ARTMAP-DS: Pattern Discrimination by Discounting Similarities

ARTMAP-DS extends fuzzy ARTMAP to discriminate between similar inputs by discounting similarities. When two or more candidate category representations are activated by a given input, features that the candidate representations have in common are ignored prior to determining the winning category. Simulations illustrate the network's ability to recognize similar inputs, such as STAR. and START, in a noisy environn1ent. 1 Focusing Attention on Small Differences ARTMAP-DS is a supervised neural network for learning and recognition. The network extends fuzzy AR.TMAP (Carpenter eta!., 1992) to discriminate between similar inputs by discotmting similarities. The network ftmctions by focusing attention on differences between candidate category representations activated by a given input, and then checking to see which features are in fact present in the input, ignoring features that the candidate representations have in common. Attentional focusing is particularly needed in syllable and word recognition applications, where a primacy gradient input representation (Grossberg, 1978) may cause low-an1plitude feature representations (in the later parts of sequences) that are vulnerable to input error and processing noise. A high value of the vigilance paran1eter, p, is needed to ensure that a fuzzy ART network can distinguish between similar input sequences such as STAR. and START (Wilson, 1996; Carpenter & Wilson, 1997); but a high value also prevents the system from correctly classifying noisy inputs. The complement-coded input representation used in fuzzy ART (Carpenter, Grossberg, & Rosen, 1991) exacerbates this problem, since the contribution in the input from the phonemes or syllables that are present may be largely masked by the contribution from the larger nun1ber that are absent. With ARTMAP-DS, a difference in the later part of the input sequence is not much harder to detect than an earlier one.