Online Dictionary Learning with Group Structure Inducing Norms

• Sparse coding. • Structured sparsity (e.g., disjunct groups, trees): increased performance in several applications. • Our goal: develop a dictionary learning method, which – enables general overlapping group structures, – is online: fast, memory efficient, adaptive, – applies non-convex sparsity inducing regularization: ∗ fewer measurements, ∗ weaker conditions on the dictionary, ∗ robust (w.r.t. noise, compressibility). – can deal with missing information. Current approaches can exhibit two of these features at most.