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

2021 Fiscal Year Annual Research Report

Intensifying deep learning theory and its application to structure analysis of deep neural network

Research Project

Project/Area Number 18H03201
Research InstitutionThe University of Tokyo

Principal Investigator

鈴木 大慈  東京大学, 大学院情報理工学系研究科, 准教授 (60551372)

Project Period (FY) 2018-04-01 – 2022-03-31
Keywords深層学習 / 機械学習 / 確率的最適化 / カーネル法 / 汎化誤差解析
Outline of Annual Research Achievements

今年度は,深層学習の原理に関して統計理論と最適化の観点から以下のような研究を進めた.
(1) 高次元および無限次元入力深層ニューラルネットワークの統計理論:現実のニューラルネットワークは画像や音声などの高次元・無限次元入力を扱う.そのような状況でも汎化する機構を明らかにするために,真の関数が方向に依存した滑らかさを持つことを仮定し,その仮定の下で深層学習が次元の呪いを回避できることを示した.
(2) 平均場設定でのニューラルネットワークの最適化法:横幅の広いニューラルネットワークのいわゆる平均場設定において,理論保証有りで大域的最適解に収束する新しい手法を提案した.手法のオリジナルな発想として双対平均加法に基づく手法から始め,それを発展する形で双対確率的座標降下法に基づく方法を提案した.この手法は指数オーダーでの収束を達成するものである.
(3) 最適化を含めた良性過学習の解析:ニューラルネットワークはデータに完全にフィットしても依然として良い予測性能を示す良性過学習と言われる現象を引き起こす.この良性過学習による予測誤差を複数の最適化法の間で比較し,いかなる状況でどの最適化法が有利であるかを特徴づけた.
(4) ネットワークの学習ダイナミクスとスパース性:真が横幅の狭い二層ニューラルネットワークであるなら,適切な正則化のもと勾配法を用いたニューラルネットワークの最適化はスパース性の効果が働いて真のニューラルネットワークのパラメータに収束してゆくことを示した.

Research Progress Status

令和3年度が最終年度であるため、記入しない。

Strategy for Future Research Activity

令和3年度が最終年度であるため、記入しない。

  • Research Products

    (36 results)

All 2022 2021 Other

All Int'l Joint Research (1 results) Journal Article (20 results) (of which Int'l Joint Research: 6 results,  Peer Reviewed: 20 results,  Open Access: 20 results) Presentation (15 results) (of which Int'l Joint Research: 4 results,  Invited: 6 results)

  • [Int'l Joint Research] Vector Institute/University of Toronto(カナダ)

    • Country Name
      CANADA
    • Counterpart Institution
      Vector Institute/University of Toronto
  • [Journal Article] AutoLL: Automatic Linear Layout of Graphs based on Deep Neural Network2022

    • Author(s)
      Watanabe Chihiro、Suzuki Taiji
    • Journal Title

      IEEE Symposium Series on Computational Intelligence (SSCI 2021)

      Volume: - Pages: -

    • DOI

      10.1109/SSCI50451.2021.9659893

    • Peer Reviewed / Open Access
  • [Journal Article] Learnability of convolutional neural networks for infinite dimensional input via mixed and anisotropic smoothness2022

    • Author(s)
      Sho Okumoto and Taiji Suzuki
    • Journal Title

      ICLR2022

      Volume: 10 Pages: -

    • Peer Reviewed / Open Access
  • [Journal Article] Particle Stochastic Dual Coordinate Ascent: Exponential convergent algorithm for mean field neural network optimization2022

    • Author(s)
      Kazusato Oko, Taiji Suzuki, Atsushi Nitanda, and Denny Wu
    • Journal Title

      ICLR2022

      Volume: 10 Pages: -

    • Peer Reviewed / Open Access / Int'l Joint Research
  • [Journal Article] Understanding the Variance Collapse of SVGD in High Dimensions2022

    • Author(s)
      Jimmy Ba, Murat A Erdogdu, Marzyeh Ghassemi, Shengyang Sun, Taiji Suzuki, Denny Wu, and Tianzong Zhang
    • Journal Title

      ICLR2022

      Volume: 10 Pages: -

    • Peer Reviewed / Open Access / Int'l Joint Research
  • [Journal Article] Convex Analysis of the Mean Field Langevin Dynamics2022

    • Author(s)
      Atsushi Nitanda, Denny Wu, Taiji Suzuki
    • Journal Title

      AISTATS2022, Proceedings of Machine Learning Research

      Volume: 151 Pages: 9741--9757

    • Peer Reviewed / Open Access / Int'l Joint Research
  • [Journal Article] Gradient Descent in RKHS with Importance Labeling2021

    • Author(s)
      Tomoya Murata, and Taiji Suzuki
    • Journal Title

      AISTATS2021, Proceedings of Machine Learning Research

      Volume: 130 Pages: 1981--1989

    • Peer Reviewed / Open Access
  • [Journal Article] Exponential Convergence Rates of Classification Errors on Learning with SGD and Random Features2021

    • Author(s)
      Shingo Yashima, Atsushi Nitanda, Taiji Suzuki
    • Journal Title

      AISTATS2021, Proceedings of Machine Learning Research

      Volume: 130 Pages: 1954--1962

    • Peer Reviewed / Open Access
  • [Journal Article] Optimal Rates for Averaged Stochastic Gradient Descent under Neural Tangent Kernel Regime2021

    • Author(s)
      Atsushi Nitanda, and Taiji Suzuki
    • Journal Title

      ICLR2021

      Volume: 9 Pages: -

    • Peer Reviewed / Open Access
  • [Journal Article] When Does Preconditioning Help or Hurt Generalization?2021

    • Author(s)
      Shun-ichi Amari, Jimmy Ba, Roger Grosse, Xuechen Li, Atsushi Nitanda, Taiji Suzuki, Denny Wu, Ji Xu
    • Journal Title

      ICLR2021

      Volume: 9 Pages: -

    • Peer Reviewed / Open Access / Int'l Joint Research
  • [Journal Article] Benefit of deep learning with non-convex noisy gradient descent: Provable excess risk bound and superiority to kernel methods2021

    • Author(s)
      Taiji Suzuki, Shunta Akiyama
    • Journal Title

      ICLR2020

      Volume: 9 Pages: -

    • Peer Reviewed / Open Access
  • [Journal Article] Estimation error analysis of deep learning on the regression problem on the variable exponent Besov space2021

    • Author(s)
      Tsuji Kazuma、Suzuki Taiji
    • Journal Title

      Electronic Journal of Statistics

      Volume: 15 Pages: 1869--1908

    • DOI

      10.1214/21-EJS1828

    • Peer Reviewed / Open Access
  • [Journal Article] Sharp characterization of optimal minibatch size for stochastic finite sum convex optimization2021

    • Author(s)
      Nitanda Atsushi、Murata Tomoya、Suzuki Taiji
    • Journal Title

      Knowledge and Information Systems

      Volume: 63 Pages: 2513~2539

    • DOI

      10.1007/s10115-021-01593-1

    • Peer Reviewed / Open Access
  • [Journal Article] Selective inference for latent block models2021

    • Author(s)
      Watanabe Chihiro、Suzuki Taiji
    • Journal Title

      Electronic Journal of Statistics

      Volume: 15 Pages: 3137--3183

    • DOI

      10.1214/21-EJS1853

    • Peer Reviewed / Open Access
  • [Journal Article] Decomposable-Net: Scalable Low-Rank Compression for Neural Networks2021

    • Author(s)
      Yaguchi Atsushi、Suzuki Taiji、Nitta Shuhei、Sakata Yukinobu、Tanizawa Akiyuki
    • Journal Title

      Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence

      Volume: 13 Pages: 3249--3256

    • DOI

      10.24963/ijcai.2021/447

    • Peer Reviewed / Open Access
  • [Journal Article] Bias-Variance Reduced Local SGD for Less Heterogeneous Federated Learning2021

    • Author(s)
      Tomoya Murata, Taiji Suzuki
    • Journal Title

      ICML2021, Proceedings of Machine Learning Research

      Volume: 139 Pages: 7872--7881

    • Peer Reviewed / Open Access
  • [Journal Article] Quantitative Understanding of VAE as a Non-linearly Scaled Isometric Embedding2021

    • Author(s)
      Akira Nakagawa, Keizo Kato, Taiji Suzuki
    • Journal Title

      ICML2021, Proceedings of Machine Learning Research

      Volume: 139 Pages: 7916--7926

    • Peer Reviewed / Open Access
  • [Journal Article] On Learnability via Gradient Method for Two-Layer ReLU Neural Networks in Teacher-Student Setting2021

    • Author(s)
      Shunta Akiyama, Taiji Suzuki
    • Journal Title

      ICML2021, Proceedings of Machine Learning Research

      Volume: 139 Pages: 152--162

    • Peer Reviewed / Open Access
  • [Journal Article] Differentiable Multiple Shooting Layers2021

    • Author(s)
      Stefano Massaroli, Michael Poli, Sho Sonoda, Taiji Suzuki, Jinkyoo Park, Atsushi Yamashita, Hajime Asama
    • Journal Title

      Advances in Neural Information Processing Systems

      Volume: 34 Pages: -

    • Peer Reviewed / Open Access / Int'l Joint Research
  • [Journal Article] Deep learning is adaptive to intrinsic dimensionality of model smoothness in anisotropic Besov space2021

    • Author(s)
      Taiji Suzuki, Atsushi Nitanda
    • Journal Title

      Advances in Neural Information Processing Systems

      Volume: 34 Pages: -

    • Peer Reviewed / Open Access
  • [Journal Article] Particle Dual Averaging: Optimization of Mean Field Neural Networks with Global Convergence Rate Analysis2021

    • Author(s)
      Atsushi Nitanda, Denny Wu, Taiji Suzuki
    • Journal Title

      Advances in Neural Information Processing Systems

      Volume: 34 Pages: -

    • Peer Reviewed / Open Access / Int'l Joint Research
  • [Presentation] 深層学習の数理2022

    • Author(s)
      鈴木大慈
    • Organizer
      日本数学会企画特別講演
    • Invited
  • [Presentation] Benefit of deep learning: Efficiency of function estimation and its optimization guarantee2021

    • Author(s)
      Taiji Suzuki
    • Organizer
      KSIAM2021 (Special Session: CJK-SIAM mini-symposium I: Emerging Mathematics in AI)
    • Int'l Joint Research / Invited
  • [Presentation] 深層学習の理論解析:非線形性と最適化動力学2021

    • Author(s)
      鈴木大慈
    • Organizer
      『非線形動力学に基づく次世代AIと基盤技術』に関するシンポジウム
    • Invited
  • [Presentation] Deep Learning Theory from Statistics to Optimization2021

    • Author(s)
      Taiji Suzuki
    • Organizer
      Tutorial talk, The 6th Asian Conference on Pattern Recognition (ACPR2021)
    • Int'l Joint Research / Invited
  • [Presentation] Deep Learning Theory and Optimization2021

    • Author(s)
      Taiji Suzuki
    • Organizer
      Tutorial talk, The Online Asian Machine Learning School (OAMLS), The 13th Asian Conference on Machine Learning (ACML2021)
    • Int'l Joint Research / Invited
  • [Presentation] Optimality and superiority of deep learning for estimating functions in variants of Besov spaces2021

    • Author(s)
      Taiji Suzuki, Atsushi Nitanda, and Kazuma Tsuji
    • Organizer
      4th International Conference on Econometrics and Statistics (EcoSta2021)
    • Int'l Joint Research
  • [Presentation] ResNetのモデル圧縮手法の提案および圧縮誤差理論解析2021

    • Author(s)
      平川 雅人,鈴木 大慈
    • Organizer
      2021年度統計関連学会連合大会
  • [Presentation] 教師生徒設定における勾配法による二層ReLU ニューラルネットワークの学習可能性について2021

    • Author(s)
      秋山 俊太,鈴木 大慈
    • Organizer
      2021年度統計関連学会連合大会
  • [Presentation] 平均場ニューラルネットワークの効率的最適化法2021

    • Author(s)
      二反田 篤史,大古 一聡,Denny Wu,鈴木 大慈
    • Organizer
      2021年度統計関連学会連合大会
  • [Presentation] リンク予測におけるバイアス項によるグラフニューラルネットワークの表現力強化2021

    • Author(s)
      長谷川 貴大,鈴木 大慈
    • Organizer
      2021年度統計関連学会連合大会
  • [Presentation] Particle Stochastic Dual Coordinate Ascent: Exponential convergent algorithm for mean field neural network optimization2021

    • Author(s)
      大古 一聡, 鈴木 大慈, 二反田 篤史, Wenny Wu
    • Organizer
      第24回情報論的学習理論ワークショップ
  • [Presentation] ノイズ付き勾配法を用いた教師生徒設定における二層ReLuニューラルネットワークの学習2021

    • Author(s)
      秋山俊太, 鈴木大慈
    • Organizer
      第24回情報論的学習理論ワークショップ
  • [Presentation] 多層ニューラルネットワークモデルに基づくmatrix reordering2021

    • Author(s)
      渡邊千紘, 鈴木大慈
    • Organizer
      第24回情報論的学習理論ワークショップ
  • [Presentation] CGから実写への転移学習におけるスケーリング則2021

    • Author(s)
      Hiroaki Mikami, Kenji Fukumizu, Shogo Murai, Shuji Suzuki, Yuta Kikuchi, Taiji Suzuki, Shin-ichi Maeda, Kohei Hayashi
    • Organizer
      第24回情報論的学習理論ワークショップ
  • [Presentation] 深層ニューラルネットワークの近似理論と適応能力2021

    • Author(s)
      鈴木大慈
    • Organizer
      数値解析セミナー
    • Invited

URL: 

Published: 2022-12-28  

Information User Guide FAQ News Terms of Use Attribution of KAKENHI

Powered by NII kakenhi