Wednesday May 2, 2012, at 10h00 in Celestijnenlaan 200A (room 200A.05.001)
Automatic Discovery for Research in Machine Learning
by Dr. Francis Maes
Machine learning research aims at conceiving and analyzing algorithms to solve specific data-related problems. Researchers in this field typically use a trial-and-error approach to reach these goals, in which algorithms are progressively improved on the basis of empirical evidence and/or theoretical analysis. This research process has some important drawbacks: it is inherently biased towards solutions that seem intuitive to humans, it is time-consuming and it is rarely reproducible. In this talk, I will defend the approach which consists in formalizing this trial-and-error approach and in using automatic discovery tools to discover new high-performance algorithms. The talk is structured in two main parts. In the first one, I will formalize the search for high-performance algorithms as a multi-armed bandit (MAB) problem and illustrate the MAB approach on three different problems: discovering control policies, discovering MAB algorithms and discovering reinforcement learning algorithms. In a second time, I will consider the problem of exploring large tree-structured expression spaces and propose some alternatives to Genetic Programming for this task. The proposed approaches formalizes the problem as a "one-player game" and relies on Monte-Carlo search techniques to solve this game. I will illustrate it on symbolic regression and on automatic feature generation. Through these numerous examples, my hope is to show that automatic discovery tools are extremely relevant to research in machine learning and that these tools are a key to quickly unlock several new research avenues.