Invariance is at the heart of many recognition problems.
On images and audio, transformation groups such as translations, rotations or frequency transpositions are a major source of intra-class variability that needs to be processed efficiently. Scattering operators can be constructed on any compact transformation group by specifying a family of wavelet modulus decompositions, which are then cascaded to yield informative and stable representations.
In this talk, we will consider the problem of learning the invariance from data, mainly unlabeled. We will first consider the problem of group learning, and then we will discuss a general scheme for generalizing scattering operators, which illustrates its links with the problems of group sparsity, slow feature learning, bi-clustering
and phase recovery. We will present some recent developments as well as some open and current questions.