Reproduce “Sampling Control for Imbalanced Calibration in Semi-Supervised Learning” on OpenResearch

About this paper

Senmao Tian, Xiang Wei, Shunli Zhang

SC-SSL controls which pseudo-labeled images contribute to semi-supervised image classification, aiming to include more examples from minority classes. It estimates the unlabeled class distribution, dynamically adjusts class-wise confidence thresholds, and subtracts an estimated optimization-bias vector from output logits at inference. On CIFAR10-LT, it reaches 86.53% test accuracy in one consistent-distribution setting, but does not lead every tested setting.

Key insights

• The expansion-separation assumption and pseudo-label denoising theorem support utilizing noisy pseudo-labels for tail classes under consistency regularization.
• Dynamically adjusting class-wise confidence thresholds and logit adjustments within an expansive classifier can effectively rebalance gradient contributions and improve feature learning for minority classes.
• The bias term (b_opt) of a classifier trained with a balanced sampling strategy can be leveraged to isolate and correct optimization-induced imbalance in logits during inference.

Summary by alphaXiv.