ScholarMate
English简体中文繁體中文
客服热线:400-1616-289
登录注册
会议论文

Rethinking Attention-Model Explainability through Faithfulness Violation Test

Liu, Yibing; Li, Haoliang*; Guo, Yangyang; Kong, Chenqi; Li, Jing; Wang, Shiqihong kong polytech univ
38th International Conference on Machine Learning (ICML)2022
3阅读
引用收藏
分享

摘要

Attention mechanisms are dominating the explainability of deep models. They produce probability distributions over the input, which are widely deemed as feature-importance indicators. However, in this paper, we find one critical limitation in attention explanations: weakness in identifying the polarity of feature impact. This would be somehow misleading - features with higher attention weights may not faithfully contribute to model predictions; instead, they can impose suppression effects. With this finding, we reflect on the explainability of current attention-based techniques, such as Attention. Gradient and LRPbased attention explanations. We first propose an actionable diagnostic methodology (henceforth faithfulness violation test) to measure the consistency between explanation weights and the impact polarity. Through the extensive experiments, we then show that most tested explanation methods are unexpectedly hindered by the faithfulness violation issue, especially the raw attention. Empirical analyses on the factors affecting violation issues further provide useful observations for adopting explanation methods in attention models.

相关
加载中…
  • A review and benchmark of feature importance methods for neural networks
    Mandler, Hannes*; Weigand, BernhardACM Computing Surveys, 2024, 56(12): 318.
  • Differentiated Explanation of Deep Neural Networks With Skewed Distributions
    Fu, Weijie; Wang, Meng*; Du, Mengnan; Liu, Ninghao; Hao, Shijie; Hu, XiaIEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(6): 2909-2922.被引 1
  • AcME-Accelerated model-agnostic explanations: Fast whitening of the machine-learning black box
    Dandolo, David; Masiero, Chiara; Carletti, Mattia; Pezze, Davide Dalle; Susto, Gian Antonio*EXPERT SYSTEMS WITH APPLICATIONS, 2023, 214: 119115.被引 11
查看全部 →

会议信息

会议名称
38th International Conference on Machine Learning (ICML)
会议日期
2022-07-17 至 2022-07-23
论文集名
INTERNATIONAL CONFERENCE ON MACHINE LEARNING, VOL 162
发表日期
2022
资助标注
National Natural Science Foundation of China [62022002]; Hong Kong Research Grants Council General Research Fund (GRF) [11203220]; Hong Kong Innovation and Technology Commission (InnoHK Project CIMDA); CityU Teaching Start-up Grant [6000801]; CityU Strategic Interdisciplinary Research Grant (SIRG) [7020055]; CityU New Research Initiatives/Infrastructure Support from Central [APRC 9610528]
收录
ISTP

作者

  • L

    Liu, Yibing (第一作者)

  • L

    Li, Haoliang (通讯)

查看全部 6 位作者收起作者
  • G

    Guo, Yangyang

  • K

    Kong, Chenqi

  • L

    Li, Jing

    hong kong polytech univ

  • W

    Wang, Shiqi

登录解锁更多功能

  • 收藏这篇,随时在“我的 · 收藏”找回
  • 认领你的成果,生成推广素材
  • 关注作者获取最新成果
  • 加入项目组、协同写作
注册 / 登录

产品服务

  • 科研之友
  • 创新城
  • 科创云

服务支持

  • 帮助中心
  • 隐私政策
  • 服务条款

联系方式

在线客服:【】
客服热线:400-1616-289
电子邮箱:support@scholarmate.com

关注或下载科研之友

微信二维码
微信公众号
客户端下载二维码
下载客户端
科研成果科研人员科研机构科研动态爱瑞思软件

©2026 深圳市科研之友网络服务有限公司

公安备案图标粤公网安备 44030502000213
粤ICP备 16046710 号粤B2-20110417