A normal form projection algorithm for associative memory
Bill Baird, Frank H. Eeckman · 1993
this paper is contained in the projection theorem, which details the associative memory capabilities of networks utilizing the normal form projection algorithm for storage of periodic attractors. The algorithm was originally designed, using dynamical systems theory, to allow learning and pattern recognition with oscillatory attractors in models of olfactory cortex. Here we concentrate on mathematical analysis and engineering oriented applications of the algorithm, and briefly discuss biological models at the end. We focus attention on the storage of periodic attractors, since that is the best understood unusual capability of this system. The storage of static and chaotic attractors are discussed as variations on this theme. We hope to give intuitive discussion and geometric perspectives to compliment and clarify the formal analysis. Other approaches to oscillatory memory may be found in [26, 17, 45, 33, 37]. The normal form projection algorithm provides one solution to the problem of storing analog attractors in a recurrent neural network. Associative memory storage of analog patterns and continuous periodic sequences in the same network is analytically guaranteed. For a network with N nodes, the capacity is N