Abstract
We propose a new data augmentation technique for semi-supervised learning settings that emphasizes learning from the most challenging regions of the feature space. Starting with a fully supervised reference model, we first identify low confidence predictions. These samples are then used to train a Variational AutoEncoder (VAE) that can generate an infinite number of additional images with similar distribution. Finally, using the originally labeled data and the synthetically generated labeled and unlabeled data, we retrain a new model in a semi-supervised fashion. We perform experiments on two benchmark RGB datasets: CIFAR-100 and STL-10, and show that the proposed scheme improves classification performance in terms of accuracy and robustness, while yielding comparable or superior results with respect to existing fully supervised approaches
Abstract (translated)
URL
https://arxiv.org/abs/2209.08174
