Image Classification Using Deep Autoencoders

Deep learning refers to computational models comprising of several processing layers that permit display of data with a compound level of abstraction. Till date, several deep learning architectures have been developed, and notable results are attained. The best result often involves an unsupervised pre-training phase followed by supervised learning task. In this work, a particular implementation of deep autoencoders with SVM (Support Vector Machine) layer as a classification layer on the top of the encoding layer is explored. A comparison is made on MNIST dataset with softmax regression function layer and SVM layer as a classification layer with 2 layers and 3 layers SAE (Stack Autoencoders) respectively. Experimental results are evaluated using SAE. It is observed that SVM as classification layer obtains 99.8% accuracy with 0.2% error rate and outperforms softmax regression layer as a classification layer in autoencoders. Further, affect of varying number of neurons in the hidden layers of the autoencoders on the network performance with regard to classification accuracy is also studied.