Performance Evaluation of a Convolutional Neural Network for Handwritten Digit Recognition

Deep Learning involves learning multiple representation as well as abstraction levels that help make sense of image, audio, or text data. The MNIST dataset of handwritten digit images is frequently used for training, testing, and validating CNN models. This paper focuses on an efficient CNN model with a series of convolutions, ReLU and pooling layers. The model is developed and tested on MNIST data set with 99% accuracy. Further, similar versions of this model are subjected to various parameter variations and the corresponding impact on output parameters are studied.