The consolidation of learning during sleep: comparing the pseudorehearsal and unlearning accounts

We suggest that any brain-like (artificial neural network based) learning s ystem will need a sleep-like mechanism for consolidating newly learned info rmation if it wishes to cope with the sequential/onging learning of signifi cantly new information. We summarise and explore two possible candidates fo r a computational account of this consolidation process in Hopfield type ne tworks. The "pseudorehearsal" method is based on the relearning of randomly selected attractors in the network as the new information is added from so me second system. This process is supposed to reinforce old information wit hin the network and protect it from the disruption caused by learning new i nputs. The "unlearning" method is based on the unlearning of randomly selec ted attractors in the network after new information has already been learne d. This process is supposed to locate and remove the unwanted associations between information that obscure the learned inputs. We suggest that as a c omputational model of sleep consolidation, the pseudorehearsal approach is better supported by the psychological, evolutionary, and neurophysiological data tin particular accounting for the role of the hippocampus in consolid ation). (C) 1999 Elsevier Science Ltd. All rights reserved.