Inferring Generative Model Structure with Static Analysis.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Training complex machine learning models requires substantial labeled data, a significant bottleneck.
• Weak supervision sources are often combined using generative models, but their structure is hard to learn without ground truth labels.

Purpose of the Study:

• To introduce Coral, a novel paradigm for inferring generative model structure by analyzing heuristic code.
• To reduce the amount of data required for learning model structure.

Main Methods:

• Coral employs static code analysis to infer generative model structure from programmatic weak supervision sources.
• Theoretical analysis proves Coral's sample complexity scales quasilinearly with heuristics and relations, outperforming exponential complexity.

Main Results:

• Coral matches or surpasses traditional structure learning methods by up to 3.81 F1 points.
• In radiology data labeling, Coral improved performance by 3.07 accuracy points over fully supervised models by modeling dependencies.

Conclusions:

• Coral offers an efficient method for learning generative model structure from code, overcoming limitations of ground truth data.
• The paradigm significantly enhances machine learning pipelines, particularly in data-scarce or weakly supervised scenarios.