An experimental evaluation of the optimal capacity of the Hopfield associative memory

Summary form only given, as follows. The author presents results from a series of simulation experiments indicating that if, in storing binary patterns in a Hopfield associative memory, the matrix of connection weights is optimally constructed, a network of n neurons is with high probability able to store very close to n random patterns without error. This is in contrast to the behavior of such memories under the outerproduct storage rule, which yields a storage capacity of only O(n/log n) patterns.<>