CSAIL Event Calendar: Previous Series
Efficient and Principled Learning Algorithms for Real World Problems
Speaker: Francesco Orabona , Università degli Studi di Milano
Relevant URL: http://cbcl.mit.edu/
Most of the research in machine learning has been directed to the problem of binary classification, given a training set and a test set acquired in a very controlled way. Even if this is a fundamental problem, still it does not fit well important real-world tasks. In this talk I will show some of the results I have presented in the literature in the last years, that aim at trying to solve interesting real-world problems with new theoretically motivated algorithms. I will present results in the IID setting and in the adversarial one. In particular, for the first setting, I will present a new algorithm for transfer learning, that automatically selects the relevant sources of prior information and uses them to bootstrap the performance in a new task with few labeled samples. For the second one, I will introduce a general framework for online learning with potential functions, and instantiations of this framework to the problem of multi kernel learning. As a last algorithm, I will talk about how to do active learning in the adversarial setting, presenting an algorithm able to work with minimal hypothesis, with a sub-linear regret bound and with a bound on the rate of queries controlled by the user.