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

Information Criterion WBIC and its improvement

Research Project

Project/Area Number 21K12025
Research Category

Grant-in-Aid for Scientific Research (C)

Allocation TypeMulti-year Fund
Section一般
Review Section Basic Section 61030:Intelligent informatics-related
Research InstitutionTokyo Institute of Technology

Principal Investigator

Watanabe Sumio  東京工業大学, 情報理工学院, 教授 (80273118)

Project Period (FY) 2021-04-01 – 2024-03-31
Project Status Completed (Fiscal Year 2023)
Budget Amount *help
¥1,820,000 (Direct Cost: ¥1,400,000、Indirect Cost: ¥420,000)
Fiscal Year 2023: ¥520,000 (Direct Cost: ¥400,000、Indirect Cost: ¥120,000)
Fiscal Year 2022: ¥780,000 (Direct Cost: ¥600,000、Indirect Cost: ¥180,000)
Fiscal Year 2021: ¥520,000 (Direct Cost: ¥400,000、Indirect Cost: ¥120,000)
Keywords特異学習理論 / 周辺尤度 / 情報量規準 / 代数幾何 / 自由エネルギー / WBIC / 特異モデル / 実対数閾値 / 調整された交差検証 / 調整された情報量規準 / WAIC / 汎化誤差
Outline of Research at the Start

データサイエンスや機械学習を現実の世界の課題に適用しようとするとき、未知の分布に対して確率モデルや事前分布が適切であるかどうかを確かめるための方法が必要になる。その際にしばしば用いられる規準が汎化損失と自由エネルギーである。本研究では、主として自由エネルギーの計算法を研究し、データが独立でなくても精度よく推定値を求める方法を構成する。汎化損失を推定するための基準WAIC1が世界中で広くい用いられているように、本研究の成果もそのようになることが目標である。

Outline of Final Research Achievements

In statistical inference and statistical learning, the free energy which is equal to the minus log marginal likelihood is defined for a given probabilistic model, a prior distribution, and a sample. In this research project, the following results were obtained. (1) The conventional theory was generalized for exchangeable random variables so that WBIC can be employed to estimate the free energy. (2) The free energies for the convolutional neural networks using a ReLU function with skip connection and without skip connection were clarified. And (3) the free energy of a learning model which has hierarchical structure or hidden variables were clarified when the data-generating distribution is not realizable.

Academic Significance and Societal Importance of the Research Achievements

確率モデル・事前分布・サンプルが与えられたときに自由エネルギーの値を算出することは、データ生成分布に対する確率モデルと事前分布の適切さを判断する際に重要な役割を果たすことが知られているが、事後分布が正規分布で近似できない場合には、その値を求めるには大きな演算量が必要であった。本研究では、自由エネルギーの値を求めるために提案されていた情報量規準WBICについて研究を行い、より一般な条件下で利用可能であるように拡張し、階層モデルを持つモデル出の挙動を解明した。これらの結果は、実問題のデータ分析におけるモデリングの基盤のひとつであり、また人工知能アライメントのための数学的基礎を構成するものである。

Report

(4 results)
  • 2023 Annual Research Report   Final Research Report ( PDF )
  • 2022 Research-status Report
  • 2021 Research-status Report
  • Research Products

    (12 results)

All 2024 2023 2022 2021 Other

All Journal Article (6 results) (of which Peer Reviewed: 6 results,  Open Access: 3 results) Presentation (5 results) (of which Int'l Joint Research: 3 results,  Invited: 3 results) Remarks (1 results)

  • [Journal Article] Free Energy of Bayesian Convolutional Neural Network with Skip Connection2024

    • Author(s)
      S.Nagayasu, S.Watanabe
    • Journal Title

      Proceedings of the 15th Asian Conference on Machine Learning

      Volume: 222 Pages: 927-942

    • Related Report
      2023 Annual Research Report
    • Peer Reviewed / Open Access
  • [Journal Article] Mathematical Theory of Bayesian Statistics for Unknown Information Source2023

    • Author(s)
      Sumio Watanabe
    • Journal Title

      Philosophical Transactions of the Royal Society A

      Volume: - Issue: 2247

    • DOI

      10.1098/rsta.2022.0151

    • Related Report
      2022 Research-status Report
    • Peer Reviewed
  • [Journal Article] Mathematical Theory of Bayesian Statistics where all models are wrong2022

    • Author(s)
      Sumio Watanabe
    • Journal Title

      Handbook of Statistics, Elsevior

      Volume: 47 Pages: 209-238

    • DOI

      10.1016/bs.host.2022.06.001

    • ISBN
      9780323952682
    • Related Report
      2022 Research-status Report
    • Peer Reviewed
  • [Journal Article] Recent Advances in Algebraic Geometry and Bayesian Statistics2022

    • Author(s)
      Sumio Watanabe
    • Journal Title

      Information Geometry

      Volume: - Issue: S1 Pages: 187-209

    • DOI

      10.1007/s41884-022-00083-9

    • Related Report
      2022 Research-status Report
    • Peer Reviewed
  • [Journal Article] Algebraic Geometry and Statistical Theory2021

    • Author(s)
      渡辺澄夫
    • Journal Title

      Bulletin of the Japan Society for Industrial and Applied Mathematics

      Volume: 31 Issue: 3 Pages: 7-14

    • DOI

      10.11540/bjsiam.31.3_7

    • NAID

      130008135911

    • ISSN
      2432-1982
    • Year and Date
      2021-09-22
    • Related Report
      2021 Research-status Report
    • Peer Reviewed / Open Access
  • [Journal Article] Information criteria and cross validation for Bayesian inference in regular and singular cases2021

    • Author(s)
      Sumio Watanabe
    • Journal Title

      Japanese Journal of Statistics and Data Science

      Volume: 4 Issue: 1 Pages: 1-19

    • DOI

      10.1007/s41237-021-00133-z

    • NAID

      210000185598

    • Related Report
      2021 Research-status Report
    • Peer Reviewed / Open Access
  • [Presentation] Singular Learning Theory, Deep Learning, and AI Alignment2024

    • Author(s)
      Sumio Watanabe
    • Organizer
      Tokyo Deep Learning 2024
    • Related Report
      2023 Annual Research Report
    • Int'l Joint Research / Invited
  • [Presentation] Bias and Variance of Bayes Cross Validation in Singular Learning Theory2023

    • Author(s)
      Sumio Watanabe
    • Organizer
      Bayesian Statistics and Statistical Learning
    • Related Report
      2023 Annual Research Report
    • Int'l Joint Research / Invited
  • [Presentation] Singularity Theory in Statistical Science2022

    • Author(s)
      Sumio Watanabe
    • Organizer
      MSJ-SI: Deepening and Evolution of Applied Singularity Theory
    • Related Report
      2022 Research-status Report
    • Int'l Joint Research / Invited
  • [Presentation] 入力分布が低次元超平面上にあるときの縮小ランク回帰の実対数閾値2022

    • Author(s)
      広瀬青、渡辺澄夫
    • Organizer
      信学技報, vol. 121, no. 419, IBISML2021-32,
    • Related Report
      2021 Research-status Report
  • [Presentation] Asymptotic Behavior of Bayesian Generalization Error in Multinomial Mixtures2022

    • Author(s)
      Takumi Watanabe, Sumio Watanabe
    • Organizer
      arxiv.org
    • Related Report
      2021 Research-status Report
  • [Remarks] 渡辺澄夫

    • URL

      http://watanabe-www.math.dis.titech.ac.jp/users/swatanab/index-j.html

    • Related Report
      2022 Research-status Report 2021 Research-status Report

URL: 

Published: 2021-04-28   Modified: 2025-01-30  

Information User Guide FAQ News Terms of Use Attribution of KAKENHI

Powered by NII kakenhi