Prototypical Networks

The foundational discipline that enables Prototypical Networks to function. Metric learning algorithms learn a distance function that maps inputs into an embedding space where semantically similar samples cluster together. In the context of few-shot modulation recognition, this means learning to project IQ samples such that BPSK signals are geometrically close to each other but far from QAM64 signals. The network is trained using a contrastive or triplet loss to minimize intra-class distance while maximizing inter-class separation, creating a structured space where a simple Euclidean distance to a class prototype becomes a highly effective classifier.

A preferred distance metric for comparing query embeddings to class prototypes in high-dimensional spaces. Unlike Euclidean distance, cosine similarity measures the angle between two vectors, normalizing for magnitude. This is critical in signal classification where the power of a received transmission can vary dramatically due to channel effects, but the underlying modulation structure—captured by the vector's direction—remains consistent. The similarity is computed as the dot product of L2-normalized vectors, producing a score between -1 and 1. Prototypical Networks often use cosine similarity over Euclidean distance to achieve power-invariant classification of radio frequency signals.

The meta-training strategy that explicitly simulates low-data test conditions to optimize Prototypical Networks. Training is structured into episodes, each a mini-classification problem with a small support set and query set. An episode might task the model with distinguishing 5 modulation types (5-way) using only 1 example each (1-shot). By solving thousands of these artificial few-shot tasks during training, the network learns a generalizable embedding function rather than memorizing specific training classes. This aligns the training objective precisely with the deployment scenario: recognizing rare or emerging signal types from very limited labeled captures.

An evolution of Prototypical Networks that replaces the fixed distance function with a learnable nonlinear comparator. Instead of computing Euclidean or cosine distance, a Relation Network concatenates the feature maps of a query sample and a class prototype, then passes them through a neural 'relation module' to output a similarity score. This allows the model to learn complex, task-specific distance metrics that can capture intricate interactions between signal features. For modulation classification, this means the network can learn to compare not just the mean of support embeddings but also higher-order statistical relationships that a simple distance metric might miss.

A reasoning mode that significantly boosts Prototypical Network accuracy by leveraging the unlabeled query set during classification. Standard inductive inference classifies each query independently. Transductive inference considers the entire query batch jointly, using the marginal distribution of queries to refine class prototypes. This is implemented via iterative k-means clustering or label propagation on the query embeddings. In spectrum monitoring, where a burst of unlabeled signals arrives simultaneously, transductive inference exploits the fact that queries from the same modulation cluster together, sharpening decision boundaries without requiring additional labeled data.

A statistical augmentation technique that enhances Prototypical Networks by generating high-quality synthetic feature vectors for novel classes. It first estimates the feature distribution of base classes, then calibrates these statistics using the limited support samples of a new class to create a more robust prototype. This combats the primary weakness of Prototypical Networks: when the support set is extremely small (e.g., 1-shot), the empirical mean is a noisy estimate of the true class centroid. Distribution calibration smooths this estimate, leading to tighter class boundaries and improved generalization for rare modulation types.