Image Classification of partially labeled Datasets using the Latent Space Representation of Autoencoders

Machine learning algorithms often require large amounts of labeled data to work efficiently. Various methods have been developed to utilize unlabeled data as an aid in training. In this paper, a method is explored of using the latent space generated by autoencoders to perform classification of images, using only a subset of training data as labeled. The labeled data are mapped to their latent space position, and then an SVM model is fitted to perform segmentation of the latent space based on those mappings and perform classification on new images. Experimental results are perform on the MNIST database, and it is shown that the accuracy of the classifier falls gradually as the percentage of data used as labeled data is reduced.