Description
Transfer learning re-uses a model trained on one source distribution to accelerate or enable learning on a related target distribution. In WiFi sensing it is the practical answer to the data-scarcity-plus-environment-shift problem: pretrain on a large synthetic / multi-room corpus, fine-tune on a small target site. It overlaps with fine-tuning (one tactic) and domain-adaptation (one specific transfer setting).
When it's used
- Bringing a CSI HAR model from a labelled source room to a new deployment
- Pretraining on synthetic crowd simulations before fitting real CSI data
- Multi-site federated CSI sensing benchmarks
Limitations
- Negative transfer is real if source and target distributions are too different
- Choice of which layers to freeze / fine-tune is empirical
- Catastrophic forgetting of source task if not regularised
Source Papers
- meegahapola2024_a321 ↗ — transfer learning across CSI sensing domains
- yang2023_a34a ↗ — transfer learning in WiFi sensing roadmap
- chen2023_5cbd ↗ — transfer learning in CSI generalisation taxonomy
- wang2026_2758 ↗ — transfer-aware CSI sensing
- cakoni2023_7150 ↗ — transfer learning across radar configurations