Self-Supervised Learning

The core mechanism of SSL where a model solves an artificially constructed task to learn data structure. The supervisory signal is generated automatically from the input data itself.

• Rotation Prediction: The model must determine the rotation angle applied to a constellation diagram, forcing it to learn the geometric structure of modulation schemes.
• Temporal Shuffling: Segments of an IQ sample stream are reordered, and the model must predict the correct sequence, learning temporal dependencies.
• Inpainting: A portion of a spectrogram or time-series signal is masked, and the model must reconstruct the missing content, learning local and global signal features.
• Contrastive Instance Discrimination: Each signal sample is treated as its own class, and the model learns to identify augmented versions of the same sample while distinguishing them from others.

A dominant SSL paradigm that learns representations by comparing sample pairs in a latent space. The objective is to maximize agreement between differently augmented views of the same data point (positive pairs) while minimizing agreement with other points (negative pairs).

• SimCLR: Uses large batch sizes and a composition of data augmentations to create positive pairs, employing an NT-Xent loss to pull similar representations together.
• MoCo (Momentum Contrast): Maintains a dynamic dictionary as a queue of encoded samples, using a momentum-updated encoder to ensure consistent negative representations without massive batch sizes.
• BYOL (Bootstrap Your Own Latent): Eliminates the need for negative pairs entirely by training a target network to predict the output of an online network, preventing representational collapse through a slow-moving average of the online parameters.

Inspired by BERT in NLP, this approach masks a random subset of an input signal and trains the model to reconstruct the missing content from the visible context. This forces the model to learn bidirectional dependencies and the underlying structure of the waveform.

• Masked Autoencoding: A Vision Transformer (ViT)-style architecture masks a high percentage of input patches (e.g., 75% of a spectrogram) and uses an asymmetric encoder-decoder to reconstruct the original signal.
• IQ Masking: Random time-steps in an IQ sample sequence are zeroed out, and the model must predict the original complex values, learning the temporal and phase relationships inherent to the modulation.
• Feature Prediction: Instead of reconstructing raw signal values, the model predicts the latent features of the masked content as computed by a teacher network, focusing the learning on high-level semantic features.

The practical value of SSL lies in the transferability of the learned representations. After pretraining on a massive unlabeled corpus, the model is adapted to the specific downstream task of modulation classification.

• Linear Probing: A linear classifier is trained on top of the frozen, pretrained backbone. This evaluates the intrinsic quality of the learned representations without modifying them.
• Full Fine-Tuning: The entire pretrained model is unfrozen and trained on a small, labeled modulation dataset. This allows the model to adapt its low-level feature detectors to the specific channel conditions and modulation set of the target domain.
• Low-Data Regime Efficacy: SSL-pretrained models consistently outperform purely supervised models when the number of labeled training examples is severely limited (e.g., fewer than 100 samples per modulation class), a common scenario in electronic warfare and spectrum monitoring.

The choice of data augmentation defines the invariances the model learns. For RF signals, augmentations must preserve the modulation identity while simulating realistic channel and hardware impairments.

• Channel Impairments: Adding simulated AWGN, phase offset, frequency offset, and multipath fading forces the model to learn representations invariant to standard propagation effects.
• Hardware Impairments: Simulating IQ imbalance, DC offset, and sampling clock drift teaches the model to be robust to receiver-specific artifacts.
• Transformations: Applying random time shifts, amplitude scaling, and small frequency rotations creates positive pairs for contrastive learning without altering the modulation scheme.
• Critical Constraint: Augmentations must not break the signal's defining characteristics. An excessive frequency shift could morph a QPSK signal into something unrecognizable, creating a false positive pair.

A central challenge in SSL is representational collapse, where the encoder learns a trivial, constant representation for all inputs, achieving zero loss but learning nothing. Architectures employ specific mechanisms to prevent this.

• Negative Pairs: Contrastive methods like SimCLR explicitly push representations of different samples apart, preventing the encoder from mapping everything to a single point.
• Momentum Encoder: BYOL and MoCo use a slowly updating target encoder to provide a stable prediction target, breaking the symmetry that would allow both networks to collapse simultaneously.
• Asymmetric Architecture: Using a predictor head on the online network but not on the target network creates an information bottleneck that prevents the trivial solution.
• Batch Normalization: The implicit contrastive effect of batch statistics across different samples can also help prevent collapse, though it can lead to unintended information leakage.

The model is trained to predict a synthetic phase rotation applied to a baseband signal. To succeed, the network must learn the underlying geometric structure of the constellation diagram.

• Input: IQ samples rotated by 0°, 90°, 180°, or 270°
• The encoder internalizes rotational equivariance properties
• Produces features that are inherently robust to phase offset
• Particularly effective for PSK and QAM family recognition

The network receives two randomly sampled segments from the same signal and must predict their temporal order. This forces the model to learn the sequential structure and symbol-rate dynamics of the transmission.

• Learns temporal coherence without explicit symbol timing
• Captures baud rate and pulse-shaping characteristics
• Distinguishes between constant-envelope and amplitude-varying schemes
• Effective pre-training for LSTM and Transformer backbones

A powerful pretext framework where the model learns to predict future latent representations from past ones while distinguishing them from negative samples. This maximizes mutual information between context and future observations.

• Uses InfoNCE loss to contrast positive and negative pairs
• Learns slow features that capture modulation statistics
• Naturally handles variable-length signal bursts
• State-of-the-art for low-SNR representation learning

The time-series signal is divided into shuffled segments, and the network must predict the correct permutation. This forces holistic understanding of signal structure across multiple timescales.

• Segments are extracted from the complex baseband representation
• Model learns inter-symbol relationships and burst structure
• Encourages learning of cyclostationary features implicitly
• Robust to moderate multipath fading conditions

Inspired by BERT in NLP, random portions of the IQ sequence are masked, and the model must reconstruct the missing samples. This teaches the network the statistical regularities of modulated waveforms.

• Masking ratio typically 15-30% of time steps
• Reconstruction in complex-valued domain preserves phase info
• Learns both local symbol transitions and global burst envelope
• Transfers well to both classification and denoising tasks

Combining multiple pretext tasks during pre-training yields richer representations than any single task alone. A shared encoder trunk feeds task-specific heads optimized jointly.

• Example: Rotation prediction + Relative positioning + Jigsaw
• Prevents the encoder from overfitting to one heuristic
• Produces features that generalize across channel impairments
• Requires careful loss weighting to balance gradient contributions

A dominant self-supervised paradigm that trains a model to pull similar data points together and push dissimilar points apart in an embedding space. In modulation recognition, this means learning to cluster IQ samples from the same modulation scheme while separating different schemes—all without explicit labels. Frameworks like SimCLR and MoCo use heavy data augmentation to create positive pairs, teaching the encoder to learn augmentation-invariant features that transfer directly to downstream classification tasks.

The engineered self-supervised objective that forces the model to learn meaningful representations by solving a surrogate problem on unlabeled data. Common pretext tasks for signal processing include:

• Rotation prediction: Classifying the applied phase rotation to a constellation diagram
• Temporal ordering: Determining the correct sequence of shuffled IQ sample segments
• Masked signal modeling: Reconstructing intentionally corrupted portions of a waveform The quality of the learned representation depends entirely on how well the pretext task captures the underlying structure of modulated signals.

A self-supervised technique where a large fraction of the input signal is randomly masked, and the model learns to reconstruct the missing portions. Inspired by BERT in NLP and MAE in vision, this approach forces the encoder to learn the statistical regularities and spectral structure of modulated waveforms. For IQ samples, masking contiguous time-frequency blocks teaches the model to infer coherent signal structure from partial observations, producing representations that excel at modulation classification even under severe channel impairment.

The fundamental goal of self-supervised learning: transforming raw, high-dimensional signal data into a compact, semantically meaningful latent space where similar modulation types cluster together. A well-trained representation exhibits:

• Linear separability: Modulation classes become distinguishable with a simple linear classifier
• Robustness: Features remain stable under noise, fading, and hardware impairments
• Transferability: The same representation generalizes to unseen channel conditions and receiver hardware This decouples feature extraction from task-specific fine-tuning, enabling rapid adaptation to new modulation recognition challenges.

Neural architectures designed specifically for self-supervised learning that process two or more views of the same input through parallel encoders. Key variants include:

• Siamese networks: Weight-shared encoders with a prediction head that maps one view to another, as in BYOL and SimSiam, eliminating the need for negative samples
• Cross-modal embeddings: Aligning representations from different signal transformations, such as IQ samples and spectrograms, to learn modality-invariant features These architectures prevent representational collapse—where the encoder outputs a constant vector—through stop-gradient operations or asymmetric architectural design.

The process of adapting a self-supervised pre-trained model to the specific task of modulation classification using a small set of labeled examples. The standard protocol involves:

• Linear probing: Freezing the pre-trained encoder and training only a linear classifier on top, evaluating the raw quality of learned representations
• Full fine-tuning: Unfreezing all weights and training end-to-end with a low learning rate, achieving maximum task performance
• Partial fine-tuning: Selectively unfreezing later layers while keeping early feature extractors frozen, balancing adaptation speed with catastrophic forgetting prevention This paradigm achieves near-supervised performance with a fraction of the labeled data.