Description
Supervised learning trains a model from labelled (input, target) pairs, optimising a loss that quantifies prediction error. It is the default training paradigm in CSI sensing — every HAR / occupancy / gesture benchmark assumes labelled CSI windows. The thesis treats supervised learning as the comparator against which self-supervised-learning, few-shot-learning, and BLE-calibrated weak labels are evaluated.
When it's used
- Standard CSI HAR / occupancy classifier training
- Density-map regression with annotated head positions
- Reading-pipeline ablations against weakly-supervised alternatives
Limitations
- Labels are expensive to collect for CSI (no easy crowdsourcing)
- Generalisation is bounded by labelled-domain coverage
- Active learning / weak supervision are routinely needed in practice