Dynamically stable associative learning (DYSTAL): a biologically motivated artificial neural network

Abstract

A biologically motivated computation model of associative learning has been developed. The DYSTAL model is an artificial neural network that extends connectionist ideas by introducing essential features of biological associative learning systems into the model. DYSTAL differs from other artificial neural networks in two principal respects: (1) the rules governing the modification of the connection weights do not involve the output of the element whose input weights are being changed, either implicitly or explicitly; and (2) the rules governing the modification of a connection weight depend on whether or not that connection propagated a signal in immediately prior steps. The essential feature of temporal pairing of stimuli in associative learning is modeled by a learning rule that requires successive firings before synaptic weights are changed. Consequences of the basic properties of DYSTAL include: (1) linear scaling of computational requirements with network size; (2) an architecture that permits feedforward, feedback, and lateral connections, both excitatory and inhibitory; (3) exceptionally rapid learning without an external 'teacher'; and (4) ability to independently associate different ensembles of inputs, thereby functioning simultaneously as an associator and a classifier.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>