Wavelet Attribution Method

WAM visualizations

The figure below displays additional visualizations of WAM. From left to right: original image, WAM in the wavelet domain, heatmap in the pixel domain, filtered image illustrating the image regions that need to be well defined for the model to predict the image label correctly.

Minimal images

The figure below displays examples of minimal images obtained after meaningfully perturbing the input image in the wavelet domain.

Explanations on audio spectrograms

We refer the reader to this page for a tutorial on audio spectrograms.

Noise Experiment. We add 0~dB white noise on the audio of the target class ('Crow') to form the input to the classifier. Relevant parts are highlighted with the green box.

Additional visualizations

WAM visualizations

Minimal images

Explanations on audio spectrograms

Noise Experiment. We add 0~dB white noise on the audio of the target class ('Crow') to form the input to the classifier. Relevant parts are highlighted with the green box.

Overlap Experiment. Audio of 'Dog' mixed with 'Rooster'. Relevant parts are highlighted.

Overlap Experiment. Audio of 'Dog' mixed with 'Cat'. Relevant parts are highlighted.

Explanations on 3D shapes

3DMNIST.

3DMNIST.

3DMNIST.

VesselMNIST3D.

VesselMNIST3D.

VesselMNIST3D.