Pipelining of Fuzzy Artmap (fam) without Match Tracking

Fuzzy ARTMAP (FAM) is a neural network architecture that can establish the correct mapping between real-valued input patterns and correct labels in a variety of classification problems. Nevertheless, as the size of the dataset grows to thousands and hundreds of thousands, FAM’s convergence time slows down considerably. In this paper we focus on a FAM variant called no-match tracking FAM (NMT-FAM). We propose a coarse grain parallelization of the NMT-FAM, based on a pipeline, and show that the parallelization strategy achieves linear speed-up in the order of p (number of processors). Experiments on the Covertype database support our results. We have also shown, but not included in this paper, that the parallelized NMT-FAM is equivalent to the sequential NMT-FAM, it also possesses a number of good properties. Our work in this paper is an effort in the direction of demonstrating that FAM can, through appropriate parallelization strategies, be used to mine data from large databases.