Projected Belief Networks with Discriminative Alignment for Classifying Marine Mammals

Combining generative and discriminative classifiers has the potential to attain large performance improvements because they operate in fundamentally different ways, and tend to make independent errors. However this potential is difficult to realize because generative classifiers tend to perform poorly. In this paper, we create a generative classifier using a projected belief network (PBN) in conjunction with discriminative alignment (DA) that rivals a CNN and achieves significant performance improvements when combined. New classification experiments using marine mammal vocalizations are shown in which the error rate is cut in half by combining with a conventional CNN.