References & Citations
Statistics > Machine Learning
Title: Online Multiple Kernel Learning for Structured Prediction
(Submitted on 13 Oct 2010)
Abstract: Despite the recent progress towards efficient multiple kernel learning (MKL), the structured output case remains an open research front. Current approaches involve repeatedly solving a batch learning problem, which makes them inadequate for large scale scenarios. We propose a new family of online proximal algorithms for MKL (as well as for group-lasso and variants thereof), which overcomes that drawback. We show regret, convergence, and generalization bounds for the proposed method. Experiments on handwriting recognition and dependency parsing testify for the successfulness of the approach.
Submission history
From: André Filipe Torres Martins [view email][v1] Wed, 13 Oct 2010 20:48:30 GMT (183kb,D)
Link back to: arXiv, form interface, contact.