Learning Additive Models Online with Fast Evaluating Kernels

Abstract. We develop three new techniques to build on the recent advances in online learning with kernels. First, we show that an exponential speed-up in prediction time per trial is possible for such algorithms as the Kernel-Adatron, the Kernel-Perceptron, and ROMMA for specific additive models. Second, we show that the techniques of the recent algorithms developed for online linear prediction when the best predictor changes over time may be implemented for kernel-based learners at no additional asymptotic cost. Finally, we introduce a new online kernelbased learning algorithm for which we give worst-case loss bounds for the -insensitive square loss.

• Type: Conference
• Year: 2001
• Where: COLT
• Authors: Mark Herbster
• Keywords: COLT 2001, Machine Learning, Online Linear Prediction, Specific Additive Models, Worst-case Loss Bounds