Relaxing backpropagation networks as associative memories

One of the most attractive features of associative memories (AM) is their abilities of associative recall, especially recall by incomplete or noisy inputs. However, most existing neural network AM models suffer from their limited storage capacities. In this paper a new model for autoassociative memory based on a novel use of backpropagation (BP) networks is proposed. In this model, recording is done by establishing pattern auto-associations using BP learning, and recall by continuously feeding back the output to the input until the network relaxes into a stable state. The convergence of the recall process is analyzed. The validity of the model is tested by computer experiments. The experiment results show that O(2ⁿ) patterns of length n can be correctly stored, and the recall quality with noisy inputs compares very favorably to conventional AM models