Snorkel MeTaL: Weak Supervision for Multi-Task Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Real-world machine learning often involves complex, multi-level sub-tasks and requires extensive labeled data.
• Existing methods struggle with problems requiring hierarchical task structures and large labeled datasets.

Purpose of the Study:

• To introduce Snorkel MeTaL, an end-to-end system for multi-task learning using weak supervision.
• To enable users to define hierarchically related sub-tasks and provide weak supervision via labeling functions.
• To automatically train hierarchical multi-task networks by learning from these labeling functions.

Main Methods:

• Developed Snorkel MeTaL, an end-to-end system for multi-task learning.
• Leveraged weak supervision at multiple levels of granularity through user-defined labeling functions.
• Learned a re-weighted model of labeling functions to train a hierarchical multi-task network.

Main Results:

• Achieved average accuracy gains of 11.2 points over a baseline supervised approach.
• Demonstrated an improvement of 9.5 accuracy points over user-provided labeling functions.
• Successfully applied MeTaL to radiology report triage and fine-grained news classification tasks.

Conclusions:

• Snorkel MeTaL effectively addresses challenges in multi-task learning with hierarchical structures and limited labeled data.
• The system significantly improves model performance by intelligently combining weak supervision signals.
• Weak supervision at multiple granularity levels is a viable strategy for enhancing machine learning model accuracy.