A Novel Algorithm for Training Polynomial Networks

Modern recognition systems require high-accuracy low-complexity methods. We explore a polynomial network approach to recognition problems. Traditionally, polynomial networks have been difficult to use for recognition because of either low accuracy or problems associated with large training sets. We detail a new training method that solves these problems. For identification, the method partitions the problem to dramatically reduce resources needed. For verification, the method conglomerates all anti-class data into a fixedsize background model. We illustrate the method by applying it to the problem of speaker recognition using voice.