Evaluation of Convolutional Architectures for Offline Handwritten Digit Recognition

Deep learning implementations have resulted in significant performance improvements in several application domains and as such several network architectures have been developed to facilitate their methods. This paper presents a comparative study of two architectures among those which are implemented for handwriting recognition, Highway CNN and LeNet-5. The evaluation is performed on two separate machines for both CPU (Intel-i5 3250M) and a GPU (Nvidia GTX-1060). We compared them not only on the basis of their accuracy, but also their training time, recognition time and their memory requirements. Our experiments demonstrate the advantage of global training and feature mapping on the MNIST dataset.