Domain Adaptation

Explicitly measures and minimizes the statistical distance between source and target domain feature distributions using metrics like Maximum Mean Discrepancy (MMD) or Correlation Alignment (CORAL). A loss term penalizes distributional divergence, forcing the network to learn domain-invariant representations.

• MMD: Compares kernel embeddings of source and target distributions in a reproducing kernel Hilbert space
• CORAL: Aligns second-order statistics by minimizing the difference in covariance matrices
• Applied after intermediate network layers to create a shared feature space

Employs a gradient reversal layer (GRL) and a domain discriminator network in a minimax game. The feature extractor learns to produce representations that confuse the discriminator, making source and target domains indistinguishable.

• Domain-Adversarial Neural Network (DANN): The canonical architecture using a GRL to invert gradients during backpropagation
• Adversarial Discriminative Domain Adaptation (ADDA): Uses separate source and target encoders with an asymmetric training procedure
• Particularly effective when target domain labels are entirely absent

Leverages pretext tasks on unlabeled target data to learn domain-specific structure before or jointly with the main classification objective. Common pretext tasks include rotation prediction, jigsaw puzzle solving, and contrastive predictive coding.

• SimCLR-style contrastive learning: Pulls augmented views of the same target sample together while pushing apart different samples
• BYOL (Bootstrap Your Own Latent): Eliminates negative pairs by using a momentum encoder and a predictor network
• Enables feature learning without any target domain labels

Deliberately varies simulation parameters—such as channel impulse response, noise floor, carrier frequency offset, and multipath profiles—during training to expose the model to extreme diversity. The network learns to treat domain-specific variations as irrelevant nuisance factors.

• Uniform sampling: Randomizes parameters within predefined ranges for each training batch
• Curriculum-based randomization: Gradually increases randomization difficulty as training progresses
• Originated in sim-to-real robotics but highly effective for RF channel generalization

Updates model parameters at inference time using only the incoming unlabeled target sample or batch. Techniques include batch normalization recalibration and entropy minimization to adapt to instantaneous channel conditions without retraining.

• TENT: Adjusts batch normalization statistics by minimizing prediction entropy on each test batch
• SHOT: Uses self-supervised pseudo-labeling and feature clustering for source-free adaptation
• Critical for mobile SDR platforms moving between diverse electromagnetic environments

When a small number of labeled target domain samples are available, parameter-efficient fine-tuning methods adapt a pre-trained source model without catastrophic forgetting. Techniques include LoRA (Low-Rank Adaptation) and adapter modules.

• LoRA: Injects trainable low-rank matrices into frozen pre-trained weights, updating only a fraction of parameters
• Prototypical networks: Compute class prototypes from few-shot target examples for nearest-neighbor classification
• Balances adaptation speed with preservation of source-domain knowledge