Learning Multiple Classifiers with Dirichlet Process Mixture Priors

Introduction A real world classification task can often be viewed as consisting of multiple subtasks. In remote sensing, for example, one may have multiple sets of radar images, each collected at a particular geographical location, with the aim of designing classifiers for detecting objects of interest in images at all locations. In this situation, one can either learn a single classifier from simple pooling of images from different locations; or learn multiple classifiers, each for a particular location and based on using images from that location only. Unfortunately, neither of the two are optimal, because the first ignores the difference between different locations and the second ignores the analogy between them.