Description
Few-shot learning trains a model that can generalise from only a handful of labelled examples per class, usually via meta-learning over related tasks (MAML, prototypical networks) or via metric-learning embeddings. In CSI sensing it is the practical workaround for sites where each gesture / activity can be demoed only a few times.
When it's used
- New-user / new-room calibration with minimal labelling effort
- Open-set CSI gesture vocabularies
- Cross-domain CSI transfer with only few shots in the target
Limitations
- Performance is highly sensitive to support-set choice
- Overfit on the meta-training task family
- Worse than full supervision when data is plentiful