jin's blog

  • 홈
  • 태그
  • 미디어로그
  • 위치로그
  • 방명록

Quantifying Attention Flow in Transformers 1

[Transformer] Quantifying Attention Flow in Transformers

Input Token의 상대적인 relevance에 따라 attention weight을 이용할때, attention weight, attention롤아웃 및 흐름을 고려해서 Input Token에 대한 attention 을 근사하는 2가지 방법을 제안하는 논문. 해당 방법은 정보의 흐름에 대한 보완적인 view를 제공하였음. 또한 오리지널 attention과 비교했을 때, ablation method 및 input gradient (두 방법 모두) 를 이용하여 얻은 Input Token의 importance score와 높은 상관관계를 생성했음을 보여주었음 $y=a_x$ 단수/복수 예측하는 verb number 예측 해당 task와 데이터셋은 arxiv.org/pdf/2005.00928.pdf

IT/Paper 2021.03.13
이전
1
다음
더보기
프로필사진

jin's blog

Endure

  • 분류 전체보기 (50)
    • IT (44)
      • Paper (23)
      • Reinforcement Learni.. (0)
      • Probability (0)
      • Deep learning (6)
      • Spark (5)
      • Python (4)
      • Computer vision (4)
      • Data Structure (1)
    • 관심사 (2)
      • 낚시 (0)
      • 피아노 (2)
      • 일상 (0)

Tag

Fast R-CNN, They Are Features, TCAV, RL논문, smoothGrad, Never Give Up, Regularizing Trajectory Optimization with Denoising Autoencoders, Learning Directed Exploration Strategies, Interpretability Beyond Feature Attribution:Quantitative Testing with Concept Activation Vectors, R-CNN, vision transformer, Axiomatic Attribution for Deep Networks, intergrated gradient, Concept vector, XAI, Adversarial Examples Are Not Bugs, Paper리뷰, Quantifying Attention Flow in Transformers, CAV, Deconvolution Network,

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