Adaptive resonance theory

Adaptive resonance theory, abbreviated ART, is a theory of how the brain processes information, developed by Stephen Grossberg and Gail Carpenter, and the name also covers a family of artificial neural network models that follow it. The models use supervised and unsupervised learning and are applied to problems such as pattern recognition and prediction. The central idea is that recognition comes from the interaction of top-down expectations with bottom-up sensory input. An expectation takes the form of a memory template that is compared with the features actually sensed, and while the mismatch stays within a threshold called the vigilance parameter the input is accepted as a member of the expected class. A basic ART network has a comparison field, a recognition field of neurons with lateral inhibition, a vigilance parameter and a reset module, and it commits a fresh neuron when no existing category matches. This answers the plasticity and stability problem of learning new things without erasing old ones. Variants include ART 1, ART 2, ARTMAP, Fuzzy ART, Gaussian ART and TopoART, and critics note that the learned categories of Fuzzy ART and ART 1 depend on the order of the training data.