Wednesday November 30 , 14h00 - 15h00 , room F107 , Seminar

Vojtech Franc (Center for Machine Perception , Czech Technical University)

Efficient Parallel Algorithm for Structured Output Learning

Many machine learning algorithms are special instances of a convex regularized risk minimization problem. Solving these convex minimization problems can be demanding in real life applications where large data are to be processed. The Bundle Method for Risk Minimization (BMRM) is a generic algorithm for solving the risk minimization problems which provides convergence guarantees. Unfortunately, the BMRM algorithm can be slow on large problems. We propose a parallelized variant of the BMRM algorithm which not only distributes the computations over N processes but which also significantly decreases the number of iterations. This is achieved by approximating the risk function by several cutting plane models instead of using only a single one like in the standard BMRM algorithm.