Ideally a solution to the catastrophic forgetting problem
would facilitate the training of ANNs in a “sequential” or
cumulative way, allowing new information to be learned by a
network at any time. In this paper we briefly review the
pseudorehearsal method proposed by Robins [1995] and
extended in [Robins, 1996; Frean and Robins, 1997; French,
1997; Robins and McCallum, 1997], which provides a
practical and interesting solution to the catastrophic
forgetting problem. We then explore the impact of the
pseudorehearsal mechanism on generalisation.