Structure-based neural network learning

摘要

We present a new learning algorithm for the structure of recurrent neural networks. It is shown that any m linearly independent n-dimensional vectors can be stored in at most (n+m-2)-dimensional symmetric network. A storage procedure which satisfies this bound is presented. We propose a new learning procedure for the domain of attraction which preserves both the equilibrium set and the stability property of the original system. It is shown that previously learned attraction regions remain invariant under the proposed learning rule, Our emphasis throughout this brief is on the design of associative memories and classifiers