Description
Deep Learning is the umbrella for multi-layer neural-network methods that learn hierarchical representations directly from raw or lightly-engineered features. In CSI sensing it has displaced almost every hand-crafted classifier: convolutional, recurrent, and Transformer architectures dominate occupancy, HAR, gesture, and crowd-counting benchmarks. The thesis treats deep learning as a default backbone and focuses on what additional inductive bias (calibration, physics) it needs to generalise.
When it's used
- End-to-end CSI → label pipelines for occupancy / HAR / gesture
- Crowd density-map regression
- Pretext tasks for
self-supervised-learning - Feature backbones combined with classical estimators
Limitations
- Data-hungry — small CSI datasets demand augmentation or transfer learning
- Domain shift across rooms / hardware destroys naive accuracy
- Black-box: no built-in conservation laws or physical guarantees