A back-propagation associative memory for both positive and negative learning

Summary form only given, as follows. A method is proposed for using a multilayer network, such as one trained using backpropagation as an associative memory. Such networks may be used for a variety of purposes, of which two principal applications could be a nonlinear associative memory with fast convergence, or as a means for testing multilayered systems after training. The basic principle involves the use of an output error signal as an energy function. Gradient descent with simulated annealing can then be used to reconstruct the inputs. The use of a ´quality´ hint neuron also allows some input patterns to be inhibited, while others are encouraged.<>