Haar Scattering Transform

We will introduce a data-adapted version of scattering transform using Haar wavelets, where the Haar bases are learnt in an unsupervised way. The resulted Haar scattering transform makes use of the affinity of features to build a multi-layer feed-forward network, where at each layer of the network only close neurons are paired together to produce output in the next layer. The Haar scattering network computes invariant representations of non-grid data, by which we mean that, for example, where each coordinate of the data vector corresponds to a node on an unknown graph. The method provides an unsupervised benchmark for deep learning tasks of such datasets.