The online learning framework is a powerful tool to design fast-learning algorithms able to work in both the stochastic and the adversarial setting. In this talk I will show two examples from each of these settings.

The selection of publicity banners on a webpage is an example of the adversarial setting. Given a customer's profile, the system has to learn which banner to show, taking into account the different and limited positions on the webpage. The only feedback received is the actual action of the user. In particular the system does not observe what would have been the user's actions had other ads been shown, or had the same ads been placed in a different order within the webpage. I will present an online algorithm to efficiently solve this problem, with optimal theoretical guarantees. Empirical results will also be shown to validate the theoretical results.

In the second part of the talk, I will show an application of online learning methods to obtain a parameter-free stochastic gradient descent procedure. Contrary to existing algorithms, it always achieves an optimal convergence rate, without the need to tune any parameter nor any prior knowledge about the optimal solution.

Francesco Orabona is a Research Assistant Professor at the Toyota Technological Institute at Chicago. His research interests are in the area of online learning and transfer learning, with applications to robotics and computer vision. He received a PhD degree in Electrical Engineering at the University of Genoa, in 2007. He was a post-doctoral researcher with Barbara Caputo and Nicolo' Cesa-Bianchi. He is co-author of more than 40 peer-reviewed papers.