Self-refreshing memory in artificial neural networks: learning temporal sequences without catastrophic forgetting

While humans forget gradually, highly distributed connectionist networks forget catastrophically: newly learned information often completely erases previously learned information. This is not just implausible cognitively, but disastrous practically. However, it is not easy in connectionist cognitive modelling to keep away from highly distributed neural networks, if only because of their ability to generalize. A realistic and effective system that solves the problem of catastrophic interference in sequential learning of ‘static’ (i.e. non-temporally ordered) patterns has been proposed recently (Robins 1995 Robins, AV. 1995. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7: 123–146. [Taylor & Francis Online] [Google Scholar], Connection Science, 7: 123–146, 1996, Connection Science, 8: 259–275, Ans and Rousset 1997 Ans, B and Rousset, S. 1997. Avoiding catastrophic forgetting by coupling two reverberating neural networks. CR Académie des Sciences Paris, Life Sciences, 320: 989–997.  [Google Scholar], CR Académie des Sciences Paris, Life Sciences, 320: 989–997, French 1997 French, RM. 1997. Pseudo-recurrent connectionist networks: an approach to the ‘sensitivity–stability’ dilemma. Connection Science, 9: 353–379. [Taylor & Francis Online] [Google Scholar], Connection Science, 9: 353–379, 1999, Trends in Cognitive Sciences, 3: 128–135, Ans and Rousset 2000 Ans, B and Rousset, S. 2000. Neural networks with a self-refreshing memory: knowledge transfer in sequential learning tasks without catastrophic forgetting. Connection Science, 12: 1–19. [Taylor & Francis Online], [Web of Science ®] [Google Scholar], Connection Science, 12: 1–19). The basic principle is to learn new external patterns interleaved with internally generated ‘pseudopatterns’ (generated from random activation) that reflect the previously learned information. However, to be credible, this self-refreshing mechanism for static learning has to encompass our human ability to learn serially many temporal sequences of patterns without catastrophic forgetting. Temporal sequence learning is arguably more important than static pattern learning in the real world. In this paper, we develop a dual-network architecture in which self-generated pseudopatterns reflect (non-temporally) all the sequences of temporally ordered items previously learned. Using these pseudopatterns, several self-refreshing mechanisms that eliminate catastrophic forgetting in sequence learning are described and their efficiency is demonstrated through simulations. Finally, an experiment is presented that evidences a close similarity between human and simulated behaviour.