Learning the Prediction Distribution for Semi-Supervised Learning with Normalising Flows

As data volumes continue to grow, the labelling process increasingly becomes\na bottleneck, creating demand for methods that leverage information from\nunlabelled data. Impressive results have been achieved in semi-supervised\nlearning (SSL) for image classification, nearing fully supervised performance,\nwith only a fraction of the data labelled. In this work, we propose a\nprobabilistically principled general approach to SSL that considers the\ndistribution over label predictions, for labels of different complexity, from\n"one-hot" vectors to binary vectors and images. Our method regularises an\nunderlying supervised model, using a normalising flow that learns the posterior\ndistribution over predictions for labelled data, to serve as a prior over the\npredictions on unlabelled data. We demonstrate the general applicability of\nthis approach on a range of computer vision tasks with varying output\ncomplexity: classification, attribute prediction and image-to-image\ntranslation.\n