Few-Shot vs Single-Shot vs Zero-Shot Learning

Few-shot, one-shot, and zero-shot learning refer to different paradigms of machine learning where models are trained with limited or no labeled examples.

1. Few-Shot Learning (FSL)

• The model learns a new task with only a few labeled examples (e.g., 5 or 10 samples per class).
• Often implemented using meta-learning, where the model is pre-trained on many tasks to generalize quickly to new ones.
• Example: A model trained on multiple languages learns a new language with just a few sentences.

2. One-Shot Learning

• A special case of few-shot learning where the model is trained with only one labeled example per class.
• Often used in face recognition (e.g., learning a person’s identity from a single photo).
• Techniques like Siamese Networks and Prototypical Networks are commonly used.

3. Zero-Shot Learning (ZSL)

• The model generalizes to new tasks without seeing any labeled examples.
• Relies on semantic relationships (e.g., word embeddings) to transfer knowledge.
• Example: A model trained on dog and cat images can classify a zebra using descriptions like “striped horse-like animal.”