Sinkhorn Label Allocation: Semi-Supervised Classification via Annealed Self-Training

Self-training is a standard approach to semi-supervised learning where the's own predictions on unlabeled data are used as supervision during. In this paper, we reinterpret this label assignment process as an transportation problem between examples and classes, wherein the cost assigning an example to a class is mediated by the current predictions of classifier. This formulation facilitates a practical annealing strategy for assignment and allows for the inclusion of prior knowledge on class via flexible upper bound constraints. The solutions to these problems can be efficiently approximated using Sinkhorn iteration, enabling their use in the inner loop of standard stochastic optimization. We demonstrate the effectiveness of our algorithm on the CIFAR-10, -100, and SVHN datasets in comparison with FixMatch, a state-of-the-art-training algorithm. Our code is available at: //github. com/stanford-futuredata/sinkhorn-label-allocation.