PIPELINING OF ART ARCHITECTURES (FAM, EAM, GAM) WITHOUT MATCH TRACKING

摘要

Adaptive Resonance theory was introduced by Grossberg to address the stability versus plasticity dilemma. That is, how can one design a learning system that is plastic enough to learn new information, and at the same time stable enough not to forget old, important information that it has already learned. In the past two decades a number of ART neural network architectures were introduced in the literature, based on the ART theory. These architectures can solve clustering and classification problems. Our focus in this paper is ART architectures that function like classifiers. ART classifiers have a number of desirable properties, such as guaranteed convergence to a solution for any classification problem of interest, fast convergence to a solution (i.e., they converge in a few epochs, where epoch is a single presentation of all the training data), they can be trained in an on-line fashion, they have the ability to recognize novel inputs, and they can explain the answers that they produce. One of their limitations is that for large database problems, where inevitably a lot of categories (clusters) are created to represent the input data, the convergence to a solution becomes excruciatingly slow, since ART's complexity is proportional to the product of the input patterns and the number of categories created. To address this problem, Castro had suggested a parallel implementation of Fuzzy ARTMAP (one of the ART classifiers) on a Beowulf cluster. Castro's implementation was efficient and general enough to apply to other ART architectures, such as Ellipsoidal ARTMAP and Gaussian ARTMAP, which are two other examples of ART classifiers. In this paper we validate this claim, that EAM and GAM can be implemented effectively on a Beowulf cluster, and we verify this claim by presenting appropriate experimental results. What is also worth noting is that Castro's Fuzzy ARTMAP Beowulf implementation can also be applied to other competitive classifiers, neural network based or not.