Few-Shot Class-Incremental Learning

FSCIL structures learning as a sequence of sessions. In each session, the model encounters a new set of N novel modulation classes with only K labeled examples each (an N-way K-shot task). The model must learn to discriminate these new classes while maintaining its ability to recognize all classes from previous sessions. This directly mirrors real-world spectrum operations where new emitters or waveforms appear incrementally over time.

The central tension in FSCIL is balancing stability (resisting catastrophic forgetting of old modulation types) against plasticity (the capacity to rapidly absorb new signal classes from limited data). A model that is too rigid fails to learn new waveforms; one that is too plastic overwrites its existing knowledge. Solutions often involve frozen feature extractors, weight regularization, or rehearsal-based methods to navigate this trade-off.

A dominant FSCIL strategy uses prototypical networks as the backbone. The model maintains a fixed, pre-trained embedding function and stores a single prototype vector (the mean embedding) for each class seen so far. When a new session arrives, it computes prototypes for the novel classes and appends them to the existing gallery. Classification is performed via nearest-prototype matching in the embedding space, naturally avoiding forgetting since old prototypes remain unchanged.

Unlike traditional incremental learning, FSCIL often operates under strict data privacy and storage constraints that prohibit storing raw signal samples from previous sessions. This exemplar-free requirement forces models to retain knowledge without a replay buffer. Techniques include knowledge distillation from a frozen copy of the old model, feature-space hallucination to generate synthetic representations of old classes, or parameter isolation where dedicated sub-networks are allocated per session.

FSCIL assumes a large, labeled base session with abundant data for many modulation classes before the incremental phase begins. This initial training phase is critical: the model learns a rich, generalizable embedding space that can separate signal types it has never seen. Techniques like supervised contrastive learning or cosine-similarity-based classifiers during base training create a feature manifold where future novel classes will naturally form distinct, separable clusters.

FSCIL performance is measured across multiple dimensions beyond simple accuracy. Key metrics include:

• Average Incremental Accuracy: The mean classification accuracy across all classes after each session.
• Forgetting Rate: The performance drop on old classes after learning new ones.
• Harmonic Mean: A combined score penalizing imbalance between old and new class accuracy.
• Area Under the Curve (AUC): The integral of accuracy over the sequence of sessions, capturing the full learning trajectory.

The primary challenge addressed by class-incremental learning. It refers to the phenomenon where a neural network's performance on previously learned tasks or classes degrades abruptly after being trained on new data. This occurs because gradient updates for new classes overwrite the weights essential for recognizing old ones. In dynamic spectrum environments, forgetting a rare but critical modulation type is a mission-compromising failure.

A foundational technique for mitigating forgetting. A copy of the model from the previous incremental state serves as a teacher network. During training on new classes, a distillation loss forces the current student model to match the teacher's output logits or feature representations for old classes. This preserves the decision boundaries of prior tasks without requiring access to the original training data.

A strategy that stores a small, representative subset of training samples—exemplars—from each previously learned class in a fixed-size memory buffer. When learning new classes, these exemplars are interleaved with the new data to jointly optimize the model. The herding algorithm is commonly used to select the most prototypical samples that best approximate the class mean in the feature space.

An alternative to storing raw exemplars that preserves data privacy. Instead of saving actual signal samples, the model stores only the mean vector (prototype) and covariance statistics for each old class in the embedding space. During incremental training, synthetic feature vectors are sampled from these Gaussian distributions and replayed alongside new class features to prevent the classifier bias from shifting entirely toward the new classes.

A regularization-based approach that identifies the synaptic weights most critical for previously learned tasks by computing the Fisher Information Matrix. During incremental training, a quadratic penalty term constrains changes to these important parameters. This allows the model to learn new modulation types using the remaining free parameters while physically anchoring the knowledge of old tasks in the network's structure.

The standard benchmarking protocol adapted for incremental settings. After each new class is learned, the model is evaluated on a balanced test set containing all classes seen so far. A typical metric is the average incremental accuracy, which measures the model's performance across all sessions. This exposes whether the model is maintaining a stable representation or catastrophically biasing toward the most recent classes.