Applying Pseudo-Labels to Partially Labeled Datasets Using the Latent Space Representation of Autoencoders for Image Classification

In this paper the feature extraction capabilities of aut encoders were leveraged to perform classification of data when only a very limited number of training datapoints are labeled. The performance of the autoencoder was improved by using the Nearest Neighbor algorithm to iteratively assign labels to the unlabeled datapoints, and thus increase the pool of labeled data that was subsequently used to improve the classifier. Experiments were performed with the MNIST and Fashion-MNIST datasets, and the proposed method provided good results when using less than 4 elements per class.