Application of Sparse auto-encoder in Handwritten Digit Recognition

Abstract

Deep learning and non-supervised learning methods have a wide range of applications in image feature extraction. This article uses MATLAB to train a deep neural network to classify handwritten digital pictures. The deep neural network is formed by stacking multiple sparse auto-encoders, training the data in an unsupervised manner, initializing the weights of the network, and then fine-tuning the network with a reciprocal propagation algorithm. Finally, the images is classified using the soft-max classifier. Sparse reduces the number of dimensions effectively, and the back propagation algorithm is optimized on the cost function, leading to the accuracy rate has been greatly improved, and completing the classification of handwritten numbers.