English

Revisiting Consistency Regularization for Semi-Supervised Learning

Computer Vision and Pattern Recognition 2021-12-14 v1

Abstract

Consistency regularization is one of the most widely-used techniques for semi-supervised learning (SSL). Generally, the aim is to train a model that is invariant to various data augmentations. In this paper, we revisit this idea and find that enforcing invariance by decreasing distances between features from differently augmented images leads to improved performance. However, encouraging equivariance instead, by increasing the feature distance, further improves performance. To this end, we propose an improved consistency regularization framework by a simple yet effective technique, FeatDistLoss, that imposes consistency and equivariance on the classifier and the feature level, respectively. Experimental results show that our model defines a new state of the art for various datasets and settings and outperforms previous work by a significant margin, particularly in low data regimes. Extensive experiments are conducted to analyze the method, and the code will be published.

Keywords

Cite

@article{arxiv.2112.05825,
  title  = {Revisiting Consistency Regularization for Semi-Supervised Learning},
  author = {Yue Fan and Anna Kukleva and Bernt Schiele},
  journal= {arXiv preprint arXiv:2112.05825},
  year   = {2021}
}

Comments

Published at GCPR2021 as a conference paper

R2 v1 2026-06-24T08:12:57.774Z