Speeding up the decision making of support vector classifiers

In this paper, we propose a new approach for speeding up the decision making of support vector classifiers (SVC) in the context of multi-class classification. A two-stage system embedded within a probabilistic framework is presented. In the first stage we pre-estimate the posterior probabilities with a model-based approach and we re-estimate only the highest probabilities with appropriate SVCs in the second stage. We have tested our system on the benchmark database MNIST and the results show that our dynamic classification process allows to speedup the full "pairwise coupling" SVCs by a factor of 7.7 while preserving the accuracy. In addition, although the "one against all" strategy estimate slightly betters probabilities, our modular architecture seems more adapted to large multi-class problems.