Query Set

The query set contains unlabeled examples that the model has never seen during meta-training. These samples simulate real-world deployment where novel signal types must be classified without prior exposure.

• Each query sample belongs to one of the N classes defined in the support set
• No ground-truth labels are available to the model during inference
• Classification accuracy on the query set is the primary performance metric

In an N-way K-shot episode, the query set is sampled alongside the support set to form a complete classification task. A typical split uses 15 query examples per class.

• For a 5-way 1-shot task: 5 support samples, 75 query samples
• Query samples are drawn from the same classes as the support set
• The model must generalize from K examples to classify many more query points

Unlike inductive inference that classifies each query independently, transductive inference processes the entire query set as a batch. The model leverages the marginal distribution of unlabeled queries to refine decision boundaries.

• Mutual information between query samples improves accuracy
• Particularly effective when the support set is extremely small (1-shot)
• Requires access to all query samples simultaneously during inference

In metric-based meta-learners like Prototypical Networks, each query sample is classified by computing its distance to class prototypes derived from the support set. The query embedding is compared using a fixed similarity function.

• Common metrics: Euclidean distance, cosine similarity
• The query is assigned to the class of the nearest prototype
• Embedding space quality directly determines query classification accuracy

In optimization-based methods like MAML, the query set serves as the target for computing the outer-loop loss. The model takes a few gradient steps on the support set, then its performance is evaluated on the query set.

• Query loss drives meta-parameter updates during training
• The model learns initial weights that enable rapid adaptation
• Query set difficulty directly shapes the learned initialization

The query set provides the ground truth for measuring not just accuracy but calibration quality. Well-calibrated models produce confidence scores that match empirical accuracy on query samples.

• Expected Calibration Error (ECE) is computed on query predictions
• Critical for open-set rejection of unknown modulation types
• Bayesian meta-learners explicitly model query prediction uncertainty

In transductive inference, the entire query set is processed as a single batch rather than as isolated samples. The model leverages the marginal distribution of all unlabeled queries to refine decision boundaries.

• Exploits cluster assumptions: unlabeled samples often form distinct groups in the embedding space
• Improves accuracy by 5-15% over inductive methods when query sets contain balanced class distributions
• Particularly effective for signal constellation analysis where geometric structure emerges across multiple samples
• Requires careful handling of class imbalance within the query batch to avoid biased prototypes

The number of samples in the query set directly impacts both classification accuracy and the reliability of confidence estimates. Larger query sets provide more robust statistical evidence for evaluating model performance.

• Standard few-shot benchmarks use 15-25 query samples per class
• Insufficient query sizes lead to high-variance accuracy estimates and unreliable benchmarking
• Bayesian meta-learning methods produce well-calibrated uncertainties that improve with query set size
• In operational SIGINT systems, query set size reflects the dwell time available for signal collection

In prototypical networks, each query sample is classified by computing its distance to class prototypes derived exclusively from the support set. The query set never influences prototype computation during inductive inference.

• Prototype = mean of embedded support samples for each class
• Distance metric: typically Euclidean or cosine similarity in the learned embedding space
• Query samples are classified independently; no information flows between queries
• Softmax over distances produces a distribution over N classes for each query
• The embedding function must be trained episodically to generalize to novel modulation types

The composition of query sets in evaluation must reflect operational signal environments to provide meaningful performance estimates. Synthetic or curated query sets often overestimate real-world accuracy.

• Include channel impairments: fading, noise, frequency offset, and phase rotation
• Vary signal-to-noise ratio (SNR) across query samples to test robustness
• Introduce class imbalance to simulate real spectrum occupancy patterns
• Include out-of-distribution samples to measure false classification rates
• Cross-domain query sets (synthetic training, over-the-air testing) reveal true generalization capability

Relation networks replace fixed distance functions with a learned deep nonlinear comparator. Each query-support pair is concatenated and passed through a relation module that outputs a scalar similarity score.

• The relation module learns to compare IQ sample embeddings adaptively
• Query set size scales quadratically: N classes × K support × M queries = N×K×M comparisons
• Eliminates the need to choose between Euclidean or cosine distance
• Particularly effective when modulation classes have hierarchical similarities (e.g., QAM variants)
• Computationally more expensive than prototypical networks for large query sets

Query samples with low maximum softmax probability can be rejected as unknown rather than forced into a known class. This is critical for open-set spectrum monitoring where novel signals appear regularly.

• Set a confidence threshold τ on the predicted class probability
• Samples below τ are labeled as unknown modulation for analyst review
• Threshold calibration requires a held-out validation query set with known unknowns
• Balances the trade-off between coverage (samples classified) and precision (correct classifications)
• Essential for electronic warfare support where missing a novel threat signal is unacceptable