An MDL-based wavelet scattering features selection for signal classification

Wavelet scattering is a redundant time-frequency transform that was shown to be a powerful tool in signal classification. It shares the convolutional architecture with convolutional neural networks, but it offers some advantages, including faster training and small training sets. However, it introduces some redundancy along the frequency axis, especially for filters that have a high degree of overlap. This naturally leads to a need for dimensionality reduction to further increase its efficiency as a machine learning tool. In this paper, the Minimum Description Length is used to define an automatic procedure for optimizing the selection of the scattering features, even in the frequency domain. The proposed study is limited to the class of uniform sampling models. Experimental results show that the proposed method is able to automatically select the optimal sampling step that guarantees the highest classification accuracy for fixed transform parameters, when applied to audio/sound signals.