Semi supervised online learning: train only when you need
Universita’ degli Studi di Milano
DATE: June 30, 2009 at 12.00 A.M.
LOCATION: Room 20 – Faculty of Science – Povo (TN)
An important feature of automatic categorization systems is the ability of computing a level of confidence about the classification of the current instance. A small confidence is viewed as an indication that further training is needed for keeping the desired level of accuracy. In this talk we describe classification algorithms that occasionally ask for human supervision achieving a good performance with far less labels than their fully supervised counterparts. Theoretically, these algorithms are shown to dynamically trade off accuracy with number of requested labels, implicitly learning the unknown structure of the data source. We study this trade-off under different assumptions of the data-generation mechanism (from purely stochastic to purely adversarial) and using a simple family of kernel-based classifiers.
CONTACT: Giuseppe Riccardi