论文标题:
CIGMA: Causal Information-Gain Mechanistic Attribution of Attention Heads in Vision Transformers
发表日期:
未找到日期
发表单位:
University of Oklahoma
原文链接:
https://openaccess.thecvf.com/content/CVPR2026/papers/Maliha_CIGMA_Causal_Information-Gain_Mechanistic_Attribution_of_Attention_Heads_in_Vision_CVPR_2026_paper.pdf
开源代码链接:
原文提到“our code is available here”,但当前材料未给出具体链接
引言
这篇论文盯上的不是“模型准不准”,而是一个更扎心的问题——模型到底是不是在看该看的东西。很多ViT和大视觉语言模型在分类时,表面上答对了,实际却可能偷偷靠背景“抄答案”。比如把一只鸟认成某类,不一定是因为鸟本身,而是因为树枝、天空、草地这些背景太像“标准答案”了。一旦场景换了,背景没了,模型就容易当场“翻车”。很多剪枝方法只关心算力和准确率,像是把一个学生的错题本撕了几页,成绩可能没掉太多,但到底错在哪儿还是不知道。CIGMA 就是来补这个洞的:它不只问“能不能剪”,还问“该剪哪几个头,才能把背景捷径连根拔起”。图1:背景影响比例对比。CIGMA 的 BIR 最低,说明它最少依赖背景线索;别的方法虽然也在做剪枝,但很多时候只是“少算一点”,并没有真正把背景捷径处理掉。这篇工作最有意思的地方在于,它把“解释模型”这件事从看热力图,往前推到因果层面。不是只看某个注意力头“好像在看背景”,而是直接把它切掉,再看模型行为到底变没变。要是变了,而且背景依赖明显下降,那这个头大概率就是“坏头”;要是没啥变化,那它可能没那么重要。
Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016.Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE International Conference on Computer Vision, pages 3429–3437, 2017.Ruth Fong, Mandela Patrick, and Andrea Vedaldi. Understanding deep networks via extremal perturbations and smooth masks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019.Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5797–5808, Florence, Italy, 2019.Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. DynamicViT: Efficient vision transformers with dynamic token sparsification. Advances in Neural Information Processing Systems, 34:13937–13949, 2021.Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. InternVL: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024.Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296–26306, 2024.