The author examines the problem of catastrophic forgetting-the overwriting of old information-in neural networks. He notes that R. Ratcliff's (1990) experiments with rehearsal regimes are a possible solution to catastrophic forgetting and describes sweep rehearsal-a much more effective regime. The use of sweep rehearsal, however, eventually encounters practical limits as the ability to recognize learned items begins to diminish. The author suggests that sweep rehearsal extends the approach of rehearsal mechanisms as far as is practicable, and exposes their eventual limitations.
展开▼