Description
Self-supervised learning trains models on pretext tasks derived from the data itself (contrastive pairs, masked reconstruction, augmentation invariance) so that learned representations transfer to downstream supervised tasks with few labels. It is the most promising answer to CSI-label scarcity: massive amounts of unlabeled CSI exist on every commodity AP, and SSL turns that volume into useful pretraining.
When it's used
- CSI foundation-model pretraining
- Bringing a baseline up to speed with only a few labelled windows
- Cross-environment representations using augmentation invariance
Limitations
- Pretext-task design strongly biases what is learned
- Augmentation strategies for CSI are still poorly studied
- Compute-heavy at pretraining time