Description
Bayesian inference computes the posterior p(θ | data) ∝ p(data | θ) p(θ) given a prior and a likelihood. In WiFi sensing the term covers a spectrum: closed-form Gaussian-conjugate updates inside Kalman filters, MCMC / variational posteriors over neural-network weights, and Bayesian model selection between competing CSI representations. The thesis uses Bayesian framing whenever calibration uncertainty must propagate into downstream decisions.
When it's used
- Uncertainty quantification on CSI inferences
- Hyperparameter inference for crowd-model calibration
- Probabilistic indoor-positioning fusion
- Bayesian neural networks for safety-aware sensing
Limitations
- Posterior approximations (variational, MCMC) are expensive
- Sensitive to prior choice on small datasets
- Calibration of approximate posteriors is itself non-trivial