Semi-Supervised Learning

Using Unlabeled Data

Leverage unlabeled examples with labeled.

Self-Training

Train on labeled, predict on unlabeled. Add high-confidence predictions to labeled set. Repeat.

Consistency Regularization

Label-preserving transformations. Ensure model gives same prediction on augmented versions. Dropout as noise.

Approaches

Pseudo-labeling, entropy regularization. MixMatch, UDA, FixMatch.

Key Takeaways

1. Use unlabeled data with labeled
2. Consistency regularization effective
3. State-of-art: FixMatch