Better Self-training for Image Classification Through Self-supervision

Self-training is a simple semi-supervised learning approach: Unlabelled examples that attract high-confidence predictions are labelled with their and added to the training set, this process being repeated multiple times. Recently, self-supervision—learning without manual supervision by solving an automatically-generated pretext task—has gained prominence in deep learning. This paper investigates three different ways of incorporating self-supervision into self-training improve accuracy image classification: as pretraining only, performed exclusively first iteration self-training, every self-training. Empirical results on SVHN, CIFAR-10, PlantVillage datasets, using both from scratch, Imagenet-pretrained weights, show applying only can greatly accuracy, for modest increase computation time.