Group-structured and independent subspace based dictionary learning
PhD thesis, Eötvös Loránd University.
Thanks to the several successful applications, sparse signal representation has become one of the most actively studied research areas in mathematics. However, in the traditional sparse coding problem the dictionary used for representation is assumed to be known. In spite of the popularity of sparsity and its recently emerged structured sparse extension, interestingly, very few works focused on the learning problem of dictionaries to these codes.
In the first part of the paper, we develop a dictionary learning method which is (i) online, (ii) enables overlapping group structures with (iii) non-convex sparsity-inducing regularization and (iv) handles the partially observable case. To the best of our knowledge, current methods can exhibit two of these four desirable properties at most. We also investigate several interesting special cases of our framework and demonstrate its applicability in inpainting of natural signals, structured sparse non-negative matrix factorization of faces and collaborative filtering. Complementing the sparse direction we formulate a novel component-wise acting, epsilon-sparse coding scheme in reproducing kernel Hilbert spaces and show its equivalence to a generalized class of support vector machines. Moreover, we embed support vector machines to multilayer perceptrons and show that for this novel kernel based approximation approach the backpropagation procedure of multilayer perceptrons can be generalized.
In the second part of the paper, we focus on dictionary learning making use of independent subspace assumption instead of structured sparsity. The corresponding problem is called independent subspace analysis (ISA), or independent component analysis (ICA) if all the hidden, independent sources are one-dimensional. One of the most fundamental results of this research field is the ISA separation principle, which states that the ISA problem can be solved by traditional ICA up to permutation. This principle (i) forms the basis of the state-of-the-art ISA solvers and (ii) enables one to estimate the unknown number and the dimensions of the sources efficiently. We (i) extend the ISA problem to several new directions including the controlled, the partially observed, the complex valued and the nonparametric case and (ii) derive separation principle based solution techniques for the generalizations. This solution approach (i) makes it possible to apply state-of-the-art algorithms for the obtained subproblems (in the ISA example ICA and clustering) and (ii) handles the case of unknown dimensional sources. Our extensive numerical experiments demonstrate the robustness and efficiency of our approach.