A02 Scattering transforms of sparse signals

The scattering transform is based on a designed convolutional neural network using a wavelet filter bank structure. In previous work, alternative filter banks have been proposed as the basis for the scattering transform. While the impact of these choices was initially unclear, our findings show that the network's response behavior is strongly influenced by the interplay between the structured sparsity of both the input signal and the selected filters. To date, this insight is primarily based on analyzing the energy distribution within the corresponding scattering networks. We are now extending this investigation to determine whether similar dependencies also manifest in other network characteristics.