One Wave To Explain Them All: A Unifying Perspective On Feature Attribution

Accepted at ICML 2025

Gabriel Kasmi 1,2, Amandine Brunetto 1, Thomas Fel3, Jayneel Parekh4
1Mines Paris - PSL University, 2RTE France, 3Kempner Institute, Harvard University, 4ISIR, Sorbonne Université,
ArXiv Code

Abstract

A major limitation of deep neural networks in safety-critical contexts is their lack of transparency. Feature attribution methods address this by identifying which features influence a model’s decisions. For high-dimensional inputs like images, attributions often appear as pixel-based heatmaps. However, extending pixel-based methods to non-image domains, leads to suboptimal explanations. This paper introduces the Wavelet Attribution Method (WAM), which replaces pixels with wavelet coefficients as features for attribution. Wavelet coefficients, defined for square-integrable signals across dimensions, provide a mathematically grounded and semantically richer domain. As the wavelet domain is defined for any input dimension, it provides a unification of the feature space for attribution. We demonstrate WAM's effectiveness on audio, images, and volumes, where it quantitatively matches or outperforms existing gradient-based methods. Explanations generated by WAM leverage the spatial and scale-localized properties of wavelet coefficients, capturing both the where and what of a model's focus. We illustrate this by showing how WAM relates to model robustness, enables deeper volume interpretation, and identifies key audio components in noisy samples.

WAM overview

Method Overview

WAM explains any input modality by decomposing the model’s decision in the wavelet domain. It computes the gradient of the model’s prediction with respect to the wavelet coefficients of the input modality (audio, images, volumes). Unlike pixels, wavelet coefficients preserve structural information about the input signal, offering deeper insights into the model’s behavior and going beyond where it focuses.

Explanation of different modalities

BibTeX

@inproceedings{
        kasmi2025WAM,
        title={One Wave To Explain Them All: A Unifying Perspective On Feature Attribution},
        author={Kasmi, Gabriel and Brunetto, Amandine and Fel, Thomas and Parekh, Jayneel},
        booktitle={Forty-second International Conference on Machine Learning},
        year={2025},
        }