This paper presents VeriX, a tool for generating optimal robust explanations and counterfactuals along decision boundaries of deep neural networks.

Specifically, its perturbations are bounded.

The pseudocode is shown below.

Overall, VeriX iterates through each feature in $\mathbf{x}$, checking whether the feature satisfies the following condition: a bounded perturbation on $\mathbf{B} \cup \{i\}$ dose not alter the model’s prediction. If the condition is met, the feature is considered unimportant; otherwise, it is deemed important. In my view, while VeriX successfully identifies important feature, it does not appear to provide an importance score for each feature.

References

Wu, M., Wu, H., & Barrett, C. VeriX: Towards Verified Explainability of Deep Neural Networks. In Thirty-seventh Conference on Neural Information Processing Systems.