Have you considered evolution? In my own humble efforts I continue to find that evolution can "fit" NNs to problems which escape resolution by training. I haven't tried it on ART yet but am certain that the phenomenon generalizes. Regarding implementation ... I always use a fine-grained, biologically-inspired OO approach sans matrices. PEs become members of a set (cluster of processors) and each weight becomes an independent entity in a set of afferent or efferent connections. This makes it easy to realize generalized patterns of connectivity, shortcuts, and sparse networks (with an 'eye' directed towards minimizing a hardware implementation).

Prashant,

you might want to look at :

A Massively Parallel Architecture for a Self-Organizing Neural Pattern Recognition Machine by Gail A. Carpenter and Grossberg (page 89)

In it the weights are referred to as traces. Traces obey the update rule of:

L/((L − 1) + |X|)

*L is a parameter greater than one and is usually equal to 2

*|X| is the summation of the Output vector from Comparison layer (see page 79 for the meaning).

You can verify this in a much simpler explanation in:

Adaptive Resonance Theory 1 (ART1) Neural Network Based Horizontal and Vertical Classification of 0-9 Digits Recognition by Mbaitiga Zacharie

I also verified the interpretation using the book by Pandia and the book by Laurene Faussett "Fundamentals of Neural Networks). In these two references they add an extra term which I could not match to the original paper.

I have to admit that reading Carpenter's and Grossberg's papers is not as easy as other papers out there. I usually read another author's paper on ART and verify it with the original source.

Hope this helps