Topics in neural networks
Some aspects of the behavior of several neural networks are considered. The original Hopfield Associative Memory (HAM) is examined, and a lower bound on the number of spurious minima is derived when the stored memories are orthogonal. Two locally interconnected variations of the basic HAM network are proposed in which the maximum distance between two neurons that can be connected is upper bounded by B. It is shown that for such locally interconnected networks containing N neurons, if B/N {yields} 0 as N {yields} {infinity} then the capacity of the network is determined by B and is independent of N. A macroscopic-analysis technique first proposed by Amari for networks with random, nonsymmetric connection weights is modified to show that HAMs must have either one or two macroscopic stable states. The analysis and simulations show that the macroscopic behavior of networks with symmetric and nonsymmetric connections are qualitatively similar. A new class of neural networks derived from the trellis-graph representation of a convolutional code is proposed. Such a network can be viewed as a collection of winner-take-all networks interconnected to reflect the structure of the trellis graph.
- Research Organization:
- Princeton Univ., NJ (USA)
- OSTI ID:
- 6988011
- Country of Publication:
- United States
- Language:
- English
Similar Records
Optical neural computing for associative memories
Phase transitions in dilute, locally connected neural networks