WaveX: Frequency-Domain Explainability of CNNs
Explaining what a CNN sees one frequency band at a time, in the Fourier and wavelet domains.
- Period
- January–May 2026
- Kind
- research
- Stack
- Integrated Gradients, Grad-CAM, Wavelets, Fourier analysis
- Guide
- Prof. Amit Sethi, IIT Bombay
- Frequency-aware Integrated Gradients in the Fourier and wavelet domains. Fourier-filtered baselines, and subband interpolation with IDWT reconstruction, isolate which spatial-frequency bands carry a class’s attribution.
- Activation-matching masks trained with a Grad-CAM-guided loss, giving compact explanations that preserve the class.
- Extended that mask optimisation to polar-frequency Fourier masks and level-wise wavelet masks.