N-way K-shot is the standard notation describing the structure of few-shot learning tasks, where N specifies the number of classes and K specifies the number of labeled examples per class.

Notation Breakdown

• N-way: The classification task has N classes to distinguish between. Higher N means more classes and harder discrimination.
• K-shot: Each class has K labeled support examples available. Higher K provides more information per class.
• Example: 5-way 5-shot = 5 classes, 5 examples each = 25 total support examples.

Common Configurations

How Difficulty Scales

• Increasing N (more classes): Harder — more classes to distinguish means higher chance of confusion between similar categories. Random baseline accuracy = $1/N$.
• Increasing K (more examples): Easier — more examples provide better class representations, capture intra-class variation, and reduce noise from atypical examples.
• 5-way 1-shot vs. 5-way 5-shot: Typical accuracy gap of 10–20 percentage points — more examples significantly help.

Episode Structure

• Support Set: $N \times K$ labeled examples total.
• Query Set: $N \times Q$ examples to classify (Q typically 10–20 per class).
• Total Examples Per Episode: $N \times (K + Q)$.

Benchmark Results (miniImageNet)

• 5-way 1-shot: State-of-the-art ~65–75% accuracy.
• 5-way 5-shot: State-of-the-art ~80–88% accuracy.
• Random Baseline: 20% for 5-way (1/N).

Variations

• Variable-Way Variable-Shot: N and K vary across episodes (used in Meta-Dataset). More realistic — real-world scenarios rarely have exactly 5 classes with exactly 5 examples each.
• Class-Imbalanced: Different classes have different numbers of examples within an episode — some classes have 2 examples, others have 10.
• Transductive N-way K-shot: The model can jointly reason about all query examples, exploiting test-set structure for better predictions.
• Generalized Few-Shot: Test episodes include both seen base classes AND unseen novel classes — the model must handle both simultaneously.

Reporting Standards

• Average Accuracy: Mean accuracy over 600–10,000 randomly sampled test episodes.
• Confidence Interval: 95% CI reported — typically ±0.2–0.5% for well-sampled evaluations.
• Reproducibility: Report random seed, episode sampling strategy, and exact train/val/test class splits.

The N-way K-shot framework provides a standardized language for comparing few-shot learning methods — ensuring fair comparison by specifying exactly how much data the model has access to for each task.