Adaptive Resonance Theory

Adaptive resonance theory makes use of two terms used in the study of brain behavior:

(1) stability and

(2) plasticity.

The stability/plasticity dilemma is the ability of a system to preserve the balance between retaining previously learned patterns and learning new patterns. Two layers of neurons are used to realize the idea: (1) a "top" layer, an output, concept layer, and (2) a "bottom" layer, an input, feature layer. Two sets of weights between the neurons in the two layers are used. The top-down weights represent learned patterns, expectations. The bottom-up weights represent a scheme for new inputs to be accommodated in the network. Patterns, associated to an output node j, are collectively represented by the weight vector of this node t_j(top-down weight vector). The reaction of the node j to a particular new input vector is defined by another weight vector bj (bottom-up weight). The key element inGrossberg's realization of the stability/plasticity dilemma is the control of the partial match between new feature vectors and ones already learned which is achieved by using a parameter called vigilance or the vigilance factor.

Vigilance controls the degree of mismatch between the new patterns and the learned (stored) patternsthat the system can tolerate.

The ART1 learning algorithm is given in figure It consists of two major phases. The first presents the input pattern and calculates the activation values of the output neurons. The winning neuron is defined. The second phase calculates the mismatch between the input pattern and the pattern currently associated with the winning neuron. If the mismatch is below a threshold (vigilance parameter), this

pattern is updated to accommodate the new one. But if the mismatch is above the threshold, the procedure continues for finding either another output neuron or a new one, as the input pattern has to be associated with a new-concept, new-output neuron.

The ART neural networks are ideal for conceptualization, clustering, and discovering existing types and number of classes in a database.

ART1 was further developed into ART2 (continuous values for the inputs) and ART3 (Carpenter and Grossberg, 1990). The ART3 model is closer to the synaptic processes in real neurons. Algorithms for supervised learning in ART3 architecture have also been developed.