It is beyond the scope of this website to fully discuss this topic.

ART networks were developed to overcome the stability-plasticity dilemma. Stability-plasticity occurs when learning new data causes unstable conditions and data is lost. Some of the output units are utilized for learning new data.

The architectures of ART 1, ART 2, and fuzzy ART networks are all based on the idea of adaptive resonant feedback between two layers of nodes. These nodes are analogous to the sensory cell groups in the cerebral cortex. The input nodes, designated by F1, respond to input. The output nodes, designated by F2, respond or react to the F1 node activity patterns. The F1 nodes do not directly interact with each other but the F2 nodes do. The F2 nodes are connected in a recurrent competitive on-center off-surround network.

There is a feed-forward and feedback connection between every node in the input and the output layers. There is a logic control and a designated control-1 and control-2 connected to each node in the layer. The logic control is used whenever a valid input is present and compares the stored data.

 There are four (4) phases used in the operation of the network.

 1)      The Initialization Phase.

 2)      The Recognition Phase.

 3)      The Comparison Phase.

 4)      The Search Phase.

To be able to mimic biological behaviour, the emphasis of ART neural networks lies at unsupervised learning and self-organization.

To gain a better understanding of this topic please read the references listed below.

Reference:

http://www.fon.hum.uva.nl/Proceedings/Proceedings21/DavidWeenink/DavidWeenink.html

http://www.geocities.com/CapeCanaveral/1624/art1.html

http://technology.open.ac.uk/tel/cor/t396/2002/hopgood_p215.pdf

http://www.uta.edu/psychology/faculty/levine/EBOOK/ART_networks.pdf

http://cialab.ee.washington.edu/Healy/Papers/Encephalon%20autonomous%20.pdf