N-way K-shot

Each episode is a self-contained classification task. The model is given a support set of N classes with K examples each, and must classify a query set of unseen examples from those same N classes.

• N-way: The number of distinct classes in the episode (e.g., 5-way means 5 different modulation types)
• K-shot: The number of labeled examples per class (e.g., 1-shot means a single IQ sample per modulation)
• Query Set: Unlabeled examples the model must classify using only knowledge from the support set

The paradigm enforces a strict separation between base and novel classes to measure true generalization.

• Meta-Training: Episodes are sampled from a large set of base classes with abundant data. The model learns how to learn from small support sets.
• Meta-Testing: Episodes are sampled from entirely novel classes never seen during training. This evaluates the model's ability to adapt to new modulation schemes with only K examples.
• No class overlap is permitted between training and testing phases.

The values of N and K directly control the difficulty of the few-shot task and the model's practical utility.

• Higher N increases class confusion and demands a more discriminative embedding space
• Lower K (e.g., 1-shot) tests extreme data efficiency, critical for intercepting rare or fleeting signals
• Standard benchmarks: 5-way 1-shot and 5-way 5-shot are common in computer vision; 5-way 5-shot to 10-way 10-shot are typical for modulation recognition
• Performance ceiling: As K increases, performance typically saturates, revealing the model's inductive bias quality

The support set functions as the model's sole source of information about the novel classes at inference time. No weight updates occur.

• Metric-based methods (e.g., Prototypical Networks): Compute a prototype—the mean embedding—for each class from its K support examples
• Optimization-based methods (e.g., MAML): Use the support set to compute a task-specific parameter update via a few gradient steps
• Memory-based methods (e.g., Matching Networks): Attend directly over the support set embeddings to classify each query
• The support set is discarded after the episode; the model retains no persistent memory of the novel classes

The N-way K-shot paradigm directly addresses the challenge of identifying rare or emerging modulation formats in electronic warfare and spectrum monitoring.

• Scenario: A SIGINT system encounters a new adversary waveform with only 5 captured IQ samples (5-shot)
• Episode Design: N includes the novel modulation plus several known confuser modulations
• Feature Extraction: Raw IQ samples or cyclostationary features are embedded into a space where Euclidean distance corresponds to modulation similarity
• Outcome: The classifier discriminates the novel signal from known types without retraining, using only the K provided examples

N-way K-shot learning differs fundamentally from standard transfer learning and fine-tuning approaches.

• Transfer Learning: Pre-train on base classes, then fine-tune all or part of the network on a target dataset. Requires a non-trivial target dataset and multiple gradient updates.
• N-way K-shot: No fine-tuning occurs at test time. The model ingests the support set and immediately classifies queries in a single forward pass.
• Key Advantage: Enables instantaneous adaptation to new modulation types without the computational cost or data volume required for fine-tuning.
• Complementary Use: A model pre-trained via transfer learning can serve as the backbone feature extractor for a meta-learning algorithm.