通过解释不变性和等变性评估可解释性方法的鲁棒性（英文版）

英文标题：Evaluating the Robustness of Interpretability Methods through Explanation Invariance and Equivariance中文摘要：通过几何深度学习的形式化方法，本文研究了神经网络的对称群不变性及其对解释性方法的影响，提出了对称性相关的鲁棒性指标和提高对称性相关解释的系统方法，并通过实验给出了 5 个可行的指南以产生稳健的解释。英文摘要：Interpretability methods are valuable only if their explanations faithfullydescribe the explained model. In this work, we consider neural networks whosepredictions are invariant under a specific symmetry group. This includespopular architectures, ranging from convolutional to graph n