Reproduce “Uncertainty-aware Sampling for Long-tailed Semi-supervised Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kuo Yang, Duo Li, Menghan Hu, Guangtao Zhai, Xiaokang Yang, Xiao-Ping Zhang
The paper introduces Uncertainty-Aware Dynamic Threshold Selection (UDTS) for long-tailed semi-supervised learning, leveraging Monte Carlo Dropout to estimate pseudo-label uncertainty and employing dynamic, class-specific thresholds for selection. This approach consistently improved accuracy and recall, especially for tail classes, across various long-tailed datasets including CIFAR-LT, STL-10-LT, and TissueMNIST.
Key insights
- Uncertainty in predictions, especially epistemic uncertainty, can effectively identify and filter out unreliable pseudo-labels that might otherwise reinforce class imbalance.
- Combining both prediction confidence and uncertainty allows for the selection of more robust pseudo-labels, particularly crucial for rare tail classes.
- Dynamic, class-specific thresholds are necessary to adapt pseudo-label selection to the evolving learning state of the model and the varying learning difficulties of different classes in long-tailed distributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.