Title: Learning with implicit regularization and sketching
Authors: Luigi Carratino - University of Genoa (Italy) [presenting]
Lorenzo Rosasco - Unige MIT IIT (Italy)
Abstract: Classically, regularization is achieved imposing explicit constraints on a data fit term, and optimization aspects are considered separately. We discuss how regularization can be controlled implicitly by an optimization method of choice. This latter approach has the advantage that training time controls at the same time statistical accuracy and time complexity of the obtained estimator. Moreover, we discuss how computations can be further reduced considering random projections. Our study bridges optimization and statistical studies.