Assessment of the Reliablity of a Model's Decision by Generalizing Attribution to the Wavelet Domain

Published: 27 Oct 2023, Last Modified: 09 Nov 2023NeurIPS XAIA 2023EveryoneRevisionsBibTeX
TL;DR: A method that generalizes attribution to the wavelet domain to assess the reliability of a prediction
Abstract: Neural networks have shown remarkable performance in computer vision, but their deployment in numerous scientific and technical fields is challenging due to their black-box nature. Scientists and practitioners need to evaluate the reliability of a decision, i.e., to know simultaneously if a model relies on the relevant features and whether these features are robust to image corruptions. Existing attribution methods aim to provide human-understandable explanations by highlighting important regions in the image domain, but fail to fully characterize a decision process's reliability. To bridge this gap, we introduce the Wavelet sCale Attribution Method (WCAM), a generalization of attribution from the pixel domain to the space-scale domain using wavelet transforms. Attribution in the wavelet domain reveals where and on what scales the model focuses, thus enabling us to assess whether a decision is reliable. Our code is accessible here: \url{https://github.com/gabrielkasmi/spectral-attribution}.
Submission Track: Full Paper Track
Application Domain: Computer Vision
Survey Question 1: To safely deploy deep learning models in real-world vision applications, tools to assess the reliability of their prediction are necessary. Reliability means that the prediction relies on relevant features and that these features are robust to alterations (e.g., image corruptions or distribution shifts). We introduce a method that enables such an assessment by decomposing a model's decision into scales, which are interpretable by design and correspond to frequencies whose robustness is quantifiable.
Survey Question 2: To assess the reliability of a prediction, existing attribution methods (e.g., class activation mapping, saliency maps) indicate where the model sees on the image, but not what it sees. Therefore, these methods provide a partial assessment of the reliability of the prediction. Our work introduces a method that aims at bridging this gap.
Survey Question 3: We benchmark our method with several explainability methods (Saliency, GradCAM, VarGrad, Rise, Sobol attribution method) to show that we retrieve similar results when focusing on assessing where the model sees on the input image.
Submission Number: 10
Loading