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

A Unified Approach for Explaining Deep Neural Networks

Research Project

Project/Area Number 18K18106
Research Category

Grant-in-Aid for Early-Career Scientists

Allocation TypeMulti-year Fund
Review Section Basic Section 61030:Intelligent informatics-related
Research InstitutionOsaka University

Principal Investigator

Hara Satoshi  大阪大学, 産業科学研究所, 准教授 (40780721)

Project Period (FY) 2018-04-01 – 2021-03-31
Project Status Completed (Fiscal Year 2020)
Budget Amount *help
¥4,160,000 (Direct Cost: ¥3,200,000、Indirect Cost: ¥960,000)
Fiscal Year 2019: ¥1,820,000 (Direct Cost: ¥1,400,000、Indirect Cost: ¥420,000)
Fiscal Year 2018: ¥2,340,000 (Direct Cost: ¥1,800,000、Indirect Cost: ¥540,000)
Keywords機械学習 / 深層学習 / 説明可能AI / 解釈性 / 人工知能
Outline of Final Research Achievements

Deep neural network models are inherently complex, which hinder us from inferring the underlying mechanisms or the evidences that the models rely on when making decisions. It is therefore essential to develop "explanation methods" that can reveal such mechanism or evidence so that we can understand the decisions of the models. In this research, we focused on a unification of the popular explanation methods, the explanation by important features and the explanation by similar/relevant instances. Through the research, we deepen and improved the methodologies for each explanations individually, and we then developed a unification framework that can taken into account the advantages of the both of the explanations.

Academic Significance and Societal Importance of the Research Achievements

深層学習モデルは高い予測・認識精度を誇る一方で、一般にとても複雑な構造をしており、モデルの判断根拠をユーザが窺い知ることは困難である。このため、深層学習モデルは一般に”ブラックボックス”とされる。”ブラックボックス”性のために深層学習モデルをそのまま人間の重要な意思決定の補助(e.g ローン審査や医療診断など)に用いることは困難である。本研究で開発した説明法はこのような深層学習モデルの”ブラックボックス”性を緩和することができる。これにより、ユーザは高精度な深層学習モデルを、その判断根拠を窺いながら意思決定補助に用いることができるようになる。

Report

(4 results)
  • 2020 Annual Research Report   Final Research Report ( PDF )
  • 2019 Research-status Report
  • 2018 Research-status Report
  • Research Products

    (13 results)

All 2020 2019 2018 Other

All Int'l Joint Research (5 results) Presentation (6 results) (of which Int'l Joint Research: 6 results) Remarks (2 results)

  • [Int'l Joint Research] Universite du Quebec a Montreal(カナダ)

    • Related Report
      2020 Annual Research Report
  • [Int'l Joint Research] University of Iowa(米国)

    • Related Report
      2020 Annual Research Report
  • [Int'l Joint Research] Universite du Quebec a Montreal(カナダ)

    • Related Report
      2019 Research-status Report
  • [Int'l Joint Research] University of Iowa(米国)

    • Related Report
      2019 Research-status Report
  • [Int'l Joint Research] Universite du Quebec a Montreal(カナダ)

    • Related Report
      2018 Research-status Report
  • [Presentation] Interpretable Companions for Black-Box Models2020

    • Author(s)
      Danqing Pan, Tong Wang, Satoshi Hara
    • Organizer
      The 23rd International Conference on Artificial Intelligence and Statistics
    • Related Report
      2020 Annual Research Report
    • Int'l Joint Research
  • [Presentation] Faking Fairness via Stealthily Biased Sampling.2020

    • Author(s)
      Kazuto Fukuchi, Satoshi Hara, Takanori Maehara
    • Organizer
      34th AAAI Conference on Artificial Intelligence
    • Related Report
      2019 Research-status Report
    • Int'l Joint Research
  • [Presentation] Fairwashing: the risk of rationalization2019

    • Author(s)
      Ulrich Aivodji, Hiromi Arai, Olivier Fortineau, Sebastien Gambs, Satoshi Hara, Alain Tapp
    • Organizer
      36th International Conference on Machine Learning
    • Related Report
      2019 Research-status Report
    • Int'l Joint Research
  • [Presentation] Convex Hull Approximation of Nearly Optimal Lasso Solutions2019

    • Author(s)
      Satoshi Hara, Takanori Maehara
    • Organizer
      16th Pacific Rim International Conference on Artificial Intelligence
    • Related Report
      2019 Research-status Report
    • Int'l Joint Research
  • [Presentation] Data Cleansing for Models Trained with SGD2019

    • Author(s)
      Satoshi Hara, Atsuhi Nitanda, Takanori Maehara
    • Organizer
      Neural Information Processing Systems
    • Related Report
      2019 Research-status Report
    • Int'l Joint Research
  • [Presentation] Maximally Invariant Data Perturbation as Explanation2018

    • Author(s)
      Satoshi Hara, Kouichi Ikeno, Tasuku Soma, Takanori Maehara
    • Organizer
      2018 ICML Workshop on Human Interpretability in Machine Learning
    • Related Report
      2018 Research-status Report
    • Int'l Joint Research
  • [Remarks] Interpretable Companions for Black-Box Models

    • URL

      https://arxiv.org/abs/2002.03494

    • Related Report
      2019 Research-status Report
  • [Remarks] Fairwashing: the risk of rationalization

    • URL

      https://arxiv.org/abs/1901.09749

    • Related Report
      2018 Research-status Report

URL: 

Published: 2018-04-23   Modified: 2022-01-27  

Information User Guide FAQ News Terms of Use Attribution of KAKENHI

Powered by NII kakenhi