Domain Adaptation and Transfer

Definition:
Domain Adaptation

When source and target distributions differ:

Supervised DA: labelled target data available
Unsupervised DA: only unlabelled target data
Domain-adversarial training: learn features that are discriminative for the task but invariant to the domain

class DomainAdversarial(nn.Module):
    def __init__(self, backbone, classifier, domain_disc):
        super().__init__()
        self.backbone = backbone
        self.classifier = classifier
        self.domain_disc = domain_disc  # GRL + discriminator

Example: Sim-to-Real Transfer for Wireless

Transfer a model trained on simulated channels to real measurements.

Solution

Approach

Train on simulated data (large, cheap)
Fine-tune on small real-world measurements
Use domain adaptation to align feature distributions

Pre-Trained Denoisers for Inverse Problems Model Export and Inference Optimization