Weakly Supervised Learning

Learning with Weak Labels

Use incomplete or noisy labels.

Types

Incomplete: partially labeled. Inexact: coarse labels. Noisy: incorrect labels.

Approaches

Label propagation: propagate to similar examples. Snorkel: programmatic labeling. Self-training: high-confidence predictions.

Data Programming

Generate labeling functions. Combine with Snorkel. Train on probabilistic labels.

Key Takeaways

1. Weak supervision: incomplete, inexact, noisy
2. Label functions for programmatic labeling
3. Snorkel provides framework