Home // FUTURE COMPUTING 2011, The Third International Conference on Future Computational Technologies and Applications // View article

The Hopfield-type Memory Without Catastrophic Forgetting

Authors:
Iakov Karandashev
Boris Kryzhanovsky
Leonid Litinskii

Keywords: Associative memory; catastrophic forgetting; quasi-Hebbian matrix.

Abstract:
We analyzed a Hopfield-like model of artificial memory that reproduces some features of the human memory. They are: a) the ability to absorb new information when working; b) the memorized patterns are only a small part of a set of patterns that are written down in connection matrix; c) the more the pattern was shown during the learning process, the better the quality of its recognition. We used the Hebb rule, but each pattern was supplied with its own weight. The weight is proportional to the frequency of the pattern showing during the learning of the network. As a result unlimited number of patterns can be written down in the connection matrix, and that would not lead to the memory destroying (as it has place in the standard Hopfield model). However, only the patterns that were shown rather frequently would be recognized: their weights have to be larger a critical value. For analyzed variants of the weights distribution the storage capacity was estimated as ~0.05N-0.06N, where N is the dimensionality of the problem.

Pages: 57 to 61

Copyright: Copyright (c) IARIA, 2011

Publication date: September 25, 2011

Published in: conference

ISSN: 2308-3735

ISBN: 978-1-61208-154-0

Location: Rome, Italy

Dates: from September 25, 2011 to September 30, 2011