Neural Approach to the Problem of Pattern Recognition of the Ten Digits

For the problem of pattern recognition of the ten digits, four learning algorithms of the supervised Hebbian learning, the pseudo-inverse learning, the Widrow-Hoff learning and the backpropagation learning are studied with neural nets. Progressively, these learning algorithms as well as network structures are investigated to improve the performance of the neural networks for the recognition task. For a linear associator with the supervised Hebbian learning rule and the concept of autoassociative memory, the performance of the network is not as good as expected. Although the pseudo inverse method can improve the recognition rate, it cannot stand firm from the attack of noise. For the Widrow-Hoff learning with the Adaline network, it is able to perform the task of pattern recognition, but is still suffered from noise corruption in the input patterns. With hidden layer structure, the performance of the multi-layer neural network using the back-propagation algorithm is over that of the other three single-layer neural networks trained with the other algorithms. Noised patterns are tested. The recognition rate of one-bit corrupted patterns is 100%, and for two bits corrupted, the identification rate can still reach more than 98%. If more bits in the input patterns are corrupted, more information is lost from the patterns, and neural nets may get more difficult to perform the task of pattern recognition correctly.