HyunMaru Self_Study

  • 홈
  • 태그
  • 방명록

출력값에 대한 지식증류 1

Decoupled Knowledge Distillation 논문 리뷰

논문 제목 : Decoupled Knowledge Distillation Contribution 기존 KD를 TCKD(Target Class Knowledge Distillation), NCKD(Non-Target Class Knowledge Distllation) 로 나누어 logit distllation의 insight 제공 한쌍의 공식으로 되어있는 기존 KD Loss의 한계점을 드러냄. Teacher의 Confidence Score와의 Coupling NCKD로 지식 전이의 효과가 억압됨. 그리고 TCKD와 NCKD의 결합은 두 부분의 균형 유연성을 제한시킴 효과적인 logit Distillation DKD를 제안 Feature-based Distillation 방법과 비교해서 DKD의 featur..

논문 리뷰 2024.02.23
이전
1
다음
더보기
프로필사진

연세대학교 일반대학원 석사과정 Yonsei Univ. AISLAB

  • 분류 전체보기 (28)
    • 밑바닥부터 시작하는 딥러닝 (2)
    • 앱개발 with Flutter (4)
    • 논문 리뷰 (7)
    • 백준에 대가리 깨지는중 (15)
    • Network Coding With Python (0)

Tag

Distillation from logits, 지식증류, Knowledge Distillation, 2021 CVPR, 백준, Class-aware Information for Logit-based Knowledge Distillation, 경량화, Spectral Convolution, Decoupled Knowledge Distillation, Logit-based Distillation, Laplacian Matrix, 실버1, Graph Fourier Transform, 지식 증류, 가장 가까운 세 사람의 심리적 거리, Chebyshev Polynomials, 출력값에 대한 지식증류, 딥러닝, Spatial Convolution, Distilling Knowledge via Knowledge Review,

최근글과 인기글

  • 최근글
  • 인기글

최근댓글

공지사항

페이스북 트위터 플러그인

  • Facebook
  • Twitter

Archives

Calendar

«   2025/05   »
일 월 화 수 목 금 토
1 2 3
4 5 6 7 8 9 10
11 12 13 14 15 16 17
18 19 20 21 22 23 24
25 26 27 28 29 30 31

방문자수Total

  • Today :
  • Yesterday :

Copyright © Kakao Corp. All rights reserved.

티스토리툴바