Class Aware Auto Encoders for Better Feature Extraction

Abstract

In this work, a modified operation of Auto Encoder has been proposed to generate better features from the input data. General autoencoders work unsupervised and learn features using the input data as a reference for output. In our method of training autoencoders, we include the class labels into the reference data so as to gear the learning of the autoencoder towards the reference data as well as the specific class it belongs to. This ensures that the features learned are representations of individual data points as well as the corresponding class. The efficacy of our method is measured by comparing the accuracy of classifiers trained on features extracted by our models from the MNIST dataset, the CIFAR-10 dataset, and the UTKFace dataset. Features extracted by our brand of autoencoders enable classifiers to obtain higher accuracy in comparison to the same classifiers trained on features extracted by traditional autoen-coders.