Feature Hallucination

A dedicated generator network learns to produce realistic feature vectors for novel classes by modeling the underlying distribution of the embedding space. Unlike raw signal augmentation, this process operates on compact, high-level representations.

• The generator is conditioned on the limited available support samples
• Synthesized features are calibrated to lie near the true class manifold
• Prevents the classifier from overfitting to the few available examples
• Often implemented as a lightweight MLP or attention-based module

By populating sparse regions of the embedding space with hallucinated features, the classifier learns smoother and more robust decision boundaries. This directly combats the brittleness that arises when class prototypes are estimated from only a handful of samples.

• Expands the occupied volume of underrepresented classes
• Reduces the risk of query samples falling into ambiguous regions
• Improves N-way K-shot accuracy, especially for high K values
• Acts as a form of regularization in the feature space

A statistical technique that calibrates the feature distribution of base classes (with abundant data) to estimate the distribution of novel classes. This enables the generation of high-quality synthetic features without requiring a complex adversarial training setup.

• Assumes a shared statistical structure between base and novel classes
• Uses mean and covariance statistics from base classes to model novel ones
• Generated features are sampled from calibrated Gaussian distributions
• Provides a simple yet effective alternative to GAN-based hallucination

Manifold Mixup performs linear interpolations on learned hidden representations rather than raw inputs. When applied to few-shot modulation tasks, it creates a continuous spectrum of synthetic features between existing support samples.

• Encourages the classifier to behave linearly between training examples
• Produces smoother decision boundaries with fewer sharp transitions
• Can be combined with generator-based hallucination for additive benefit
• Particularly effective for modulation schemes with gradual parameter variations

Feature hallucination operates in the embedding space, while traditional data augmentation transforms raw IQ samples. This distinction is critical for computational efficiency and semantic validity.

• Feature-space: Low-dimensional, fast, semantically meaningful
• Raw-space: High-dimensional, computationally expensive, may produce invalid signals
• Hallucinated features preserve the discriminative structure learned by the encoder
• Raw augmentations risk introducing artifacts that confuse the classifier
• Hybrid approaches apply augmentation at both levels for maximum robustness

Feature hallucination can improve cross-domain few-shot learning where base training classes and novel test classes come from different distributions—such as synthetic vs. over-the-air signals.

• Calibrated generation adapts to distribution shifts in the embedding space
• Reduces reliance on domain-specific raw signal characteristics
• Enables models trained on clean simulations to generalize to noisy real-world captures
• Critical for deploying few-shot classifiers in dynamic spectrum environments

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 in the learned embedding space using Euclidean distance. This simple, non-parametric approach provides a strong baseline for few-shot modulation recognition and directly motivates feature-space augmentation strategies like hallucination to refine class boundaries.

A statistical technique that estimates the feature distribution of novel classes by leveraging statistics from base classes with abundant data. It assumes that the mean and covariance of a novel class can be approximated by transferring distributional properties from similar base classes. This calibrated distribution is then sampled to generate synthetic feature vectors, directly addressing the same class-boundary refinement problem as feature hallucination but through explicit statistical modeling rather than learned generation.

A regularization and augmentation method that performs linear interpolations not on raw input signals but on learned hidden layer representations. By blending feature vectors from different classes and assigning soft labels proportional to the interpolation ratio, it encourages smoother decision boundaries. In few-shot modulation recognition, manifold mixup provides a complementary approach to hallucination by expanding the coverage of the embedding space without requiring a separate generator network.

A learned, lower-dimensional vector representation where semantically similar inputs—such as signals sharing modulation characteristics—are mapped to nearby points. The quality of this space is critical: a well-structured embedding enables metric-based classifiers to succeed with simple distance measures. Feature hallucination operates entirely within this space, synthesizing plausible vectors to densify underrepresented regions and sharpen class boundaries without returning to the raw signal domain.

A meta-training strategy that structures learning into a series of N-way K-shot episodes, each simulating a low-data test scenario. The model is optimized not for a fixed set of classes but for the ability to rapidly adapt to any novel task. This paradigm is the foundation upon which feature hallucination generators are trained—the generator learns to produce useful augmentations by observing the performance gains they provide across thousands of simulated few-shot episodes.

The task of identifying test samples that differ fundamentally from the training data distribution. In open-set modulation recognition, a classifier must not only identify known schemes but also reject unknown signal types. Feature hallucination can complicate this boundary by generating synthetic features that may inadvertently overlap with true out-of-distribution regions, making robust OOD detection a critical companion technique when deploying hallucination-based classifiers in real spectrum environments.