Snorkel DryBell

Snorkel DryBell leverages weak supervision to programmatically generate and manage labeled training datasets from diverse organizational knowledge sources for scalable classifier development.

Key Features:

• Weak Supervision Utilization: Generates labels by applying weak supervision from diverse organizational knowledge sources, reducing the need for manual annotation.
• Flexible Ingestion System: Uses a template-based ingestion system to accommodate a wide range of organizational knowledge formats into the labeling workflow.
• Cross-Feature Production Serving: Supports cross-feature production serving to integrate and utilize multiple features across different datasets for model serving.
• Scalable Execution: Executes at scale without sampling, able to process millions of data points in tens of minutes.

Scientific Applications:

• Google case studies: Applied to three classification tasks at Google, producing classifiers with quality comparable to models trained with tens of thousands of hand-labeled examples and converting non-servable organizational resources into servable models with an average performance improvement of 52%.

Methodology:

Programmatic construction and management of training datasets via heuristic labeling functions; integration of diverse organizational knowledge sources; scalable execution processing millions of data points without sampling.