query set
**The query set** in few-shot learning contains the **test examples** used to evaluate model performance after the model has been given the labeled support set. It serves as the episode's internal test set — measuring how well the model learned from the few provided examples.
**Role in a Few-Shot Episode**
- **Support Set**: The K labeled examples per class that the model uses to "learn" the task. Analogous to training data.
- **Query Set**: Additional examples from the **same N classes** that the model must classify. Analogous to test data.
- **Evaluation**: Predictions on query examples are compared against true labels to compute **episode accuracy**.
**Example: 5-Way 5-Shot Episode**
| Component | Content | Purpose |
|-----------|---------|--------|
| Support Set | 5 classes × 5 examples = 25 labeled images | "Learn" these classes |
| Query Set | 5 classes × 15 examples = 75 unlabeled images | Classify using support knowledge |
| Output | Accuracy on query predictions | Evaluate few-shot performance |
**Query Set in Meta-Training vs. Meta-Testing**
- **During Meta-Training**: Query set loss drives **model parameter updates**. The model learns to perform well on queries after seeing only the support set. Gradients flow through both support processing and query classification.
- **During Meta-Testing**: Query set provides the **final evaluation metric**. No parameter updates — this is the true test of few-shot generalization.
**Key Properties**
- **Disjoint from Support**: Query and support sets must be **completely non-overlapping** — the same example cannot appear in both. This ensures unbiased evaluation of generalization.
- **Same Classes**: Query examples come from the **same N classes** as the support set — the model must classify queries into one of the N support classes.
- **Typical Size**: Usually 10–20 query examples per class, though this varies by benchmark. More queries provide more stable accuracy estimates.
**Query Set in Different Methods**
- **Prototypical Networks**: Compute class prototypes from support set, then classify each query by **nearest prototype** using Euclidean distance.
- **MAML**: Adapt model parameters using support set gradients, then evaluate adapted model on queries.
- **Matching Networks**: Each query attends to all support examples via learned similarity, producing a weighted classification.
**Inductive vs. Transductive Processing**
- **Inductive**: Each query example is classified **independently** — no information flows between query examples.
- **Transductive**: All query examples are processed **jointly** — the model can use the distribution and structure of the query set to improve predictions. Typically improves accuracy by 2–5%.
**Impact on Evaluation**
- **Episode Accuracy**: Fraction of correctly classified query examples within a single episode.
- **Reported Accuracy**: Average accuracy across hundreds or thousands of test episodes, typically reported with **95% confidence intervals**.
- **Query Count Sensitivity**: More query examples per class provide more stable accuracy estimates but increase computational cost per episode.
The query set is the **measurement instrument** of few-shot learning — it reveals how effectively the model has learned to generalize from the few support examples to new instances of the same classes.