentation (18 results) (of which Int'l Joint Research: 4 results,  Invited: 8 results) Book (1 results) Remarks (2 results)

  • [Int'l Joint Research] チューリッヒ工科大学(スイス)

    • Related Report
      2022 Annual Research Report
  • [Journal Article] Deep learning in random neural fields: Numerical experiments via neural tangent kernel2023

    • Author(s)
      Watanabe, Kaito and Sakamoto, Kotaro and Karakida, Ryo and Sonoda, Sho and Amari, Shun-ichi
    • Journal Title

      Neural Networks

      Volume: 160 Pages: 148

    • DOI

      10.1016/j.neunet.2022.12.020

    • Related Report
      2022 Annual Research Report
    • Peer Reviewed
  • [Journal Article] Attention in a family of Boltzmann machines emerging from modern Hopfield networks2023

    • Author(s)
      Toshihiro Ota, Ryo Karakida
    • Journal Title

      Neural Computation

      Volume: -

    • Related Report
      2022 Annual Research Report
  • [Journal Article] Learning Curves for Continual Learning in Neural Networks: Self-Knowledge Transfer and Forgetting2022

    • Author(s)
      Ryo Karakida, Shotaro Akaho
    • Journal Title

      International Conference on Learning Representations

      Volume: - Pages: 1-27

    • Related Report
      2022 Annual Research Report
    • Peer Reviewed / Open Access
  • [Journal Article] The Spectrum of Fisher Information of Deep Networks Achieving Dynamical Isometry2021

    • Author(s)
      Tomohiro Hayase, Ryo Karakida
    • Journal Title

      Proceedings of AISTATS

      Volume: -

    • Related Report
      2021 Research-status Report
    • Peer Reviewed / Open Access
  • [Journal Article] Pathological spectra of the fisher information metric and its variants in deep neural networks2021

    • Author(s)
      Karakida Ryo, Akaho Shotaro, Amari Shun-ichi
    • Journal Title

      Neural Computation

      Volume: 33 Issue: 8 Pages: 2274

    • DOI

      10.1162/neco_a_01411

    • Related Report
      2021 Research-status Report
    • Peer Reviewed / Open Access
  • [Journal Article] Understanding approximate Fisher information for fast convergence of natural gradient descent in wide neural networks*2021

    • Author(s)
      Karakida Ryo, Osawa Kazuki
    • Journal Title

      Journal of Statistical Mechanics: Theory and Experiment

      Volume: 2021 Issue: 12 Pages: 124010-124010

    • DOI

      10.1088/1742-5468/ac3ae3

    • Related Report
      2021 Research-status Report
    • Peer Reviewed / Open Access
  • [Journal Article] Self-paced data augmentation for training neural networks2021

    • Author(s)
      Tomoumi Takase, Ryo Karakida, Hideki Asoh
    • Journal Title

      Neurocomputing

      Volume: 442 Pages: 296-306

    • DOI

      10.1016/j.neucom.2021.02.080

    • Related Report
      2021 Research-status Report
    • Peer Reviewed / Open Access
  • [Journal Article] Understanding Approximate Fisher Information for Fast Convergence of Natural Gradient Descent in Wide Neural Networks2020

    • Author(s)
      Ryo Karakida, Kazuki Osawa
    • Journal Title

      Proceedings of Conference on Neural Information Processing Systems (NeurIPS)

      Volume: 33

    • Related Report
      2020 Research-status Report
    • Peer Reviewed / Open Access
  • [Journal Article] Collective dynamics of repeated inference in variational autoencoder rapidly find cluster structure2020

    • Author(s)
      Nagano Yoshihiro、Karakida Ryo、Okada Masato
    • Journal Title

      Scientific Reports

      Volume: 10 Issue: 1

    • DOI

      10.1038/s41598-020-72593-4

    • Related Report
      2020 Research-status Report
    • Peer Reviewed / Open Access
  • [Journal Article] Universal Statistics of Fisher Information in Deep Neural Networks: Mean Field Approach2019

    • Author(s)
      Ryo Karakida, Shotaro Akaho, Shun-ichi Amari
    • Journal Title

      Proceedings of Conference on Artificial Intelligence and Statistics

      Volume: -

    • Related Report
      2019 Research-status Report
    • Peer Reviewed / Open Access
  • [Journal Article] Fisher Information and Natural Gradient Learning in Random Deep Networks2019

    • Author(s)
      Shun-ichi Amari, Ryo Karakida, Masafumi Oizumi
    • Journal Title

      Proceedings of Conference on Artificial Intelligence and Statistics

      Volume: -

    • Related Report
      2019 Research-status Report
    • Peer Reviewed / Open Access
  • [Journal Article] Statistical neurodynamics of deep networks: geometry of signal spaces2019

    • Author(s)
      Shun-ichi Amari, Ryo Karakida, Masafumi Oizumi
    • Journal Title

      Nonlinear Theory and Its Applications, IEICE

      Volume: 10 Issue: 4 Pages: 322-336

    • DOI

      10.1587/nolta.10.322

    • NAID

      130007722599

    • ISSN
      2185-4106
    • Related Report
      2019 Research-status Report
    • Peer Reviewed
  • [Journal Article] The Normalization Method for Alleviating Pathological Sharpness in Wide Neural Networks2019

    • Author(s)
      Ryo Karakida, Shotaro Akaho, Shun-ichi Amari
    • Journal Title

      Proceedings of Conference on Neural Information Processing Systems

      Volume: -

    • Related Report
      2019 Research-status Report
    • Peer Reviewed / Open Access
  • [Presentation] 対角線形ネットにおける勾配正則化の陰的バイアス2023

    • Author(s)
      唐木田亮, 高瀬朝海, 早瀬友裕, 大沢和樹
    • Organizer
      日本物理学会2023年春季大会
    • Related Report
      2022 Annual Research Report
  • [Presentation] カーネル法の統計力学的解析とそれによる継続学習の評価2022

    • Author(s)
      唐木田亮
    • Organizer
      統計物理と統計科学のセミナー
    • Related Report
      2022 Annual Research Report
    • Invited
  • [Presentation] Neural tangent kernel regimeにおける継続学習の学習曲線2022

    • Author(s)
      唐木田亮
    • Organizer
      日本応用数理学会2022年度年会
    • Related Report
      2022 Annual Research Report
    • Invited
  • [Presentation] 継続学習における自己知識転移と忘却2022

    • Author(s)
      唐木田亮
    • Organizer
      第51回統計的機械学習セミナ-
    • Related Report
      2022 Annual Research Report
    • Invited
  • [Presentation] 効率的な勾配正則化アルゴリズムとその陰的バイアスの解析2022

    • Author(s)
      唐木田亮, 高瀬朝海, 早瀬友裕, 大沢和樹
    • Organizer
      IBIS2022
    • Related Report
      2022 Annual Research Report
  • [Presentation] The Spectrum of Fisher Information of Deep Networks Achieving Dynamical Isometry2021

    • Author(s)
      Tomohiro Hayase, Ryo Karakida
    • Organizer
      AISTATS
    • Related Report
      2021 Research-status Report
    • Int'l Joint Research
  • [Presentation] Improving the trainability of deep neural networks: A perspective from the infinite width limit2021

    • Author(s)
      Ryo Karakida
    • Organizer
      4th international conference on econometrics and statistics
    • Related Report
      2021 Research-status Report
    • Int'l Joint Research / Invited
  • [Presentation] 継続学習における転移と忘却: NTK regimeのレプリカ解析2021

    • Author(s)
      唐木田 亮, 赤穗 昭太郎
    • Organizer
      IBIS2021
    • Related Report
      2021 Research-status Report
  • [Presentation] Understanding Approximate Fisher Information for Fast Convergence of Natural Gradient Descent in Wide Neural Networks2021

    • Author(s)
      Ryo Karakida
    • Organizer
      Math Machine Learning Seminar MPI MIS + UCLA
    • Related Report
      2020 Research-status Report
    • Invited
  • [Presentation] 幅無限大深層モデルにおける近似自然勾配法の収束解析2021

    • Author(s)
      唐木田 亮
    • Organizer
      日本物理学会 第76回年次大会
    • Related Report
      2020 Research-status Report
  • [Presentation] 深層学習の数理: ランダム行列と統計力学的視点2020

    • Author(s)
      唐木田 亮
    • Organizer
      Random Matrices, Free Probability, and Machine Learning ワークショップ
    • Related Report
      2020 Research-status Report
    • Invited
  • [Presentation] 深層学習の数理: 統計力学的アプローチ2020

    • Author(s)
      唐木田 亮
    • Organizer
      ディープラーニングと物理学2020 オンライン
    • Related Report
      2020 Research-status Report
    • Invited
  • [Presentation] 深層モデルにおいて高速に収束する近似自然勾配法の理論解析2020

    • Author(s)
      唐木田 亮, 大沢 和樹
    • Organizer
      IBIS2020
    • Related Report
      2020 Research-status Report
  • [Presentation] Understanding Approximate Fisher Information for Fast Convergence of Natural Gradient Descent in Wide Neural Networks2020

    • Author(s)
      Ryo Karakida
    • Organizer
      Conference on Neural Information Processing Systems (NeurIPS)
    • Related Report
      2020 Research-status Report
    • Int'l Joint Research
  • [Presentation] ランダムなBackpropagation学習における巨視的ダイナミクスの生成汎関数法的解析2020

    • Author(s)
      唐木田亮
    • Organizer
      日本物理学会 第75回年次大会
    • Related Report
      2019 Research-status Report
  • [Presentation] Fisher Information of Deep Neural Networks With Random Weights2019

    • Author(s)
      Ryo Karakida
    • Organizer
      The 11th ICSA international conference
    • Related Report
      2019 Research-status Report
    • Int'l Joint Research / Invited
  • [Presentation] Universal Statistics of Fisher Information in Deep Neural Networks: Mean Field Approach2019

    • Author(s)
      Ryo Karakida, Shotaro Akaho, Shun-ichi Amari
    • Organizer
      Conference on Artificial Intelligence and Statistics
    • Related Report
      2019 Research-status Report
  • [Presentation] The Normalization Method for Alleviating Pathological Sharpness in Wide Neural Networks2019

    • Author(s)
      Ryo Karakida, Shotaro Akaho, Shun-ichi Amari
    • Organizer
      Conference on Neural Information Processing Systems
    • Related Report
      2019 Research-status Report
  • [Book] 数理科学 深層神経回路網の幾何~ 統計神経力学とのつながり ~2020

    • Author(s)
      唐木田 亮
    • Total Pages
      7
    • Publisher
      サイエンス社
    • Related Report
      2020 Research-status Report
  • [Remarks] 研究者が作成したwebページ

    • URL

      https://sites.google.com/view/ryokarakida/english

    • Related Report
      2021 Research-status Report
  • [Remarks] 唐木田亮 website

    • URL

      https://sites.google.com/view/ryokarakida/

    • Related Report
      2020 Research-status Report

URL: 

Published: 2019-04-18   Modified: 2024-01-30  

Information User Guide FAQ News Terms of Use Attribution of KAKENHI

Powered by NII kakenhi