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

Theory for unified expression of recognition mechanisms and its application to machine learning

Research Project

Project/Area Number 20H04206
Research Category

Grant-in-Aid for Scientific Research (B)

Allocation TypeSingle-year Grants
Section一般
Review Section Basic Section 61010:Perceptual information processing-related
Research InstitutionTokyo Institute of Technology

Principal Investigator

Yukihiko Yamashita  東京工業大学, 教育革新センター, 教授 (90220350)

Co-Investigator(Kenkyū-buntansha) 杉山 将  東京大学, 大学院新領域創成科学研究科, 教授 (90334515)
田中 聡久  東京農工大学, 工学(系)研究科(研究院), 教授 (70360584)
鷲沢 嘉一  電気通信大学, 大学院情報理工学研究科, 准教授 (10419880)
Project Period (FY) 2020-04-01 – 2023-03-31
Project Status Completed (Fiscal Year 2022)
Budget Amount *help
¥16,640,000 (Direct Cost: ¥12,800,000、Indirect Cost: ¥3,840,000)
Fiscal Year 2022: ¥4,420,000 (Direct Cost: ¥3,400,000、Indirect Cost: ¥1,020,000)
Fiscal Year 2021: ¥4,420,000 (Direct Cost: ¥3,400,000、Indirect Cost: ¥1,020,000)
Fiscal Year 2020: ¥7,800,000 (Direct Cost: ¥6,000,000、Indirect Cost: ¥1,800,000)
Keywords認識機構 / ファイバー束 / パターン認識 / 機械学習 / 脳信号処理 / オンライン教育
Outline of Research at the Start

本研究では,まず,深層学習の理論的な解釈やその先の学習理論の確立のために,信号の確率的性質や特徴量などを同時に扱うことができるファイバー束を用いて,認識機構を統一的・幾何学的に表現する理論を構築する。その中で,ベクトルをファイバーとするベクトル束ばかりでなく,行列(群)をファイバーとするファイバー束を用いて,多様体上の正規分布を定義するために求めた局所等方独立の理論を展開し,認識機構における特徴抽出,画像変換,脳信号処理の理論を構築する共に,ポスト深層学習に向け,深層学習の理論を解釈する。さらに,新たな機械学習の理論を構築し計算アルゴリズムを開発し,得られた理論を実証する。

Outline of Final Research Achievements

We conducted this research to express recognition mechanism by using fiber bundles and apply the expression and knowledge obtained in this research to machine learning. The major results of this research are as follows. (1) We defined the local isotropic independent equation for the fiber bundle whose fiber has an algebraic structure, and investigated its properties. (2) We proposed the concept of local metric preserving learning and showed it can solve the problem of vanishing gradients in deep learning. (3) For the problem with positive and unlabeled data, we defined a loss function using the ratio of probability density functions and showed its detailed properties and advantages. (4) We established a transfer learning theory in the Riemannian manifold and showed its advantages for the epileptic seizure detection. (5) We applied the algorithm used or improved in the research to other fields.

Academic Significance and Societal Importance of the Research Achievements

今日では,機械学習の理論は,自然言語処理,個人認証,文字認識などの広い分野で応用され,現代社会では欠かすことができないものになっている。さらに,生成系AIや自動車の自動運転などは今までの社会を一変させる可能性を秘めている。しかしながら,それらの中で使われている深層学習は,理論的に十分解明されているとは言えない。それを多様体を拡張したファイバー束を使って解明しようとする研究は非常に先進的であり,得られた成果は意義のあるものである。そして,その研究から得られた知見は,現在の機械学習の性能を向上させており,社会的意義も大きい。

Report

(4 results)
  • 2022 Annual Research Report   Final Research Report ( PDF )
  • 2021 Annual Research Report
  • 2020 Annual Research Report
  • Research Products

    (18 results)

All 2023 2022 2021 2020

All Journal Article (15 results) (of which Int'l Joint Research: 5 results,  Peer Reviewed: 13 results,  Open Access: 3 results) Presentation (3 results)

  • [Journal Article] Gaussian Mixture Bandpass Filter Design for Narrow Passband Width by Using a FIR Recursive Filter2023

    • Author(s)
      YAMASHITA Yukihiko
    • Journal Title

      IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences

      Volume: E106.A Issue: 10 Pages: 1277-1285

    • DOI

      10.1587/transfun.2022EAP1108

    • ISSN
      0916-8508, 1745-1337
    • Year and Date
      2023-10-01
    • Related Report
      2022 Annual Research Report
    • Peer Reviewed / Open Access
  • [Journal Article] Representation learning for continuous action spaces is beneficial for efficient policy learning2023

    • Author(s)
      Zhao Tingting、Wang Ying、Sun Wei、Chen Yarui、Niu Gang、Sugiyama Masashi
    • Journal Title

      Neural Networks

      Volume: 159 Pages: 137-152

    • DOI

      10.1016/j.neunet.2022.12.009

    • Related Report
      2022 Annual Research Report
    • Peer Reviewed / Int'l Joint Research
  • [Journal Article] Active selection of source patients in transfer learning for epileptic seizure detection using Riemannian Manifold2023

    • Author(s)
      T. Orihaa, K. M. Hassan, and T. Tanaka
    • Journal Title

      ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

      Volume: -

    • Related Report
      2022 Annual Research Report
    • Peer Reviewed / Int'l Joint Research
  • [Journal Article] Learning from Noisy Complementary Labels with Robust Loss Functions2022

    • Author(s)
      Hiroki Ishiguro, Takashi Ishida, Masashi Sugiyama
    • Journal Title

      IEICE Transactions on Information and Systems

      Volume: E105.D Issue: 2 Pages: 364-376

    • DOI

      10.1587/transinf.2021EDP7035

    • NAID

      130008149868

    • ISSN
      0916-8532, 1745-1361
    • Year and Date
      2022-02-01
    • Related Report
      2021 Annual Research Report
    • Peer Reviewed / Open Access
  • [Journal Article] Class-Wise Denoising for Robust Learning under Label Noise2022

    • Author(s)
      Gong Chen、Ding Yongliang、Han Bo、Niu Gang、Yang Jian、You Jane J.、Tao Dacheng、Sugiyama Masashi
    • Journal Title

      IEEE Transactions on Pattern Analysis and Machine Intelligence

      Volume: 45 Pages: 1-1

    • DOI

      10.1109/tpami.2022.3178690

    • Related Report
      2022 Annual Research Report
    • Peer Reviewed / Int'l Joint Research
  • [Journal Article] Positive-unlabeled classification under class-prior shift: a prior-invariant approach based on density ratio estimation2022

    • Author(s)
      Nakajima Shota、Sugiyama Masashi
    • Journal Title

      Machine Learning

      Volume: 112 Issue: 3 Pages: 889-919

    • DOI

      10.1007/s10994-022-06190-z

    • Related Report
      2022 Annual Research Report
    • Peer Reviewed / Int'l Joint Research
  • [Journal Article] Morlet wavelet transform using attenuated sliding Fourier transform and kernel integral for graphic processing unit2022

    • Author(s)
      Yukihiko Yamashita and Toru Wakahara
    • Journal Title

      arXiv

      Volume: -

    • Related Report
      2021 Annual Research Report
    • Open Access
  • [Journal Article] リーマン多様体による能動学習を用いたてんかん発作検出2022

    • Author(s)
      折原俊貴, 田中聡久
    • Journal Title

      信学技報

      Volume: 121 Pages: 201-206

    • Related Report
      2021 Annual Research Report
  • [Journal Article] Steady-State Visual Evoked Potential Classification Using Complex Valued Convolutional Neural Networks2021

    • Author(s)
      Ikeda Akira, Washizawa Yoshikazu
    • Journal Title

      Sensors

      Volume: 21 Issue: 16 Pages: 5309-5309

    • DOI

      10.3390/s21165309

    • Related Report
      2021 Annual Research Report
    • Peer Reviewed
  • [Journal Article] Stabilized Calculation of Gaussian Smoothing and Its Differentials Using Attenuated Sliding Fourier Transform2021

    • Author(s)
      Yukihiko Yamashita、Toru Wakahara
    • Journal Title

      Proceedings of 2020 25th International Conference on Pattern Recognition (ICPR)

      Volume: -

    • Related Report
      2020 Annual Research Report
    • Peer Reviewed
  • [Journal Article] LocalDrop: A Hybrid Regularization for Deep Neural Networks2021

    • Author(s)
      Lu Ziqing、Xu Chang、Du Bo、Ishida Takashi、Zhang Lefei、Sugiyama Masashi
    • Journal Title

      IEEE Transactions on Pattern Analysis and Machine Intelligence

      Volume: - Pages: 1-1

    • DOI

      10.1109/tpami.2021.3061463

    • Related Report
      2020 Annual Research Report
    • Peer Reviewed
  • [Journal Article] Initial Location Search Method using Target Region Features for Region-Based Image Matching2020

    • Author(s)
      山下幸彦、若原徹
    • Journal Title

      電子情報通信学会論文誌D 情報・システム

      Volume: J103-D Issue: 4 Pages: 330-340

    • DOI

      10.14923/transinfj.2019JDP7056

    • ISSN
      1880-4535, 1881-0225
    • Year and Date
      2020-04-01
    • Related Report
      2020 Annual Research Report
    • Peer Reviewed
  • [Journal Article] Coupling-based invertible neural networks are universal diffeomorphism approximators2020

    • Author(s)
      Takeshi Teshima、Isao Ishikawa、 Koichi Tojo、 Kenta Oono、Masahiro Ikeda、Masashi Sugiyama
    • Journal Title

      Advances in Neural Information Processing Systems

      Volume: 33 Pages: 3362-3373

    • Related Report
      2020 Annual Research Report
    • Peer Reviewed
  • [Journal Article] Centroid Estimation with Guaranteed Efficiency: A General Framework for Weakly Supervised Learning2020

    • Author(s)
      Gong Chen、Yang Jian、You Jane J.、Sugiyama Masashi
    • Journal Title

      IEEE Transactions on Pattern Analysis and Machine Intelligence

      Volume: - Issue: 6 Pages: 1-1

    • DOI

      10.1109/tpami.2020.3044997

    • Related Report
      2020 Annual Research Report
    • Peer Reviewed
  • [Journal Article] Detecting EEG outliers for BCI on the Riemannian manifold using spectral clustering2020

    • Author(s)
      Yamamoto Maria Sayu、Sadatnejad Khadijeh、Tanaka Toshihisa、Islam Rabiul、Tanaka Yuichi、Lotte Fabien
    • Journal Title

      2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC)

      Volume: - Pages: 438-414

    • DOI

      10.1109/embc44109.2020.9175456

    • Related Report
      2020 Annual Research Report
    • Peer Reviewed / Int'l Joint Research
  • [Presentation] スライディングフーリエ変換によるガウス平滑化のGPU実装2022

    • Author(s)
      山下幸彦, 若原徹
    • Organizer
      電子情報通信学会総合大会
    • Related Report
      2021 Annual Research Report
  • [Presentation] Positive-unlabeled classification under class-prior shift: A prior-invariant approach based on density ratio estimation2021

    • Author(s)
      Shota Nakajima, Masashi Sugiyama
    • Organizer
      t IJCAI2021 Weakly Supervised Representation Learning Workshop
    • Related Report
      2021 Annual Research Report
  • [Presentation] AUC Maximization in Deep Neural Network Learning for Imbalanced Classification Problems2021

    • Author(s)
      Shuyu Zhang and Yoshikazu Washizawa
    • Organizer
      電子情報通信学会総合大会
    • Related Report
      2020 Annual Research Report

URL: 

Published: 2020-04-28   Modified: 2024-01-30  

Information User Guide FAQ News Terms of Use Attribution of KAKENHI

Powered by NII kakenhi