Reproduce “SimPLE: Similar Pseudo Label Exploitation for Semi-Supervised Classification” 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
Zijian Hu, Zhengyu Yang, Xuefeng Hu, Ram Nevatia
SimPLE introduces a Pair Loss to semi-supervised learning, leveraging similarities between distinct, high-confidence unlabeled samples to extend consistency regularization beyond individual samples. It improves classification accuracy on challenging datasets like CIFAR-100 and Mini-ImageNet and boosts performance in transfer learning scenarios.
Key insights
- Explicitly modeling similarities between distinct, high-confidence unlabeled data points provides valuable information for semi-supervised learning.
- The Bhattacharyya coefficient serves as an effective metric for measuring similarity between pseudo-label probability distributions, guiding the Pair Loss.
- Combining inter-sample consistency (Pair Loss) with traditional within-sample consistency and supervised learning improves model generalization and accelerates convergence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.