Improved input representation for enhancement of neural network performance

The performance of an associate memory network depends significantly on the representation of the data. For example, it has already been recognized that bipolar representation of neurons with -1 and +1 states out- perform neurons with on and off states of +1 and 0 respectively. This paper will show that a simple modification of the pattern vector to have zero bias will provide even more significant increase for the performance of an associative memory network. The higher order algorithm is used for the numerical simulation studies of this paper. To the lowest order this algorithm reduces to the Hopfield model for auto-associative memory and the bidirectional associative memory (BAM) for hetero-associative memory model respectively. 16 refs., 4 figs., 1 tabs