What is Adaptive resonance theory in unsupervised machine learning?

Adaptive resonance theory (ART) is a collection of algorithms that recognize and predict patterns. Unsupervised and supervised models can be used to partition ART.

The architecture of ART is a self-organizing neural network. The method allows you to learn new mappings while keeping your old knowledge.

ART1 can cluster data similarly to k-means, but it has the advantage of changing the number of clusters based on the data rather than having to define k at runtime.

A comparison field (used to assess how a new feature vector fits inside existing categories), a recognition field (which comprises neurons that represent active clusters), and a reset module are all included in ART1.

The comparison field identifies the cluster in which the input vector most closely fits when applied. If the input vector matches above a vigilance parameter in the recognition field, the recognition field’s connections to the neuron are changed to account for the new vector. Otherwise, a new neuron is formed in the recognition field to account for a new cluster. Current neuron weights are not altered when a new neuron is generated, maintaining their previous information.

This is repeated for all examples in the data collection until no example input vector changes cluster. The training is over at this moment.

ART offers numerous methods that work with binary or real-valued input vectors (ART2). Predictive ART is an ART1/2 variant that requires supervised training.