The Meta-Learning Problem Formulation
What you'll learn: How meta-learning structures the problem of "learning to learn" using support/query sets and episodic training.
The Core Problem Structure
Meta-learning formalizes the challenge of learning from limited data by treating tasks themselves as training examples. Instead of learning a single model for one problem, we learn a model that can quickly adapt to new tasks with minimal examples.
Support and Query Sets
Each task is split into two parts:
- Support set: The few labeled examples (e.g., 5 images per class) used to adapt or condition your model for this specific task
- Query set: Held-out examples from the same task used to evaluate how well the model adapted
Think of it like a mini train/test split within each task. The support set teaches you "what this task is about," and the query set tests if you learned it.
Meta-Training vs Meta-Testing
- Meta-training: Your model sees many diverse tasks. For each task, it uses the support set to make predictions on the query set. Gradients flow through this entire process, teaching the model how to adapt quickly
- Meta-testing: Completely new tasks (never seen during meta-training) evaluate whether your model truly learned to learn, not just memorized task patterns
The Episodic Training Paradigm
Meta-learning uses episodes (or tasks) as the fundamental training unit. Each episode samples a random N-way K-shot problem, provides a support set, and tests on a query set. This mimics the meta-testing conditions during training—a crucial principle.
Key Takeaway: Meta-learning reformulates supervised learning by treating tasks as examples, splitting each into support/query sets, and using episodic training to learn adaptation strategies rather than fixed predictions.