Learning Data Representation, Invariance and Discriminability
Lorenzo Rosasco, Ph.D.
DIBRIS, Universita’ degli Studi di Genova
LCSL, Massachusetts Institute of Technology & Istituto Italiano di Tecnologia
Artificial intelligent systems have recently achieved impressive results, enabled by machine learning approaches, that is data-driven computational techniques. The main bottleneck of current systems is the large, if not huge, amount of human supervision required, and developing unsupervised, or weakly supervised, methods to learn data representation is widely acknowledged to be key to overcome this limitation. Here data representation refers broadly to the processing of raw data in a format enabling better classification. In this talk, I will review recent attempts to learn representation from data, discussing connection with artificial neural networks and computational neuroscience models of the visual cortex.