Description
Domain adaptation is the special case of transfer-learning where the task stays the same but the input distribution shifts (different room, different antennas, different humans). The dominant technique is adversarial alignment (DANN / domain-adversarial training): a feature extractor is trained to fool a domain discriminator while serving the task classifier, producing domain-invariant features. It is the most studied antidote to environment dependence in CSI sensing.
When it's used
- Cross-room / cross-NIC CSI sensing
- BLE-anchored CSI calibration where target labels are sparse
- Cross-user gesture / gait recognition
Limitations
- Adversarial training is unstable; needs careful balancing
- Source / target alignment can hurt task performance
- Multi-source DA is much harder than the canonical two-domain setting