Self-orthogonalization associative memories

The authors present a generalized Hopfield algorithm which is based on the Gramm-Schmidt orthogonal process and the gradient descent approach. The method studies the correlation between input and stored vectors and reduces the cross-correlation noise by using the orthogonal technique. Simulation results have shown an increase in storage capacity with respect to the number of stored vectors. Although one can apply the K-L transform or the discrete cosine transform on the training patterns, the proposed model is much better for practical implementation using a neural network. A significant improvement of the signal-to-noise ratio is obtained. The model is implementable on any `inner product´ version of the Hopfield machine