Superior Accuracy of Decrementing over Incrementing Associative Networks when Operating in Initially Random Connectivities

V.G.Dobson

Vision and Networks Unit, Department of HUmna Sciences, Brunel University,
Uxbridge, Middlesex, UB8 3PH, U.K.

Abstract

This paper examines the performance of associative nets which associate binary patterns by cutting or decrementing links. A simple model of pattern retrieval using links implementing Boolean NOT operations is shown to be very resistant to negative noise due to missing components or input signals, and random connectivity. A more complex model, integrating signals which decay with time, confers optimal resistance to both positive and negative noise in autoassociative circuits, and thus operates as an efficient content-addressable memory implementing 'nearest neighbour' or 'best fit' searches. In random connectivities with both high and low connectivity ratios, error rates of decrementing nets were found to be significantly lower than those of the corresponding incrementing nets. This is because incrementing nets store information in arithmetical rather than logical changes in link gains, and require circuits with arithmetically regular parameter values for accurate retrieval. It is concluded that decrementing nets are simpler and more efficient, and have a more plausible developmental rationale. Recent neurological evidence suggesting that decrementing nets operate in the brain is reviewed.