Active Learning¶
Definition¶
Active learning selects the most informative unlabeled examples for human annotation — maximizing model improvement per labeled sample. Critical for eKYC where labeling is expensive (especially liveness and forensics).
Query Strategies¶
| Strategy | How It Works | Best For |
|---|---|---|
| Uncertainty sampling | Select samples where model is least confident | General — simplest approach |
| Query-by-committee | Multiple models disagree on prediction | Ensemble models |
| Expected model change | Select samples that would most change model weights | Maximum learning per sample |
| Diversity sampling | Select diverse samples covering feature space | Avoiding redundant labels |
| Core-set | Select samples that best represent unlabeled data distribution | Large unlabeled pools |
Active Learning Loop for eKYC¶
graph TD
A[Unlabeled eKYC Data<br/>1M+ images] --> B[Model Predicts<br/>Score each sample]
B --> C[Query Strategy<br/>Select most informative 1000]
C --> D[Human Annotation<br/>Expert labels selected samples]
D --> E[Retrain Model<br/>Updated with new labels]
E --> F{Performance sufficient?}
F -->|No| B
F -->|Yes| G[Deploy Model]Key Takeaways¶
Summary
- Active learning reduces labeling cost by 3-10x — label only what matters most
- Uncertainty sampling is the simplest and often most effective strategy
- Especially valuable for liveness (expert annotation needed) and forensics (subtle ground truth)
- Combine with self-supervised pre-training for maximum data efficiency