P. Irofti, “
The effect of atom replacement strategies on dictionary learning”, in
INTERNATIONAL TRAVELING WORKSHOP ON INTERACTIONS BETWEEN SPARSE MODELS AND TECHNOLOGY, Aalborg, Denmark, 2016.
AbstractThe sparse representations field presents a wide set of algorithms for learning overcomplete dictionaries. During the learning process many of the dictionary columns remain unused by the resulting representations. In this paper we present a few replacement strategies and their direct impact on a set of popular algorithms such as K-SVD. Experiments show significant reductions in the representation error and also evidentiate clear differences between the strategies.