Abstract: The support vector machine is successfully applied in many fields of pattern recognition, but there ere several limitations thereof. The relevance vector machine is a I3ayesian treatment, its mathematics model doesn't have regularination coefficient, and its kernel functions don' t need to satisfy Mercer' s condition. The relevance vector machine can present the good generalization performance, and its predictions are probabilistic. We introduced the sparse I3ayesian modes for regression and classification, regarded the relevance vector machine learning as the maximization of marginal likelihood through the model parameters inference, then we described three kinds of training methods and presented the flow of the fast sequential sparse I3ayesian learning algorithm.

Key words: Sparse bayesian model, Relevance vector machine, Support vector machine