Few-Shot Learning for Rare Defects is the application of ML techniques that can learn to recognize new defect types from just a few (1-10) labeled examples — critical for semiconductor manufacturing where new defect types emerge with process changes and collecting large labeled datasets is impractical.

Key Approaches

• Metric Learning: Learn an embedding space where similar defects cluster together (Siamese networks, prototypical networks).
• Meta-Learning: Train a model to learn quickly from few examples (MAML, Reptile).
• Data Augmentation: Generate synthetic variations of the few available examples.
• Foundation Models: Use large pre-trained vision models (CLIP, DINO) as feature extractors for few-shot classification.

Why It Matters

• New Defect Types: Every process change can introduce novel defect types with initially very few examples.
• Fast Deployment: Deploy a new defect classifier with just 5-10 labeled examples instead of hundreds.
• Continuous Learning: Incrementally add new defect classes without retraining the entire model.

Few-Shot Learning is learning defects from a handful of examples — enabling rapid deployment of classifiers for novel defect types with minimal labeling effort.