ABSTRACT
The nature of interference effects in human memory is empirically well understood. Acquisition of new information leads to gradual and limited unlearning or forgetting of previous information. However, parallel distributed processing models, or neural networks, have had mixed success in explaining the nature of interference and forgetting. On one hand, the linear TODAM model has been shown to be capable of gradual unlearning (Lewandowsky & Murdock, 1989a), but on the other hand non-linear back propagation networks have been shown to suffer from “catastrophic interference” (McCloskey & Cohen, 1989). Several simulations explored the reasons for this discrepancy and showed that even non-linear back-propagation networks do not exhibit catastrophic interference, and are capable of gradual unlearning, when stimuli and responses are represented by uncorrelated (zero-centered) vectors and when a continuous retrieval measure is used.
