Few-Shot & Zero-Shot Learning¶
Definition¶
Learning from minimal (few-shot) or no (zero-shot) task-specific examples — enabling eKYC systems to handle new document types, new attack types, or new demographics without extensive retraining.
Applications in eKYC¶
| Scenario | Approach | Example |
|---|---|---|
| New document type | Few-shot classification | New country ID with 5-10 sample images |
| New attack type | Zero-shot detection | Detect unseen mask type using learned spoof features |
| New language OCR | Few-shot adaptation | Recognize new script with minimal labeled text |
| Rare fraud pattern | Few-shot anomaly detection | Detect pattern seen only 2-3 times before |
Key Methods¶
| Method | How It Works |
|---|---|
| Prototypical Networks | Learn class prototypes from few examples, classify by nearest prototype |
| MAML | Meta-learn initialization that adapts quickly to new tasks |
| CLIP zero-shot | Use language-vision model to classify without task-specific training |
| Siamese Networks | Learn similarity function, compare new examples to references |
Key Takeaways¶
Summary
- Few-shot learning enables rapid adaptation to new document types and attack types
- CLIP-based zero-shot is emerging as powerful for document classification without training
- Meta-learning (MAML) enables quick adaptation to new domains with minimal data
- Critical for eKYC vendors supporting 200+ countries — can't collect large datasets for every document type