One Wave To Explain Them All: A Unifying Perspective On Feature Attribution
Authors: Gabriel Kasmi, Amandine Brunetto, Thomas Fel, Jayneel Parekh
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the superiority of WAM through extensive empirical evaluations across a diverse set of metrics, model architectures, and datasets, showing its robustness and versatility. We also discuss the novel insights and connections that our method permits, such as connecting feature attribution and the characterization of a model s robustness or filtering the important part of an audio signal without requiring a to train an explanation method. Table 1 presents the performance of WAM. Our method consistently outperforms traditional attribution methods, especially for images and volumes, while achieving competitive results for audio. |
| Researcher Affiliation | Collaboration | 1Mines Paris PSL University, Paris, France 2RTE France, Paris La D efense, France 3Kempner Institute, Harvard University, Cambridge, MA, United-States 4ISIR, Sorbonne Universit e, Paris, France. Correspondence to: Gabriel Kasmi <gabriel.kasmi[at]minesparis.psl.eu>. |
| Pseudocode | Yes | Figure 10. Pseudo-code for adding Gaussian noise to audio. |
| Open Source Code | No | Project page: https: //gabrielkasmi.github.io/wam/. (The URL provided is for a project page, not a direct code repository link, and there is no explicit statement about code release for the methodology described in the paper.) |
| Open Datasets | Yes | To address this gap, we designed a comprehensive evaluation framework spanning diverse datasets: ESC-50 (Piczak, 2015) for audio, Image Net (Russakovsky et al., 2015) for images, and Med MNIST3D (Yang et al., 2023) for volumes. |
| Dataset Splits | Yes | Evaluations are conducted on 400 samples from ESC-50 (fold 1), 1,000 images from Image Net s validation set and the full Adrenal MNIST3D test set (298 samples). For volumes, we considered the LATEC benchmark (Klein et al., 2024) and evaluate our method on two datasets of Med MNIST (Yang et al., 2023): Adrenal MNIST3D and Vessel MNIST. We carry out our evaluation over the complete test set. |
| Hardware Specification | No | No specific hardware details (like GPU/CPU models, processor types, or memory amounts) are mentioned in the paper for running its experiments. |
| Software Dependencies | No | We use the Python library Captum (Kokhlikyan et al., 2020) for consistently implementing existing methods on our datasets for audio and images. All models are retrieved from the Py Torch Image Models (Wightman, 2019) repository. (While software names are mentioned, specific version numbers are not provided for reproducibility.) |
| Experiment Setup | Yes | The number of samples, n, needed to compute the approximation of the smoothed gradient and the standard deviation σ2 are hyperparameters. In practice, we employ the Nadam (Dozat, 2016) optimizer, which combines the benefits of Nesterov acceleration and Adam optimization. Our approach consistently produces masks with controllable sparsity levels up to 90%, meaning that 90% of the wavelet coefficients are zeroed out, while maintaining a classification score comparable to or better than the original prediction. Results are averaged across 1,000 images optimized for 500 steps and for α ranging in [0, 100] for each image. |