A fast supervised-learning algorithm for large, multilayered neural networks

A method for training multilayered neural networks that uses genetic-algorithm techniques is discussed. Tests show that this method, called GenLearn, is significantly faster than methods that use the generalized delta rule (GDR). GenLearn also has the advantage that, unlike the GDR, it can escape local minima in its search of weight space. In a feedforward, three-layered neural net, the search for an appropriate first-layer weight matrix can be thought of as a search for powerful internal representations of the input data. GenLearn, in searching for the most fit hidden neurons, searches for a globally-optimal internal representation. The hidden neurons that are least fit do not survive and are replaced. Thus, GenLearn is based on survival of fittest hidden neurons. The biggest advantage of the GenLearn procedure over the GDR in training three-layered neural nets is that it scales up to large neural networks much better