Geometric Regularization Improves Long-Tailed Semi-Supervised Learning.

Real-world semi-supervised learning (SSL) often struggles with two major issues: highly imbalanced, long-tailed label distributions and the propagation of errors from noisy pseudo-labels. These challenges can severely limit a model's ability to generalize effectively and exacerbate confirmation bias, where the model reinforces its own incorrect predictions. To tackle these problems, researchers propose Gaussian Bridge Consistency (GBC), a new framework that leverages geometric regularization. GBC operates by creating semantic interpolation paths in the latent space, connecting unlabeled samples to reliable class prototypes stored in a dynamic Prototype Atlas. This atlas continuously updates with high-quality labeled and pseudo-labeled exemplars for each class. For every unlabeled instance, GBC forms a class-conditional Gaussian Feature Bridge, allowing the student model to smoothly transition from uncertain predictions to stable class prototypes. A bridge consistency loss is applied along this path to ensure alignment with a geometrically interpolated target distribution. Additionally, the paper introduces BridgeMix, a confidence-aware feature mixing strategy that interpolates both sample and anchor pairs, further enhancing cross-sample generalization. Extensive experiments confirm GBC's robustness and effectiveness, particularly in improving performance for under-represented long-tail classes in realistic SSL settings.