Support Set

The support set defines the classification space for a single episode or inference task. In an N-way K-shot configuration, it contains N unique classes with K labeled examples each. Unlike a traditional training set, the support set is provided at runtime, conditioning a pre-trained model to recognize classes it may have never seen before. This mirrors real-world scenarios like enrolling a new IoT device with only a few RF fingerprint captures.

In metric-based architectures like Prototypical Networks, the support set is used to compute a class prototype—the mean embedding vector of all support examples belonging to that class. Query samples are then classified by finding the nearest prototype in the embedding space. The support set's role is to establish these decision boundaries in real-time, making the quality and representativeness of each support example critical to accuracy.

For Siamese Networks trained with Triplet Loss, the support set provides the anchor and positive examples that define class membership. The model learns to pull embeddings of same-class pairs together while pushing different-class pairs apart by a defined margin. The support set must contain both positive pairs (same device, different captures) and implicit negative pairs (different devices) to establish discriminative boundaries.

In optimization-based methods like Model-Agnostic Meta-Learning (MAML), the support set is used to compute a task-specific gradient update. The model performs a small number of inner-loop gradient descent steps on the support set loss, rapidly adapting its parameters from a meta-learned initialization. The support set must be sufficiently informative to guide this adaptation without causing overfitting to the few available examples.

During episode-based training, support sets are repeatedly sampled from a larger dataset to simulate few-shot tasks. Each episode randomly selects N classes and K examples per class, training the model to generalize across diverse task distributions. This sampling strategy must match the deployment scenario—for RF fingerprinting, episodes might sample across different transmitter models, environmental conditions, or signal-to-noise ratios to ensure robustness.

In a deployed few-shot device authentication system, the support set functions as the enrollment database. When a new IoT transmitter is onboarded, a small number of RF captures form its support set, creating a unique fingerprint template. The model compares future transmissions against this support set using cosine similarity or Euclidean distance. The enrollment process must capture representative samples across expected operating conditions to minimize False Rejection Rates.

The unlabeled set of examples used to evaluate the model's performance on a specific few-shot task. After the model is conditioned on the support set, it must classify or predict the labels for the query set. The loss is computed on query set predictions, driving the meta-learning optimization. In an N-way K-shot task, the query set typically contains multiple samples per class distinct from the support examples.

A metric-based few-shot learning architecture that computes a prototype for each class by averaging the embeddings of its support set examples. Classification of a query sample is performed by finding the nearest prototype in the embedding space using Euclidean distance. This non-parametric approach requires no fine-tuning at inference time, making it computationally efficient for rapid device enrollment scenarios.

A meta-learning training strategy where each iteration simulates a complete few-shot task. An episode is constructed by randomly sampling a support set and a query set from the training data. The model is trained across thousands of episodes to learn a generalizable initialization that can rapidly adapt to new classes defined by a novel support set. This mimics the test-time conditions of few-shot device enrollment.

The standard problem formulation for few-shot learning tasks. N-way specifies the number of distinct classes in the episode, and K-shot defines the number of labeled examples per class in the support set. A 5-way 1-shot task requires the model to distinguish between 5 device classes using only a single enrollment sample per device. Higher K values generally improve accuracy by providing richer class representations.

A lower-dimensional, continuous vector space where semantically similar data points are mapped close together. The support set and query set are both projected into this space by an encoder network. The quality of the embedding space—its ability to form tight, well-separated clusters—directly determines few-shot classification accuracy. Distance metrics like cosine similarity or Euclidean distance operate within this space.

An optimization-based meta-learning algorithm that learns an initial model parameterization that can rapidly adapt to new tasks. During meta-training, the model performs a few gradient steps on the support set and is evaluated on the query set. The outer loop optimizes the initial parameters so that a small number of inner-loop updates on any support set yields strong generalization. This is ideal for quick device enrollment with minimal samples.