Gradual reduction of hidden units in the back propagation algorithm, and its application to blood cell classification

Three layer neural networks are applied to the blood cell classification. Optimization of the number of hidden layers using a gradual reduction method (GRM), which we propose, is mainly discussed. The applicability of two types of BP algorithms, standard BP plus GRM and BP with forgetting (BPF) plus GRM, are compared. From results, BPF is found to suit the above purpose better and is believed to be applicable for practical utilization.