Few-Shot Learning

Few-shot learning is typically implemented through meta-learning (learning to learn). The model is trained across many related tasks—each with its own small support set—to acquire an internal optimization strategy that generalizes rapidly to novel modulation classes. This episodic training paradigm teaches the classifier to extract transferable signal features rather than memorizing specific constellations.

A dominant architecture for few-shot modulation recognition is the prototypical network. It learns a metric embedding space where signal samples cluster around a prototype representation for each class. Classification of a new query sample is performed by finding the nearest class prototype in this learned space, enabling recognition of unseen modulation schemes without retraining.

The few-shot paradigm operates on two distinct data partitions per episode:

• Support Set: A small number of labeled examples (e.g., 1-shot or 5-shot) used to define the classification task.
• Query Set: Unlabeled samples used to evaluate how well the model has adapted to the support set's classes. This structure forces the model to reason by comparison rather than memorization.

At the core of few-shot classification is distance metric learning. The network is trained to minimize the distance between embeddings of the same modulation class while maximizing the separation between different classes. Common distance functions include Euclidean distance and cosine similarity, with the choice significantly impacting performance on high-dimensional IQ sample embeddings.

Siamese networks provide an alternative few-shot approach by learning a similarity function between pairs of signal samples. Two identical subnetworks process separate inputs, and their outputs are combined to predict whether the samples belong to the same modulation class. This pairwise comparison strategy is particularly effective for one-shot learning scenarios where only a single reference example exists.

Few-shot modulation classifiers face significant domain gap issues when the signal characteristics of the support set differ from the query set due to varying channel conditions, receiver hardware, or SNR levels. Techniques such as domain adversarial training and feature-level augmentation are employed to learn channel-invariant embeddings that maintain classification accuracy across heterogeneous deployment environments.

A learning to learn paradigm where a model is trained across a distribution of tasks, not just a single dataset. In modulation recognition, the model learns an initialization that can rapidly adapt to new signal types after seeing only a few IQ sample sequences.

• Optimization-based: MAML (Model-Agnostic Meta-Learning) finds parameters that are sensitive to task changes
• Metric-based: Learns an embedding space where distance directly corresponds to class similarity
• The meta-learner internalizes general signal structure knowledge, enabling fast fine-tuning on novel modulations

A metric-based few-shot architecture that computes a prototype vector for each modulation class by averaging the embeddings of its few support examples. Classification of a query signal is performed by finding the nearest prototype in the learned embedding space using Euclidean distance.

• Each modulation type is represented as a single point in a high-dimensional space
• The network learns a non-linear mapping where signals of the same scheme cluster tightly
• Computationally efficient at inference, requiring only a distance calculation against stored prototypes

A twin-branch neural architecture that learns a similarity function between pairs of signal samples. During training, the network is presented with pairs of IQ sequences and learns to output a low distance for same-modulation pairs and a high distance for different-modulation pairs.

• Uses shared weights between both branches to ensure symmetric processing
• Contrastive loss or triplet loss drives the embedding space organization
• Excels at one-shot learning where only a single reference example per class is available for comparison

A framework that combines attention mechanisms with external memory to classify query examples by comparing them directly to a small support set. The model uses a differentiable nearest-neighbor approach where the attention kernel is learned end-to-end.

• Full context embeddings: the support set examples are encoded conditioned on each other
• Cosine similarity-based attention weights determine the contribution of each support example
• Naturally handles variable-sized support sets without retraining, ideal for dynamic signal environments

Specialized augmentation strategies designed to synthetically expand the limited support set without introducing label-invariant distortions. For RF signals, this includes applying realistic channel impairments that preserve the underlying modulation identity.

• Hallucination-based: Generative models create plausible new examples from the few real samples
• Signal transformations: Controlled phase rotation, frequency offset, and AWGN injection
• Mixup and CutMix: Interpolating between support examples to create smooth decision boundaries
• Critical for preventing overfitting when only 1-5 labeled IQ sequences are available per novel class

A training regime that mirrors the few-shot test scenario by constructing mini-batches as mini-episodes. Each episode consists of a small support set and a query set sampled from a subset of classes, forcing the model to learn rapid adaptation strategies.

• N-way K-shot: Each episode contains N classes with K examples each
• The model never sees the same class configuration repeatedly, preventing memorization
• Aligns the training objective directly with the deployment scenario, reducing the train-test distribution gap
• Essential for meta-learning approaches like Prototypical Networks and MAML