Extensive usage of prior knowledge improves generalization performance

Neural networks, as a powerful instrument for statistical inference, can be applied to a great variety of classification and regression tasks. As a disadvantage of this generality, networks need much time and data to select a good parameter set during training. Taking handwritten digit recognition as an exemplary application, the author shows that the use of prior knowledge in the problem domain can considerably support the network in finding the relevant structures inherent in the training data and can thus improve the network´s generalization performance.