“A Self-Optimizing, Non-Symmetrical Neural Net for Content Addressable Memory and Pattern Recognition”, 1986-10 (; backlinks):
A natural, collective neural model for Content Addressable Memory (CAM) and pattern recognition is described. The model uses non-symmetrical, bounded synaptic connection matrices and continuous valued neurons. The problem of specifying a synaptic connection matrix suitable for CAM is formulated as an optimization problem, and recent techniques of Hopfield are used to perform the optimization.
This treatment naturally leads to two interacting neural nets. The first net is a symmetrically connected net (master net) containing information about the desired fixed points or memory vectors. The second net is, in general, a non-symmetric net (slave net), whose synapse values are determined by the master net, and is the net that actually performs the CAM task. The two nets acting together are an example of neural self-organization.
Many advantages of this master/slave approach are described, one of which is that non-symmetric synaptic matrices offer a greater potential for relating formal neural modeling to neurophysiology. In addition, it seems that this approach offers advantages in application to pattern recognition problems due to the new ability to sculpt basins of attraction.
The simple structure of the master net connections indicates that this approach presents no additional problems in reduction to hardware when compared to single net implementations.
…Simulations were performed on VAX780s and Crays. It is worth pointing out that machine precision has little effect on final answers, but can radically affect the time taken to converge to a final answer.