A Tour of Sparsity-Aware Learning -- Calling at: Online, Distributed, Robust And Dictionary Learning is coming at 05/23/2016 - 4:00pm
COVL 216 Mon, 05/23/2016 - 4:00pm Sergios Theodoridis Professor, Department of Informatics and Telecommunications, National and Kapodistrian University of Athens Abstract: Learning sparse models has been a topic at the forefront of research for the last ten years or so. Considerable effort has been invested in developing efficient schemes for the recovery of sparse signal/parameter vectors. Moreover, concepts that have originally been developed around the regression task have been extended to more general and difficult problems, such as low-rank matrix factorization for dimensionality reduction, robust learning in the presence of outliers, dictionary learning for “data-dependent” signal representation. Furthermore, online techniques for sparse modeling estimation are attracting an increasing interest, especially in the context of big data applications. Another area which is gaining in importance is distributed learning over graphs. An area, which was mainly inspired and born within the sensor network discipline, but now lends itself, nicely, for big data processing. In this talk, I touch upon all the previously mentioned problems. Sparse modeling of regression tasks is viewed in its online estimation facet, via convex analytic arguments, based on the set-theoretic framework; the emphasis is on very recent extensions of the theory to include non-convex related constraints, which impose sparsity on the model in a much more aggressive manner compared to the more standard, convex, l_1-norm related arguments. In spite of the involved non-convexity, still complexity per time iteration exhibits a linear dependence on the number of unknown parameters; furthermore, strong theoretical convergence results have been established. In the sequel, distributed learning techniques are reviewed with an emphasis on greedy-type batch as well as online versions. The task of robust learning in the presence of outliers is then reviewed and new methods, based on the explicit modeling of the outliers, in the context of sparsity-aware learning, will be presented. The new method, based on greedy-type arguments, enjoys a number of merits, compared to more classical techniques. Furthermore, strong theoretical results have been established, for the first time, in such a type of treatment of the robust estimation task. Finally, dictionary learning, in its very recent online and distributed processing framework, is discussed and new experimental as well as theoretical results will be presented Bio: [node:field-speaker-bio:text] Read more: http://eecs.oregonstate.edu/colloquium/tour-sparsity-aware-learning-call... [1] [1] http://eecs.oregonstate.edu/colloquium/tour-sparsity-aware-learning-calling-online-distributed-robust-and-dictionary-learning
_______________________________________________ Colloquium mailing list [email protected] https://secure.engr.oregonstate.edu/mailman/listinfo/colloquium
