CONVEX OPTIMIZATION IN MACHINE LEARNING
Gert Lanckriet, Department of Electrical Engineering and Computer Science, University of California, Berkeley, USA

Abstract: The connection between machine learning and convex optimization techniques like linear programming and convex quadratic programming has already a long history (think about Vapnik's SVMs or Mangasarian and Bennett's LP SVMs). More recent convex optimization tools like Second Order Cone Programming (SOCP) and Semi-Definite Programming (SDP) are becoming more popular because of being simple and powerful at the same time. In the talk "Convex Optimization in Machine Learning", I will outline some applications of more recent convex optimization tools in machine learning. First, I'll talk about the Minimax Probability Machine (MPM), a distribution-free approach to binary classification, which makes use of SOCP. Secondly, I will spend most of the time looking at the problem of learning the kernel matrix using SDP (e.g. optimization of the margin for SVMs). The latter is still work in progress, so I would be glad to spend time as well discussing new ideas.

This seminar was held at the Department of Computer Science, Royal Holloway, University of London on 7 January 2002.