摘要 |
<p>A method of training a neural network (2) having dynamically adjustable parameters controlled by a controller (10) which determine the response of the network (2). A set of input vectors (I1 to In) are input to network (2) at an input port (4). The corresponding set of output vectors (O'1 to O'n) provided by the network (2) are compared to a target set of output vectors (O1 to On) by an error logger (12) which provides to the controller (10) a measure of similarity of the two sets. The controller (10) is arranged to alter the dynamic parameters independence on the average number of occasions the output vectors are different from the respective target output vectors. Measuring the similarity of the whole of the output set and target set and adjusting the parameters on this global measure rather than on the similarity of pairs of individual vectors provides enhanced training rates for neural networks having a data throughput rate that can be higher than the rate at which the parameters can be adjusted.</p> |