Timezone: »
Transformer-based models are receiving increasingly popularity in the field of computer vision, however, the corresponding interpretability is limited. As the simplest explainability method, visualization of attention weights exerts poor performance because of lacking association between the input and model decisions. In this study, we propose a method to generate the saliency map concerning a specific target category. The proposed approach connects the idea of the Markov chain, in order to investigate the information flow across layers of the Transformer and combine the integrated gradients to compute the relevance of input tokens for the model decisions. We compare with other explainability methods using Vision Transformer as a benchmark and demonstrate that our method achieves better performance in various aspects. Our code is available in the anonymized repository: https://anonymous.4open.science/r/TransitionAttentionMaps-8C62.
Author Information
Tingyi Yuan (Xi'an Jiaotong University)
Xuhong Li (Baidu)
Haoyi Xiong (Baidu)
Dejing Dou (Baidu)
More from the Same Authors
-
2022 Poster: Generative Time Series Forecasting with Diffusion, Denoise, and Disentanglement »
Yan Li · Xinjiang Lu · Yaqing Wang · Dejing Dou -
2022 Poster: InterpretDL: Explaining Deep Models in PaddlePaddle »
Xuhong Li · Haoyi Xiong · Xingjian Li · Xuanyu Wu · Zeyu Chen · Dejing Dou -
2022 : SMILE: Sample-to-feature MIxup for Efficient Transfer LEarning »
Xingjian Li · Haoyi Xiong · Cheng-Zhong Xu · Dejing Dou -
2022 : A Simple Framework for Active Learning to Rank »
Qingzhong Wang · Haifang Li · Haoyi Xiong · Wen Wang · Jiang Bian · Yu Lu · Shuaiqiang Wang · zhicong cheng · Dawei Yin · Dejing Dou -
2022 : A Comparative Survey of Deep Active Learning »
Xueying Zhan · Qingzhong Wang · Kuan-Hao Huang · Haoyi Xiong · Dejing Dou · Antoni Chan -
2023 Poster: $\mathcal{M}^4$: A Unified XAI Benchmark for Faithfulness Evaluation of Feature Attribution Methods across Metrics, Modalities and Models »
Xuhong Li · Mengnan Du · Jiamin Chen · Yekun Chai · Himabindu Lakkaraju · Haoyi Xiong -
2022 Spotlight: InterpretDL: Explaining Deep Models in PaddlePaddle »
Xuhong Li · Haoyi Xiong · Xingjian Li · Xuanyu Wu · Zeyu Chen · Dejing Dou -
2022 Spotlight: Lightning Talks 1A-1 »
Siba Smarak Panigrahi · Xuhong Li · Mikhail Usvyatsov · Shaohan Chen · Sohan Patnaik · Haoyi Xiong · Nikolaos V Sahinidis · Rafael Ballester-Ripoll · Chuanhou Gao · Xingjian Li · Konrad Schindler · Xuanyu Wu · Zeyu Chen · Dejing Dou -
2022 Poster: AutoMS: Automatic Model Selection for Novelty Detection with Error Rate Control »
Yifan Zhang · Haiyan Jiang · Haojie Ren · Changliang Zou · Dejing Dou -
2020 Poster: Discriminative Sounding Objects Localization via Self-supervised Audiovisual Matching »
Di Hu · Rui Qian · Minyue Jiang · Xiao Tan · Shilei Wen · Errui Ding · Weiyao Lin · Dejing Dou