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

Theory and Application of deep learning models through the lens of computational algebraic geometry

Research Project

Project/Area Number 20K23341
Research Category

Grant-in-Aid for Research Activity Start-up

Allocation TypeMulti-year Fund
Review Section 1002:Human informatics, applied informatics and related fields
Research InstitutionChiba University (2021)
The University of Tokyo (2020)

Principal Investigator

Kera Hiroshi  千葉大学, 大学院工学研究院, 助教 (00887705)

Project Period (FY) 2020-09-11 – 2022-03-31
Project Status Completed (Fiscal Year 2021)
Budget Amount *help
¥2,860,000 (Direct Cost: ¥2,200,000、Indirect Cost: ¥660,000)
Fiscal Year 2021: ¥1,430,000 (Direct Cost: ¥1,100,000、Indirect Cost: ¥330,000)
Fiscal Year 2020: ¥1,430,000 (Direct Cost: ¥1,100,000、Indirect Cost: ¥330,000)
Keywords深層学習 / 敵対的攻撃 / 敵対的転移性 / ドメイン適応 / 計算代数幾何 / 消失イデアル / 近似計算代数幾何
Outline of Research at the Start

Deep Neural Networksに代表される深層学習モデルは現代の新たな学習モデルであり,様々な応用で目覚ましい成果を上げている.このモデルは入力に対し繰り返し線形演算と非線形演算を適用する形で構成され非常に表現能力が高い.しかし,なぜ,そしてどのくらい表現能力が高いのか,この複雑なモデルをなぜ効率よく最適化できるのかなど,様々な理論的性質が未だ明らかになっていない.本研究では,計算代数幾何という新たな視点でこれらの解明に取り組む.また理論面のみならず,近年発展した消失イデアルの近似基底計算に立脚した新たな深層学習モデルを設計し,精度とデータの性質に関する関係を定式化する.

Outline of Final Research Achievements

Deep learning models are known for their high expressive power. However, their behavior can be significantly altered by small perturbations in the input, which poses severe concerns in reliability. We have achieved two main results that mitigate and handle such small malicious perturbations. First, we showed the existence of architecture of deep learning models that is robust against malicious perturbations that adversely affect other models. We also showed a new application of the malicious perturbations in the domain adaptation of object recognition. Both papers have been accepted by an international journal, IEEE Access.

Academic Significance and Societal Importance of the Research Achievements

本研究では,敵対的攻撃から深層学習モデルを守るメカニズムをモデル構造という新たな観点から分析し,また敵対的攻撃をドメイン適応タスクでの精度向上へと繋げる新たな応用を示した.前者は学術的には深層学習で学習される関数の特性に関わり,統計的・幾何的理解が求められており,また産業的にも人工知能システムの信頼性に深く関わる問題である.後者は敵対的攻撃の手法を頑健性向上でなく精度向上へ活用している.従来は頑健性と精度にはある種のトレードオフが存在していたが,ドメイン適応というタスクでこれを回避できた点が興味深い.これらの研究を通して,深層学習の関数特性に関する基礎的・応用的貢献が行えたと考える.

Report

(3 results)
  • 2021 Annual Research Report   Final Research Report ( PDF )
  • 2020 Research-status Report
  • Research Products

    (4 results)

All 2022 2021 Other

All Int'l Joint Research (1 results) Journal Article (1 results) (of which Int'l Joint Research: 1 results,  Peer Reviewed: 1 results,  Open Access: 1 results) Presentation (2 results)

  • [Int'l Joint Research] Chulalongkorn University(タイ)

    • Related Report
      2021 Annual Research Report
  • [Journal Article] Evolving Architectures With Gradient Misalignment Toward Low Adversarial Transferability2021

    • Author(s)
      Operiano Kevin Richard G.、Pora Wanchalerm、Iba Hitoshi、Kera Hiroshi
    • Journal Title

      IEEE Access

      Volume: 9 Pages: 164379-164393

    • DOI

      10.1109/access.2021.3134840

    • Related Report
      2021 Annual Research Report
    • Peer Reviewed / Open Access / Int'l Joint Research
  • [Presentation] 敵対的訓練を用いたドメイン不変な特徴抽出2022

    • Author(s)
      藤井 一磨,計良 宥志,川本一彦
    • Organizer
      情報処理学会研究報告
    • Related Report
      2021 Annual Research Report
  • [Presentation] Reducing Transferability using Neuroevolution with Gradient Misalignment2021

    • Author(s)
      Kevin Richard Operiano, Wanchalerm Pora, 伊庭斉志,計良宥志
    • Organizer
      進化計算シンポジウム2021
    • Related Report
      2021 Annual Research Report

URL: 

Published: 2020-09-29   Modified: 2023-01-30  

Information User Guide FAQ News Terms of Use Attribution of KAKENHI

Powered by NII kakenhi