'Unlearning' has a stabilizing effect in collective memories.

Published

Journal Article

Crick and Mitchison have presented a hypothesis for the functional role of dream sleep involving an 'unlearning' process. We have independently carried out mathematical and computer modelling of learning and 'unlearning' in a collective neural network of 30-1,000 neurones. The model network has a content-addressable memory or 'associative memory' which allows it to learn and store many memories. A particular memory can be evoked in its entirety when the network is stimulated by any adequate-sized subpart of the information of that memory. But different memories of the same size are not equally easy to recall. Also, when memories are learned, spurious memories are also created and can also be evoked. Applying an 'unlearning' process, similar to the learning processes but with a reversed sign and starting from a noise input, enhances the performance of the network in accessing real memories and in minimizing spurious ones. Although our model was not motivated by higher nervous function, our system displays behaviours which are strikingly parallel to those needed for the hypothesized role of 'unlearning' in rapid eye movement (REM) sleep.

Full Text

Duke Authors

Cited Authors

  • Hopfield, JJ; Feinstein, DI; Palmer, RG

Published Date

  • July 1, 1983

Published In

Volume / Issue

  • 304 / 5922

Start / End Page

  • 158 - 159

PubMed ID

  • 6866109

Pubmed Central ID

  • 6866109

Electronic International Standard Serial Number (EISSN)

  • 1476-4687

International Standard Serial Number (ISSN)

  • 0028-0836

Digital Object Identifier (DOI)

  • 10.1038/304158a0

Language

  • eng