Maintaining stability during new learning in neural networks

A fundamental problem of neural networks and similar distributed systems is that the learning of new information potentially interferes with information already stored in the network. The author reviews the pseudorehearsal solution to this problem proposed by Robins (1995). He describes and evaluates a variant of pseudorehearsal that provides an alternative to the methods current use of significant amounts of temporary storage space. Pseudorehearsal methods allow networks to "self stabilize" in the face of new learning, providing a theoretical framework for modelling-and a practical method for implementing-continuous/ongoing learning with neural networks.