Classification of hand-written digits by a large scale neural network `CombNET-II´

The authors describe the network structure and the learning procedure of CombNET-II together with experimental results on hand-written digit classification. CombNET-II uses the self-growing neural network learning procedure for training the stem network. After training the stem network, all input data are partitioned into category groups. Then, branch networks are trained for every category group. Backpropagation is utilized to train branch networks. Each branch neural network which is a three-layered hierarchical network has only a small number of connections so that it is easy to tune up. Therefore, CombNET-II has good convergency in the learning process. CombNET-II has been applied to the classification of 6349 printed Kanji characters and to the recognition of 1000 spoken words. CombNET-II correctly classified 99.4% of previously unseen hand-written digits