Active Learning

Data-Efficient Learning

Select most informative data to label.

Query Strategies

Uncertainty: query least confident. Diversity: query representative samples. Expected model change: query would change model most.

Pool-Based

Label small initial set. Train model. Query from unlabeled pool. Repeat.

Applications

Labeling expensive: medical, scientific. Reduces labeling cost significantly.

Key Takeaways

1. Query informative samples
2. Uncertainty, diversity, change strategies
3. Reduces labeling cost