Distribution Calibration

The core mechanism of distribution calibration involves transferring statistical knowledge from base classes with abundant data to novel classes with limited examples. The process first computes the mean and covariance of feature vectors for each base class. It then assumes that the distribution shift between base and novel classes follows a predictable pattern, allowing the system to calibrate the novel class distribution by applying learned transformations to the base class statistics. This enables the generation of high-quality synthetic feature vectors that accurately represent the novel class's true distribution in the embedding space.

A critical preprocessing step that applies a power transformation to the feature vectors to make their distribution more Gaussian-like. This transformation is essential because many metric-based few-shot classifiers assume that features are normally distributed in the embedding space. The transformation parameter is learned from the base class data and then applied to novel class features, ensuring that the subsequent statistical calibration steps operate on properly normalized distributions. This improves the quality of synthetic samples generated for rare modulation types.

Once the novel class distribution is estimated, the system generates synthetic feature vectors by sampling from the calibrated Gaussian distribution. The sampling process uses the calibrated mean and a covariance matrix that combines the base class covariance with a learned calibration offset. This approach produces diverse yet realistic feature vectors that expand the support set for novel modulation types. The generated samples are then used alongside the original few-shot examples to train the final classifier, effectively increasing the number of training examples without requiring additional real-world signal captures.

The calibration set is built by selecting base classes with sufficient examples to reliably estimate their feature distributions. For each selected base class, the system computes statistics from multiple randomly sampled subsets, creating a robust estimate of the mean and covariance variability across different data configurations. This calibration set serves as the foundation for learning how distributions shift from base to novel classes. The quality of this set directly impacts the accuracy of the generated synthetic features for rare modulation schemes.

Distribution calibration is typically deployed as a preprocessing module that enhances existing metric-based classifiers like Prototypical Networks. After generating synthetic features for novel classes, the augmented support set is used to compute more accurate class prototypes. This integration significantly boosts classification accuracy in extreme low-data regimes, such as 1-shot or 5-shot scenarios. The technique is model-agnostic and can be applied to any few-shot learning framework that relies on feature-space comparisons for classification.

A key advantage of distribution calibration is its robustness to domain shifts between training and testing environments. By calibrating distributions rather than directly transferring features, the method remains effective even when novel classes come from different signal-to-noise ratio conditions or channel impairments than the base classes. This makes it particularly valuable for real-world RF deployments where over-the-air captures may differ significantly from the lab-generated training data used for base class learning.

A data augmentation strategy operating in the learned embedding space rather than the raw signal domain. A generator network synthesizes plausible feature vectors for underrepresented classes by modeling the intra-class variance of base categories.

• Directly addresses class imbalance in the feature space
• Generates vectors statistically similar to real support samples
• Often paired with a discriminator to ensure realism
• Key enabler for distribution calibration pipelines

A metric-based meta-learning algorithm that computes a prototype—the mean embedding vector—for each class from the support set. Query samples are classified by finding the nearest prototype using Euclidean distance.

• Assumes a unimodal, spherical distribution per class
• Performance degrades when class distributions are skewed
• Distribution calibration directly addresses this limitation
• Forms the baseline that calibration-enhanced methods improve upon

A learned, lower-dimensional vector representation where semantically similar inputs map to nearby points. The quality of this space determines few-shot performance.

• Trained via episodic learning on abundant base classes
• Must produce discriminative and transferable features
• Distribution calibration operates entirely within this space
• Dimensionality typically ranges from 64 to 512 for RF tasks

The standard episodic evaluation paradigm where a model must discriminate between N novel classes given only K labeled examples per class in the support set.

• Common benchmarks: 5-way 1-shot, 5-way 5-shot
• Each episode simulates a scarce-data deployment scenario
• Distribution calibration is most impactful in 1-shot settings
• Episodes are sampled randomly from held-out novel categories

A measure of similarity between two non-zero vectors computed as the cosine of the angle between them. Used as a classification metric in calibrated few-shot classifiers.

• Range: [-1, 1] where 1 indicates identical direction
• Scale-invariant, focusing on angular proximity
• Often preferred over Euclidean distance in high dimensions
• Calibrated feature norms can improve cosine-based decisions

A reasoning mode where the classifier considers the entire query set as a batch and leverages the marginal distribution of unlabeled queries to refine predictions.

• Uses query-set statistics to normalize or re-center embeddings
• Complements distribution calibration by exploiting test-time structure
• Particularly effective when query sets are class-balanced
• Converts an inductive problem into a transductive one