The Meta-Training vs Meta-Testing Split
What you'll learn: How few-shot learning separates base classes for training from novel classes for evaluation to test true generalization.
The Core Idea
In traditional machine learning, you split data into train and test sets. In meta-learning (few-shot learning), you split classes themselves into two groups:
- Meta-Training (Base Classes): A set of classes your model learns how to learn from during training
- Meta-Testing (Novel Classes): Completely different classes held out for final evaluation
Think of it like learning to learn languages. During meta-training, you practice learning Romance languages (French, Spanish, Italian). During meta-testing, you're given a Germanic language (German) you've never seen. The goal isn't to master French—it's to become good at the process of learning languages quickly, so you can tackle German with just a few examples.
Why This Split Matters
If you trained and tested on the same classes, you'd just memorize those specific categories. The meta-training/meta-testing split forces your model to generalize to completely new concepts it has never encountered.
For example:
- Meta-Training: Learn from 64 base animal classes (cats, dogs, birds...)
- Meta-Testing: Evaluate on 16 novel classes (lemurs, platypuses...) the model has never seen
During meta-training, you create episodes from base classes. During meta-testing, episodes come from novel classes with their own support and query sets.
The Test of True Few-Shot Ability
This split ensures your model doesn't just memorize specific classes—it learns a transferable strategy for recognizing new categories from minimal examples.
Key Takeaway: Meta-learning splits classes, not just data—base classes train the meta-learner, while novel classes test whether it can truly generalize to unseen categories.