LaSSL: Label-Guided Self-Training for Semi-supervised Learning

The key to semi-supervised learning (SSL) is explore adequate information leverage the unlabeled data. Current dominant approaches aim generate pseudo-labels on weakly augmented instances and train models their corresponding strongly variants with high-confidence results. However, such methods are limited in excluding samples low-confidence under-utilization of label information. In this paper, we emphasize cruciality propose a Label-guided Self-training approach Semi-supervised Learning (LaSSL), which improves pseudo-label generations from two mutually boosted strategies. First, ground-truth labels iteratively-polished pseudo-labels, instance relations among all then minimize class-aware contrastive loss learn discriminative feature representations that make same-class gathered different-class scattered. Second, top improved representations, propagate across potential data manifold at feature-embedding level, can further improve labelling reference neighbours. These strategies seamlessly integrated promoted whole training process. We evaluate LaSSL several classification benchmarks under partially labeled settings demonstrate its superiority over state-of-the-art approaches.