Catastrophic Forgetting in Connectionist Networks: Causes, Consequences and Solutions

All natural cognitive systems, and, in particular, our own, gradually forget previously learned information. Consequently, plausible models of human cognition should exhibit similar patterns of gradual forgetting old information as new information is acquired. Only rarely (see Box 3) does new learning in natural cognitive systems completely disrupt or erase previously learned information. In other words, natural cognitive systems do not, in general, forget catastrophically. Unfortunately, however, this is precisely what occurs under certain circumstances in distributed connectionist networks. It turns out that the very features that give these networks their much-touted abilities to generalize, to function in the presence of degraded input, etc., are the root cause of catastrophic forgetting. The challenge is how to keep the advantages of distributed connectionist networks while avoiding the problem of catastrophic forgetting. In this article, we examine the causes, consequences and numerous solutions to the problem of catastrophic forgetting in neural networks. We consider how the brain might have overcome this problem and explore the consequences of this solution. Introduction By the end of the 1980’s many of the early problems with connectionist networks, such as their difficulties with sequence-learning and the profoundly stimulus-response nature of supervised learning algorithms such as error backpropagation had been largely solved. However, as these problems were being solved, another was discovered by McCloskey and Cohen and Ratcliff. They suggested that there might be a fundamental limitation to this type of distributed architecture, in the same way that Minsky and Papert had shown twenty years before that there were certain fundamental limitations to what a perceptron could do. They observed that under certain conditions, the process of learning a new set of patterns suddenly and completely erased a network’s knowledge of what it had already learned. They referred to this phenomenon as catastrophic interference (or catastrophic forgetting) and suggested that the underlying reason for this difficulty was the very thing — a single set of shared weights — that gave the networks their remarkable abilities to generalize and degrade gracefully. Catastrophic interference is a radical manifestation of a more general problem for connectionist models of memory — in fact, for any model of memory —, the so-called “stability-plasticity” problem. The problem is how to design a system that is simultaneously sensitive to, but not radically disrupted by, new input. In this article we will focus primarily on a particular, widely used class of distributed neural network architectures — namely, those with a single set of shared (or partially shared) multiplicative weights. While this defines a very broad class of networks, this definition is certainly not exhaustive. In the remainder of this article we will discuss the numerous attempts over the last decade to solve this problem within the context of this type of network.