Decrementing associative networks

Decrementing networks are simple connectionist systems for performing auto- and heteroassociative learning of sparse patterns with recall from noisy data. They consist of spontaneously active elements operating upon one another entirely through negative (decrementing) processes: signals are inhibitory and powerful, thresholds are negative, signal decay is essential for retrieval of noisy patterns, and information is stored by decrementing or cutting links. These powerful inhibitory binary links represent deterministic constraints operating between pattern components. The use of binary links enables the nets to be constructed from circuits resembling conventional programmable hardware. They also allow decrementing networks to implement combinatorial logics similar to those required for production rules and hypothesis testing in conventional expert systems.<>