ABSTRACT

Early studies in dictionary learning focused on learning a basis for representation. There were no constraints on the dictionary atoms or on the loading coefficients. For problems in sparse representation, the objective is to learn a basis that can represent the samples in a sparse fashion. Dictionary learning is a synthesis formulation; that is, it learns a basis/dictionary along with the coefficients such that the data can be synthesized. There can be an alternate formulation where a basis is learnt to analyze the data to produce the coefficients. Relating transform learning to the dictionary learning formulation, dictionary learning is an inverse problem, while transform learning is a forward problem. As in dictionary learning, transform learning is solved using an alternating minimization approach.