Spectral Integrated Gradients Improve Feature Attribution
Researchers propose Spectral Integrated Gradients (SIG), a new feature attribution method that addresses limitations in the standard Integrated Gradients (IG) approach. IG is widely used for explaining model predictions but suffers from noisy gradients due to its straight-line integration path. SIG constructs integration paths using singular value decomposition (SVD) of the baseline-to-input difference, activating singular components from largest to smallest. This coarse-to-fine progression introduces global structure before fine-grained details, resulting in cleaner attribution maps with reduced noise. Evaluated on diverse image classification datasets, SIG achieves improved quantitative performance compared to existing path-based methods. The paper is available on arXiv under identifier 2605.19607.
Key facts
- Spectral Integrated Gradients (SIG) is proposed as a new feature attribution method.
- SIG uses singular value decomposition (SVD) to construct integration paths.
- The method follows a coarse-to-fine progression by activating singular components from largest to smallest.
- SIG reduces noise in attribution maps compared to standard Integrated Gradients.
- Evaluated on diverse image classification datasets.
- SIG achieves improved quantitative performance over existing path-based methods.
- The paper is available on arXiv with identifier 2605.19607.
- Standard Integrated Gradients uses a straight-line integration path that accumulates noisy gradients.
Entities
Institutions
- arXiv