Combining two methods of recognizing hand-printed digits

Hand-printed digits can be recognized quite well by a feedforward neural network that uses equality constraints between weights to achieve limited translation^ invariance. However, the net has no explicit model of what a digit looks like and this can lead it to make confident errors. An alternative approach, which incorporates much more prior knowledge, is to use explicit deformable models of the digits and to recognize a digit by finding which model fits best. We describe a system that uses learned digit models which consist of splines knowledge, and the elastic matching process is good at rejecting parts of the image that are best explained as noise. However, the elastic model is far from the actual data. So we are developing a hybrid system that combines the best aspects of both approaches. First, the slow elastic matching method is used to accurately label the training data with all the instatiation parameters of the correct digit. Then a feedforward network is trained to produce the fully instantiated digit, rather than just the class of the digit. After training, the neural net is used to initialize the elastic models, and the elastic matching is used to reject the erroneous hypotheses of the neural network.