Contrastive Learning

The core mechanism relies on creating positive pairs—two augmented versions of the same base signal—and negative pairs from different signals. For RF data, augmentations include additive Gaussian noise, small frequency shifts, and time cropping. The model is trained to pull positive pairs together in the embedding space while pushing negative pairs apart, learning representations invariant to channel impairments and hardware imperfections.

The Noise Contrastive Estimation (InfoNCE) loss is the dominant objective function. It frames representation learning as a categorical classification problem where the model must identify the true positive pair among a batch of negative distractors. The loss is computed as:

• Cross-entropy over similarity scores between embeddings
• Temperature parameter τ controls concentration of the distribution
• Lower τ creates harder negatives, improving feature discrimination This loss directly optimizes the mutual information between augmented views.

Frameworks like MoCo (Momentum Contrast) maintain a dynamic dictionary of negative samples using a slowly-evolving momentum encoder. Key architectural components:

• Query encoder: Processes the current augmented view, updated by backpropagation
• Key encoder: Processes historical views, updated via exponential moving average
• Queue: Stores thousands of negative embeddings for consistent contrast This decoupling enables large, consistent negative sets without massive batch sizes, critical for GPU-constrained RF training pipelines.

A non-linear projection head (typically a 2-3 layer MLP) maps encoder representations to a lower-dimensional space where contrastive loss is applied. After pre-training, this head is discarded—only the encoder backbone is retained for downstream tasks. Critical design considerations:

• Prevents dimensionality collapse where all representations converge to a trivial constant
• Batch normalization in the projection head is essential for training stability
• Output dimension of 128-256 balances expressiveness and computational efficiency

In dense signal environments, random negative sampling is insufficient. Hard negative mining identifies negative samples that are semantically close to the anchor but belong to different classes—for example, QPSK vs. 8PSK signals at similar SNR levels. Techniques include:

• Ring loss to enforce uniform distribution on the hypersphere
• Debiased contrastive loss to correct sampling bias from random negatives
• Mining negatives from the same modulation family to learn fine-grained discriminative features This is essential for automatic modulation classification where inter-class boundaries are subtle.

After self-supervised pre-training on unlabeled RF captures, the learned encoder is transferred to downstream tasks via:

• Linear evaluation: Freeze the encoder, train only a linear classifier on top—measures representation quality directly
• Full fine-tuning: Unfreeze all weights and train end-to-end on labeled data
• Semi-supervised fine-tuning: Use a small labeled subset with consistency regularization In RF fingerprinting, contrastive pre-training reduces labeled data requirements by up to 4x while maintaining emitter identification accuracy above 95%.

The set of stochastic transformations applied to raw IQ samples to create the positive pairs required for contrastive pre-training. Effective RF augmentations must preserve the semantic identity of the signal while introducing realistic variability:

• Additive White Gaussian Noise (AWGN): Varies the SNR
• Frequency Offset: Simulates oscillator mismatch
• Time Shifting: Introduces random delays
• Phase Rotation: Mimics channel rotation The choice and strength of augmentations critically determine the quality of the learned representations and the model's downstream robustness to channel impairments.