Steepest descent retrieval algorithm for autoassociative neural memory

The proposed steepest descent retrieval algorithm is shown to improve the recovery of stored patterns in autoassociative recurrent neural memories. The algorithm implements the steepest reduction of the computational energy function rather than an ordinary random and unqualified update. The experiments indicate that this update mode is the more efficient when compared to the conventional asynchronous update. Specifically, it performs better in recovery of vectors at smaller Hamming distance and in rejection of stable but spurious memories