A backpropagation algorithm which automatically determines the number of association units

Presents a modified backpropagation algorithm which iteratively cuts out or adds association units during the learning process, and which is expected to form networks with the minimal number of association units for given learning problems. The principle is rather simple: when the calculation by the conventional backpropagation converges, the effectiveness of each association unit is assessed by the goodness factor and the unit having the least value of it in the layer is removed. On the other hand, when it does not converge, a new unit is added. The goodness factor is defined to represent the total amount of feedforward-propagated signal by the unit, and is easily calculated. In all the learning problems tested, the minimal or nearly minimal networks for the given problems were obtained with high probability. Some examples are presented