Reproduce “Snorkel DryBell: A Case Study in Deploying Weak Supervision at Industrial Scale” on OpenResearch

About this paper

Stephen H. Bach, Daniel Rodriguez, Yintao Liu, Chong Luo, Haidong Shao, Cassandra Xia, Souvik Sen, Alexander Ratner and 5 more

Snorkel DryBell, a system developed by researchers from Google and universities, demonstrates how weak supervision can be deployed at industrial scale to effectively leverage diverse organizational knowledge for machine learning. The system allows training models with quality comparable to those derived from extensive hand-labeling, significantly reducing development time and cost.

Key insights

• Weak supervision can substantially reduce the need for manual data labeling, with weakly supervised models matching the performance of models trained with tens of thousands of hand-labeled examples.
• Leveraging existing, diverse organizational knowledge resources as weak supervision signals significantly improves model performance and allows for faster development cycles.
• The 'cross-feature transfer' approach effectively bridges the gap between rich, but impractical, training features and the compact, real-time features required for production serving.

Summary by alphaXiv.