Individual Weighting of Unlabeled Data Points via Confidence-Awareness in Semi-Supervised Learning
Farzin Ghorban, Nesreen Hasan, Jörg Velten, Anton Kummert · 2022
In existing semi-supervised learning (SSL) approaches, the contributions of labeled and unlabeled data points to the training objective are expressed as two separated terms for the respective sets. A single weight specifies how much the unlabeled samples contribute to the overall loss and is carefully scaled in the course of the training. In this work, we propose a SSL framework in which the contribution of each unlabeled sample is individually scaled based on the confidence generated by a teacher model. We demonstrate that our scaling mechanism enables the student model to efficiently learn from heavily interpolated and augmented samples and achieve state-of-the-art performance.