Improved input representation for enhancement of neural network performance

Summary form only given. An important consideration for the implementation of associative memory is the storage capacity of the network. For a Hopfield net, the memory capacity for uncorrelated patterns is approximately 0.25 N/log N, where N is the number of neurons. In general, the capacity for information storage is proportional to the number of synapses. For fully connected networks the number of synapses scales as N/sup m+1/, where m is the order of the network. Higher order networks have a much greater storage capacity than a Hopfield net for an equivalent number of neurons. Simply increasing the number of neurons will not always increase the storage capacity. This is shown by simulation results. It is also shown that a simple modification of the pattern vector to have zero bias will provide an even more significant increase in the performance of an associative memory network.<>