Description
Fine-tuning continues training a pretrained model — fully or with selected layers frozen — on a target dataset, typically with a lower learning rate. It is the workhorse tactic of transfer-learning in WiFi sensing: pretrain on a large self-supervised CSI corpus, fine-tune on a few labelled minutes from a new room.
When it's used
- Adapting pretrained CSI backbones to new venues
- Bringing foundation-model representations into specialised sensing tasks
- Continual learning across deployments
Limitations
- Risk of overfitting tiny target sets
- Hyperparameter search (LR, layers to freeze) is empirical
- Catastrophic forgetting of source task without regularisation