Pseudo-Labeling

The core gating mechanism that determines which unlabeled samples are admitted into the training set. Only predictions exceeding a predefined probability threshold (e.g., 0.95) are converted into pseudo-labels. This prevents confirmation bias, where the model reinforces its own mistakes by training on low-confidence, likely incorrect guesses. In RF modulation classification, a high threshold ensures only unambiguous signal types are used for self-training.

Pseudo-labeling is not a one-shot process. It operates in a cyclical loop:

• Step 1: Train an initial model on a small, labeled RF dataset.
• Step 2: Use the model to predict labels for a large, unlabeled dataset.
• Step 3: Select predictions above the confidence threshold and add them to the training set.
• Step 4: Retrain the model on the expanded dataset and repeat. This allows the model to progressively refine its decision boundary in sparse regions of the signal space.

The underlying mathematical principle that makes pseudo-labeling effective. By converting high-confidence soft predictions (e.g., a probability distribution) into hard labels, the model is forced to make low-entropy decisions. This pushes decision boundaries away from high-density regions of unlabeled data, acting as a form of implicit regularization that assumes data points belonging to the same class cluster together in the feature space.

A critical safeguard against confirmation bias where the model labels all unaligned data as the majority class. Advanced pseudo-labeling implementations enforce a prior distribution constraint, ensuring the proportion of pseudo-labels assigned to each class matches the expected distribution in the unlabeled set. This prevents a model trained on rare RF emitters from collapsing to predict only the dominant background noise class.

Often combined with pseudo-labeling to improve robustness. The model is forced to assign the same pseudo-label to an unlabeled RF sample even when it is perturbed with additive noise, fading, or spectrogram augmentation. This enforces a smooth manifold assumption: small perturbations to a signal should not change its classification. The combination of pseudo-labeling and consistency loss is foundational to modern semi-supervised methods like FixMatch.

A training strategy that dynamically adjusts the confidence threshold during the self-training process. Initially, a very high threshold admits only the easiest, most unambiguous samples. As the model improves, the threshold is gradually lowered to incorporate progressively harder examples. This curriculum learning approach mimics human education by starting with simple concepts before tackling edge cases in complex RF environments.

The primary selection mechanism in pseudo-labeling that only accepts model predictions exceeding a predefined probability score (e.g., >0.95) as ground truth. In RF modulation classification, a softmax output of 0.97 for QPSK would convert that unlabeled IQ sample into a labeled training example, while a 0.65 prediction would be discarded. Dynamic thresholding schedules that increase the confidence bar over training epochs prevent early confirmation bias from noisy pseudo-labels.

The iterative process where a model trained on initial labeled data generates pseudo-labels for unlabeled data, then retrains on the expanded dataset. In RF fingerprinting, this loop progressively refines emitter identification: a model trained on 100 labeled device captures can pseudo-label 10,000 additional transmissions, then retrain to capture subtle hardware impairments it initially missed. The loop repeats until performance plateaus or a maximum iteration count is reached.

A complementary technique that enforces stable predictions across perturbed versions of the same unlabeled RF input. When combined with pseudo-labeling, the model must predict the same modulation scheme for a signal passed through different channel impairment simulations (e.g., varying noise floors or Doppler shifts). This prevents the model from assigning high-confidence but brittle pseudo-labels to inputs near decision boundaries.

A graph-based alternative to direct pseudo-labeling where known labels diffuse through a similarity graph constructed from unlabeled RF samples. Each unlabeled signal's class is inferred from its nearest labeled neighbors in embedding space rather than from a model prediction. This approach excels in few-shot RF scenarios where a classifier lacks the capacity to generate reliable pseudo-labels directly but signal structure is well-captured by a pre-trained feature extractor.

The primary failure mode of pseudo-labeling where the model reinforces its own initial mistakes by confidently mislabeling ambiguous RF signals and retraining on those errors. In automatic modulation classification, an early bias toward classifying high-order QAM as QPSK can cascade through self-training iterations. Mitigation strategies include ensemble disagreement filtering, where pseudo-labels are only accepted when multiple independently trained models agree on the prediction.

Modern semi-supervised algorithms that unify pseudo-labeling with consistency regularization and Mixup augmentation. FixMatch generates pseudo-labels from weakly augmented RF signals (e.g., slight noise addition) and enforces those labels on strongly augmented versions (e.g., significant frequency shifts). This forces the model to learn robust, invariant features while expanding the training set. These methods achieve state-of-the-art performance on RadioML benchmarks with as few as 10 labeled examples per modulation class.