Building efficient learning algorithms: a computational regularization perspective - Lorenzo Rosasco

Описание: The workshop aims at bringing together researchers working on the theoretical foundations of learning, with an emphasis on methods at the intersection of statistics, probability and optimization.


Classical algorithms design in machine learning is based on minimizing an empirical objective over a class of models. The latter imposes a bias on the kind of solutions sought for, and further allows to prevent possible instability. While this approach provides a useful abstraction, in practice it requires a number of further design choices. The impact of these choices is typically assessed separately and often on an empirical basis via trial and errors. In this talk I will consider a least squares learning scenario and show how a number of different algorithmic tricks can indeed be understood within a unifying regularization framework. These observations provides new perspective on how to build algorithms that are accurately as well as resource efficient.