Support Set

The support set explicitly defines the novel classes and the decision boundaries for a single few-shot episode. It is the model's sole source of supervised information at inference time.

• Contains K labeled examples for each of the N classes to be recognized
• Defines the N-way K-shot classification problem
• Without it, the model has no basis for discriminating between unseen classes

Each support set is a mini-dataset constructed to simulate a low-data deployment scenario. The model processes an entire episode—support set plus query set—as a single atomic task.

• Drawn from classes the model has never seen during meta-training
• Paired with a corresponding query set of unlabeled examples
• Forces the model to learn how to learn rather than memorizing specific classes

In metric-based meta-learners like Prototypical Networks, the support set is embedded into a vector space where class prototypes are computed as the mean of embedded samples.

• Each support example is mapped through a learned embedding function
• A prototype c_k = (1/|S_k|) Σ f_φ(x_i) is computed per class
• Query samples are classified by nearest-prototype distance (Euclidean or cosine)

In optimization-based methods like MAML, the support set is used to compute task-specific parameter updates through one or a few gradient steps.

• The model starts from meta-learned initialization parameters θ
• Computes loss on the support set: L_{T_i}(f_θ)
• Produces adapted parameters: θ'_i = θ - α ∇θ L{T_i}(f_θ)
• The query set then evaluates the adapted model's generalization

The support set's small size creates a high-variance learning problem. With only 1-5 examples per class, every sample carries disproportionate influence.

• K=1 (one-shot): Maximum difficulty, no intra-class variance visible
• K=5 (five-shot): Provides a rough estimate of class distribution
• Imbalanced or noisy support samples can catastrophically skew class prototypes
• Effective models must extract maximal information from minimal data

The distribution of support set samples relative to query samples critically impacts performance. A mismatch between the two—known as a domain shift—degrades accuracy.

• Support and query sets should ideally be drawn from the same distribution
• In cross-domain few-shot learning, they originate from different domains (e.g., synthetic vs. over-the-air signals)
• Models must learn representations robust to this distributional gap

The standard episodic training paradigm that defines the structure of a few-shot task. A model must discriminate between N novel classes given only K labeled examples per class in the support set. For modulation recognition, a 5-way 1-shot task means identifying 5 different modulation schemes with only a single IQ sample per scheme. This paradigm explicitly trains models to generalize from minimal data.

The unlabeled examples in a few-shot episode that the model must classify by leveraging knowledge derived exclusively from the corresponding support set. In a modulation classification task, the query set contains unseen signal captures of the same N classes defined by the support set. Performance is measured by the accuracy of predictions on this set, making it the primary evaluation target for meta-learned classifiers.

A meta-training strategy that structures learning into a series of mini-datasets called episodes. Each episode simulates a low-data test scenario by sampling a support set and query set from a pool of base classes. This process explicitly optimizes the model for rapid adaptation rather than static classification. For signal intelligence applications, episodes might alternate between synthetic and over-the-air captures to improve cross-domain robustness.

A learned, lower-dimensional vector representation where semantically similar inputs map to nearby points. For few-shot modulation classification, raw IQ samples are projected into an embedding space where signals of the same modulation type cluster tightly together. This space enables distance-based comparison for metric-based classifiers. The quality of this space—its intra-class compactness and inter-class separation—directly determines few-shot accuracy.