35) Players must have a sufficient arsenal of palimpsests or forgetful learning rules. Palimpsest rules for human neural networks do not suffer from catastrophic forgetting. Instead they selectively forget older memories in order to store new patterns. Standard palimpsest learning algorithms have a capacity of up to 0:05n, where n is the size of the network. Here a new learning rule is introduced. This rule is local and incremental. It is shown that it has palimpsest properties, and it has a palimpsest capacity of about 0:25n. The price is that the number of synaptic states is not a property of the synapse, it increases with the size of the Rule. An extension to memory-based learning is described in which automatically induced Rules are used as binary features. These features have an "active" value when the left-hand side of the underlying rule applies to the instance. The Ripper Rule induction algorithm is adopted for the selection of the underlying rules. The similarity of a memory instance to a new instance is measured by taking the sum of the weights of the matching rules both instances share. Experiments indicate that (i) the method works equally well or better than Ripper Rules on various language learning and other datasets; (ii) the method does not necessarily perform better than default memory-based learning, but (iii) when multivalued features are combined with the rulebased features, some slight to significant improvements are observed.
Please be sure to Bookmark this Rule and commit to memory Key Words contained within it to use in a Search, as it has been known to spontaneously relocate in a random manner.
This message was edited by MrKlaatu on 10/09/02 01:09.34.
