Reproduce “Real-Fake: Effective Training Data Synthesis Through Distribution Matching” 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
Jianhao Yuan, Jie Zhang, Shuyang Sun, Philip Torr, Bo Zhao
Researchers from Peking University, the University of Oxford, and ETH Zurich introduced "REAL-FAKE," a principled distribution-matching framework for synthetic data generation. This approach dramatically reduces the performance gap between models trained on synthetic versus real data, achieving 70.9% Top1 accuracy on ImageNet1K with 1x synthetic data, and also enhances out-of-distribution generalization and privacy.
Key insights
- Framing synthetic data generation as a distribution matching problem for both data and class-conditional properties is crucial for generating effective training data.
- Combining MMD for latent space alignment with multi-modal conditioning and an informed prior significantly improves the distributional fidelity and utility of synthetic data.
- Synthetic data can not only close the performance gap to real data but also offer superior out-of-distribution generalization and robust privacy protection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.