ViT-CX: Causal Explanation of Vision Transformers

Authors: Weiyan Xie, Xiao-Hui Li, Caleb Chen Cao, Nevin L. Zhang

IJCAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The empirical results show that Vi T-CX produces more meaningful saliency maps and does a better job revealing all important evidence for the predictions than previous methods. The explanation generated by Vi TCX also shows significantly better faithfulness to the model. The codes and appendix are available at https://github.com/vaynexie/Causal X-Vi T. ... Visual examples and experiment results show that Vi T-CX clearly outperforms previous baselines in terms of faithfulness to model and interpretability to human users (Figure 1 and Table 1).
Researcher Affiliation Collaboration Weiyan Xie 1 , Xiao-Hui Li 2 , Caleb Chen Cao 1 and Nevin L. Zhang 1 1 The Hong Kong University of Science and Technology, China 2 Huawei Technologies Co., Ltd, China {wxieai, cao, lzhang}@ust.hk, {lixiaohui33}@huawei.com
Pseudocode No The paper describes the methods in text and uses flowcharts (Figure 3) but does not provide formal pseudocode or algorithm blocks.
Open Source Code Yes The codes and appendix are available at https://github.com/vaynexie/Causal X-Vi T.
Open Datasets Yes We use 5,000 images randomly selected from the ILSVRC2012 validation set [Deng et al., 2009].
Dataset Splits No The paper uses 5,000 images from the ILSVRC2012 validation set for evaluation, but it does not provide explicit training/test/validation dataset splits of its own data for model training/evaluation.
Hardware Specification Yes All experiments are run on an Intel Xeon E5-2620 CPU and an NVIDIA 2080 Ti GPU.
Software Dependencies No The paper mentions "Mind Spore" as a deep learning computing framework, but it does not specify a version number or provide version numbers for any other software dependencies needed to replicate the experiment.
Experiment Setup Yes To generate the masks Mvit, we use feature maps from the last transformer block for Vi T-B and Dei T-B, and choose those from the last block of the second to last stage for Swin-B; When clustering on Mvit to generate the mask set Mcx, the distance threshold δ is set to 0.1 for Vi T-B and Dei T-B, and set to 0.05 for Swin-B; The standard deviation σ of the Gaussian noise ϵi is set to 0.1.