Prototypical Networks

Each class is represented by a single prototype vector, computed as the mean of the embedded support examples belonging to that class. The embedding function, typically a deep convolutional or recurrent neural network, maps raw input data into a space where Euclidean distance reflects semantic similarity. This averaging operation provides a simple yet powerful inductive bias: the prototype sits at the center of the class's support set distribution, making it robust to individual sample noise.

Classification of a query point is performed by computing the negative squared Euclidean distance to each class prototype, followed by a SoftMax function to produce a probability distribution over classes:

• The model outputs p(y=k|x) ∝ exp(-d(f(x), c_k)) where c_k is the prototype for class k
• This formulation is equivalent to a Gaussian mixture model with equal isotropic covariance in the embedding space
• The distance metric can be extended to Mahalanobis distance for more expressive class-conditional structures

Training is structured as episodes (or mini-batches) that simulate few-shot test scenarios. Each episode samples:

• A subset of classes from the training set
• A small support set (e.g., 1 or 5 examples per class) to compute prototypes
• A query set of held-out examples from the same classes for loss computation

This meta-learning strategy ensures the model learns to generalize from limited data rather than memorizing individual training instances.

The squared Euclidean distance can be generalized to any Bregman divergence, including KL-divergence and Itakura-Saito distance. The prototype computation remains the arithmetic mean for any regular Bregman divergence, making the framework mathematically principled. This property allows Prototypical Networks to be adapted to different data modalities—for example, using cosine distance for text embeddings or specialized divergences for spectral data in RF fingerprinting applications.

Prototypical Networks naturally extend to zero-shot learning by replacing data-driven prototypes with semantic vectors derived from class metadata or attribute descriptions. Instead of averaging support embeddings, the prototype for an unseen class is generated from a pre-trained language model or attribute vector, enabling classification of classes for which no training examples exist. This is particularly valuable in open set emitter recognition where new device types must be identified from technical specifications alone.

The distance to the nearest prototype serves as a natural open set rejection score. Query samples far from all known prototypes in the learned metric space can be flagged as unknown or out-of-distribution. This is enhanced by:

• Gaussian embedding variants that model each class as a distribution rather than a point, capturing within-class variance
• Threshold calibration using Extreme Value Theory on distance distributions from validation data
• Integration with epistemic uncertainty estimates for robust novelty detection in dynamic RF environments

The foundational framework that underpins Prototypical Networks. Rather than learning a direct mapping from input to class label, the model learns an embedding function that places semantically similar samples close together and dissimilar samples far apart in a vector space.

• Euclidean distance is used by Prototypical Networks to measure proximity to class prototypes
• Cosine similarity is an alternative metric used by Matching Networks
• The choice of distance metric directly impacts the geometry of the learned embedding space
• Bregman divergences, such as squared Euclidean distance, are theoretically optimal for prototype-based classification

A precursor architecture to Prototypical Networks that also performs few-shot classification via learned similarity comparisons. Matching Networks use attention mechanisms over a support set to classify query samples.

• Employs a full-context embedding where support set samples are encoded in the context of one another using bidirectional LSTMs
• Classification is performed via a cosine similarity weighted attention over support set labels
• Prototypical Networks simplify this by collapsing support examples into a single mean prototype, achieving comparable or superior performance with less computational overhead

A family of loss functions including ArcFace and CosFace that enforce discriminative constraints on the angular space of feature embeddings. These losses are often used to pre-train or fine-tune the embedding network used in Prototypical Networks.

• Adds an additive angular margin penalty between class prototypes and samples to maximize inter-class separation
• Produces highly compact intra-class clusters, which directly improves prototype quality
• Originally developed for face recognition, but highly effective for RF emitter identification where signal differences can be subtle
• Combining angular margin loss with episodic training creates embeddings that generalize better to unseen emitter classes

The classification paradigm where a model must simultaneously identify known classes and reject unknown classes not seen during training. Prototypical Networks naturally extend to this setting through distance-based rejection.

• Unknown emitters will map to regions of the embedding space far from any known prototype
• A distance threshold can be calibrated on a validation set to reject queries whose distance to the nearest prototype exceeds a learned boundary
• This is critical for spectrum surveillance where new transmitters appear continuously
• Prototypical Networks avoid the open space risk inherent in SoftMax-based classifiers that partition the entire space into known classes

The meta-learning training paradigm used to prepare Prototypical Networks for few-shot inference. The model is trained on a distribution of mini-datasets rather than individual samples.

• Each episode consists of a small support set (e.g., 5 examples per class) and a query set for evaluation
• Classes are randomly sampled from the training set to simulate the few-shot scenario encountered at test time
• The model learns to extract features that are useful for comparing and contrasting any set of classes, not just those seen during training
• This mimics the operational reality of identifying new emitters from only a handful of intercepted transmissions

A low-dimensional vector representation of high-dimensional input data where semantic similarity is preserved as geometric proximity. The quality of the embedding function is the single most important factor in Prototypical Network performance.

• For RF signals, the embedding network typically consists of 1D convolutional layers or residual blocks processing IQ samples
• The embedding dimension is a critical hyperparameter: too small loses discriminative power, too large increases the risk of overfitting with few support examples
• Well-trained embeddings exhibit clusterable manifolds where each emitter class forms a tight, separable group
• The prototype is simply the arithmetic mean of support embeddings for a given class