Distorted handwritten Kanji character pattern recognition by a learning algorithm minimizing output variation

Presents a neural network learning algorithm which adjusts connection weights to minimize error and output variations. The introduction of output variation into the error function creates two types of output variations, increasing output and decreasing output. The algorithm learns two kinds of modified training samples. The magnitude of modification depends on the position of the training sample in the sample distribution and the extent of the learning process. Using distorted Kanji character patterns and noise-modified training samples, the algorithm is compared with two conventional algorithms, backpropagation (BP) and BP trained with noise. The algorithm has a higher recognition rate, and is much more robust