Embedding Space

The embedding function must map all variations of the same modulation type—regardless of noise, frequency offset, or timing jitter—to a tight, dense cluster. Prototypical Networks rely on this property, as the class prototype is simply the mean vector of the support set. If the cluster is diffuse, the prototype becomes a poor representative, leading to misclassification. This is enforced during training by minimizing the distance between an embedded sample and its corresponding class centroid.

Clusters representing distinct modulation classes must be pushed far apart in the vector space to create clear decision boundaries. A large inter-class margin ensures that a query sample from one class is not confused with the prototype of another. This is typically achieved through loss functions like the triplet loss or contrastive loss, which explicitly penalize overlapping distributions. For signals with similar constellations, such as QPSK and OQPSK, this separability is critical.

The latent space should be smooth, meaning that small perturbations to the input signal result in small, proportional displacements in the embedding. This property, often enforced by Manifold Mixup or explicit regularization, ensures that the space interpolates meaningfully. For example, a signal transitioning from low to high SNR should follow a continuous trajectory in the embedding space, allowing the classifier to generalize to unseen noise levels without abrupt jumps in the representation.

An effective embedding space discards irrelevant information, known as nuisance variables, such as carrier phase rotation, minor Doppler shifts, or channel gain. The encoder must learn to map signals with identical modulation but different physical-layer distortions to the same point. This is often achieved through aggressive data augmentation during training—applying random phase rotations and amplitude scaling—to force the network to become invariant to these transformations.

The embedding network must learn a hierarchy of features that capture the defining characteristics of modulation schemes. Lower layers might detect basic spectral shapes, while higher layers encode complex cyclostationary signatures and higher-order cumulants. For few-shot learning, the top-level embedding must capture features that are not just discriminative for the base training classes but are also generalizable to novel, unseen modulation formats. This requires the feature extractor to avoid overfitting to base-class idiosyncrasies.

The choice of distance function in the embedding space is critical. While Cosine Similarity is popular for its bounded range and focus on angular separation, Euclidean distance is sensitive to magnitude. An effective space is often structured so that a simple linear distance correlates strongly with semantic dissimilarity. More advanced approaches, like Relation Networks, learn a non-linear distance metric on top of the embeddings, allowing for more complex and adaptive decision boundaries than a fixed mathematical formula.

A branch of machine learning that aims to learn a distance function over objects. The objective is to map inputs into an embedding space where semantically similar samples are pulled together and dissimilar ones are pushed apart. This is the foundational principle enabling distance-based comparison in few-shot classifiers.

• Triplet Loss: Uses anchor, positive, and negative samples to enforce relative distances.
• Contrastive Loss: Minimizes distance for similar pairs and maximizes it for dissimilar pairs beyond a margin.
• Directly optimizes the geometry of the embedding space for downstream tasks like nearest-neighbor classification.

A measure of similarity between two non-zero vectors that computes the cosine of the angle between them. In an embedding space, it is often preferred over Euclidean distance because it normalizes for magnitude, focusing purely on the directional alignment of feature vectors.

• Range: Yields values from -1 (opposite) to 1 (identical).
• Formula: cos(θ) = (A · B) / (||A|| ||B||)
• Commonly used in Prototypical Networks to compare query embeddings against class prototypes.

A metric-based meta-learning algorithm that classifies query samples by computing their distance to a prototype representation in the learned embedding space. A prototype is simply the mean vector of the embedded support samples for a given class.

• Inductive Bias: Assumes a single cluster per class centered on the mean.
• Classification: Assigns the query to the class with the nearest prototype using Euclidean distance.
• The quality of the embedding space directly determines the separability of these prototypes.

A metric-based few-shot learning architecture that learns a deep nonlinear distance metric to compare query and support samples. Instead of using a fixed similarity function like cosine or Euclidean distance, it trains a relation module to score the relationship between concatenated feature embeddings.

• Flexibility: Can learn complex, non-linear decision boundaries in the embedding space.
• Architecture: Consists of an embedding module and a relation module.
• Outputs a scalar relation score between 0 and 1 for each query-support pair.

A regularization and data augmentation method that performs linear interpolations on learned hidden representations rather than raw inputs. In the context of an embedding space, it encourages smoother decision boundaries by training on convex combinations of feature vectors.

• Process: Mixes hidden states and corresponding labels from random layers.
• Effect: Flattens the representation manifold, improving generalization.
• Helps few-shot classifiers learn a more continuous and structured embedding space with fewer spurious decision boundaries.

A data augmentation strategy performed directly in the embedding space where a generator network synthesizes additional, plausible feature vectors for underrepresented classes. This is critical for few-shot learning where the support set is too small to define a robust class boundary.

• Mechanism: Trains a generator on base class feature distributions.
• Application: Generates synthetic embeddings for novel classes to augment the support set.
• Improves the estimation of class prototypes and decision boundaries without needing raw signal generation.