The Few-Shot Learning Problem
What you'll learn: Why few-shot learning is critical when you have very few labeled examples per class.
The Core Challenge
Imagine you're a botanist who needs to identify rare plant species, but you only have 1-5 photos of each species. Traditional deep learning would fail spectacularly here—it typically needs hundreds or thousands of labeled examples per class to learn effectively. This is the few-shot learning problem: how do we train models that can recognize new classes from just a handful of examples?
Why This Matters
In the real world, labeled data is often:
- Expensive: Medical imaging requires expert radiologists to label each scan
- Rare: Endangered species have few photographed instances
- Time-sensitive: New products need classification before large datasets accumulate
- Privacy-constrained: Personal data can't always be collected at scale
Standard supervised learning breaks down in these scenarios. You can't train a ResNet on 5 examples and expect it to generalize. The model would immediately overfit, memorizing those few samples without learning meaningful patterns.
The Setting
Few-shot learning is typically framed as N-way K-shot classification:
- N-way: Classify among N different classes
- K-shot: With only K labeled examples per class
For example, "5-way 1-shot" means distinguishing between 5 classes using just 1 example of each. This forces models to learn quickly from minimal data—much like humans do when shown a single photo of an unfamiliar animal.
Key Takeaway: Few-shot learning addresses the fundamental constraint that labeled data is often scarce or costly, requiring models that can generalize from just a handful of examples rather than thousands.