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

Model Switching Criteria in Dynamic Model Learning of Neural Networks

Research Project

Project/Area Number 19K12151
Research Category

Grant-in-Aid for Scientific Research (C)

Allocation TypeMulti-year Fund
Section一般
Review Section Basic Section 61040:Soft computing-related
Research InstitutionUniversity of Tsukuba

Principal Investigator

Kameyama Keisuke  筑波大学, システム情報系, 教授 (40242309)

Project Period (FY) 2019-04-01 – 2023-03-31
Project Status Completed (Fiscal Year 2022)
Budget Amount *help
¥4,420,000 (Direct Cost: ¥3,400,000、Indirect Cost: ¥1,020,000)
Fiscal Year 2021: ¥1,040,000 (Direct Cost: ¥800,000、Indirect Cost: ¥240,000)
Fiscal Year 2020: ¥2,080,000 (Direct Cost: ¥1,600,000、Indirect Cost: ¥480,000)
Fiscal Year 2019: ¥1,300,000 (Direct Cost: ¥1,000,000、Indirect Cost: ¥300,000)
Keywordsニューラルネットワーク / 動的モデル学習 / 転移学習 / 蒸留 / カリキュラム学習 / モデル選択
Outline of Research at the Start

従来,複雑な問題への適応や転移学習などの手段として用いられてきたニューラルネットワークのモデル切り替えを,固定したモデルによらない動的モデル学習法として統一的に整理するとともに,どのようにモデルを切り替えて学習していくことが学習効率や汎化能力の向上につながるのかを検討する.切り替え時点の決定に向けては,学習中に得られる有用な特徴量を見出し,それらを用いて適した切り替え時点を検出する.切り替えモデルと写像の選択に向けては,切り替え候補写像を選択するための写像間距離について検討するとともに,多様なモデル間での切り替え法を提案する.これらの後に実世界の認識問題へ適用し,その有用性を評価する.

Outline of Final Research Achievements

This project concerns the “dynamical model training” of neural networks which involves a training process that switches multiple models and maps implemented during a single learning process. It aimed to clarify the methods and conditions for obtaining a superior map in the desired model in an efficient manner. In model compression using dynamical model training in convolutional neural networks (CNNs) for image classification, methods using multi-stage switching and map selection by distillation with selective transfer of meritorious nature have been proposed. There, it was clarified that a gradual change in the model size and merit-based distillation contribute to the improvement of the performances of the compressed networks. In addition, a method for modality selection in image classification using multimodal features was proposed, and its effectiveness was shown.

Academic Significance and Societal Importance of the Research Achievements

本研究は,パターン認識系の構築における蒸留や転移学習などに見られる複数のニューラルネットワークモデルを経由した学習過程について,それらに共通するモデル切り替え時のモデルと写像の選択方法について学習効率と認識精度の観点から検討を加え,新たな学習方式を提案して学習の効率化と認識精度の向上が可能となることを示しており,単一のモデルに限定されない学習方式の実用化に貢献する成果である.このことは,一定の能力を持つパターン認識系を従来より計算リソースの限られたデバイス上に実現することを可能にするもので,タブレット端末やIoTデバイス等への高度なパターン認識系の実装に資するものである.

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

    (13 results)

All 2023 2022 2021

All Journal Article (7 results) (of which Peer Reviewed: 7 results,  Open Access: 1 results) Presentation (6 results) (of which Invited: 1 results)

  • [Journal Article] Machine Learning Curriculums Generated by Classifier Ensembles2023

    • Author(s)
      Tzu-Jui Huang and Keisuke Kameyama
    • Journal Title

      Proc. 19th IEEE International Colloquium on Signal Processing and its Applications

      Volume: 1 Pages: 117-121

    • DOI

      10.1109/cspa57446.2023.10087822

    • Related Report
      2022 Annual Research Report
    • Peer Reviewed
  • [Journal Article] CNN Model Compression by Merit-Based Distillation2023

    • Author(s)
      Takumi Morikawa and Keisuke Kameyama
    • Journal Title

      Proc. 19th IEEE International Colloquium on Signal Processing and its Applications

      Volume: 1 Pages: 122-127

    • DOI

      10.1109/cspa57446.2023.10087390

    • Related Report
      2022 Annual Research Report
    • Peer Reviewed
  • [Journal Article] LSTM-based forecasting using policy stringency and time-varying parameters of the SIR model for COVID-192023

    • Author(s)
      Pavodi Ndoyi Maniamfu and Keisuke Kameyama
    • Journal Title

      Proc. 19th IEEE International Colloquium on Signal Processing and its Applications

      Volume: 1 Pages: 111-116

    • DOI

      10.1109/cspa57446.2023.10087773

    • Related Report
      2022 Annual Research Report
    • Peer Reviewed
  • [Journal Article] Informative Band Subset Selection for Hyperspectral Image Classification using Joint and Conditional Mutual Information2022

    • Author(s)
      U. A. Md. Ehsan Ali and Keisuke Kameyama
    • Journal Title

      Proc. IEEE Symposium Series on Computational Intelligence

      Volume: 1 Pages: 573-580

    • DOI

      10.1109/ssci51031.2022.10022154

    • Related Report
      2022 Annual Research Report
    • Peer Reviewed
  • [Journal Article] Multi-Stage Model Compression using Teacher Assistant and Distillation with Hint-Based Training2022

    • Author(s)
      Takumi Morikawa and Keisuke Kameyama
    • Journal Title

      Workshop on Pervasive and Resource-constrained AI (PerConAI) part of the 20th IEEE International Conference on Pervasive Computing and Communications (PerCom 2022)

      Volume: 1 Pages: 484-490

    • Related Report
      2021 Research-status Report
    • Peer Reviewed
  • [Journal Article] Adaptive Selection of Classifiers for Person Recognition by Iris Pattern and Periocular Image2021

    • Author(s)
      Keita Ogawa and Keisuke Kameyama
    • Journal Title

      Neural Information Processing. ICONIP 2021. Lecture Notes in Computer Science

      Volume: 13111 Pages: 656-667

    • DOI

      10.1007/978-3-030-92273-3_54

    • ISBN
      9783030922726, 9783030922733
    • Related Report
      2021 Research-status Report
    • Peer Reviewed
  • [Journal Article] Pre-Training Acquisition Functions by Deep Reinforcement Learning for Fixed Budget Active Learning2021

    • Author(s)
      Yusuke TAGUCHI, Hideitsu HINO and Keisuke KAMEYAMA
    • Journal Title

      Neural Processing Letters

      Volume: - Issue: 3 Pages: 1945-1962

    • DOI

      10.1007/s11063-021-10476-z

    • Related Report
      2020 Research-status Report
    • Peer Reviewed / Open Access
  • [Presentation] 3DCNN と MLP の特徴量交換を用いた口唇動画を併用した音声認識2023

    • Author(s)
      河内信誠,亀山啓輔
    • Organizer
      2023年電子情報通信学会総合大会
    • Related Report
      2022 Annual Research Report
  • [Presentation] 敵対的サンプルの生成法を活用したデータ拡張2023

    • Author(s)
      朴潤花,亀山啓輔
    • Organizer
      2023年電子情報通信学会総合大会
    • Related Report
      2022 Annual Research Report
  • [Presentation] Joint-Conditional Mutual Information based Feature Subset Selection for Remotely Sensed Hyperspectral Image Classification2022

    • Author(s)
      U. A. Md Ehsan ALI and Keisuke KAMEYAMA
    • Organizer
      信学技報 IBISML2022-16
    • Related Report
      2022 Annual Research Report
  • [Presentation] 分類器の学習的な組み合わせによる虹彩と眼周囲情報を用いたマルチモーダル個人認証2021

    • Author(s)
      小川恵太, 亀山啓輔
    • Organizer
      第20回情報科学技術フォーラム(FIT)
    • Related Report
      2021 Research-status Report
  • [Presentation] Teacher Assistant及び中間層を模倣するDistillationによるニューラルネットワークのモデル圧縮2021

    • Author(s)
      森川拓海, 亀山啓輔
    • Organizer
      第20回情報科学技術フォーラム(FIT)
    • Related Report
      2021 Research-status Report
  • [Presentation] 虹彩認証における高次スペクトル特徴量の利用2021

    • Author(s)
      亀山啓輔
    • Organizer
      信学技報, IEICE-MICT2021-35
    • Related Report
      2021 Research-status Report
    • Invited

URL: 

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

Information User Guide FAQ News Terms of Use Attribution of KAKENHI

Powered by NII kakenhi