Corrective learning of autoassociative neural network for handwritten numeral recognition

摘要

Autoassociative neural network is a network having the same number of neurons in input and output layers, and the less in the hidden layers. The network is trained using the input vector itself as the desired output. This training leads to minimize the mean square error between the input and the output, however it does not necessary minimize the classification error. This paper proposes a new corrective learning method which moves away the hypersurface of a class to which an input pattern is misclassified. The efficiency of the proposed method is evaluated and studied by handwritten numeral recognition using three layered and five layered autoassociative neural networks.