83 Novelty Detection with Probabilistic ARTMAP Neural Networks

In a wide range of applications, designing robust classification system involve detecting patterns sampled from unfamiliar classes. In this paper, extensions to four probabilistic ARTMAP neural networks — ARTMAP-PI, Probabilistic Fuzzy ARTMAP (PFAM), PROBART and Gaussian ARTMAP (GAM) - are proposed to allow for category-based novelty detection. The performance of these extended networks is compared to the ARTMAP-FD and Near-Enough-Neighbor (NEN) algorithms, in terms accuracy and computational complexity. Performance is assessed through a comprehensive set of computer simulations, using a PSO-based training strategy. The pattern recognition problems considered for simulations consist of synthetic data with overlapping class distributions, and with complex decision boundaries with no overlap. When classifying data from familiar classes, simulation results indicate that PFAM generally achieves a classification rate that is significantly higher than or comparable to the other ARTMAP networks. It always provides networks that require a lower compression (thus fewer computational resources). When detecting data from unfamiliar classes, PFAM provides the better performance with data having complex decision bounds, while ARTMAP-PI tends to provide the better performance with data having overlapping decision bounds.