Title:The Extreme Value Machine

Abstract:This paper provides a novel characterization of the max-margin distribution in input space. The use of the statistical Extreme Value Theory (EVT) is introduced for modeling margin distances, allowing us to derive a scalable non-linear model called the Extreme Value Machine (EVM). Without the need for a kernel, the EVM leverages a margin model to estimate the probability of sample inclusion in each class. The EVM selects a near-optimal subset of the training vectors to optimize the gain in terms of points covered versus parameters used. We show that the problem reduces to the NP-hard Set Cover problem which has a provable polynomial time approximation. The resulting machine has comparable closed set accuracy (i.e., when all testing classes are known at training time) to optimized RBF SVMs and exhibits far superior performance in open set recognition (i.e., when unknown classes exist at testing time). In open set recognition performance, the EVM is more accurate and more scalable than the state of the art.