Introduction
A weather forecast that says "20 °C tomorrow" is useful. One that says "20 °C, give or take 1 degree" or "somewhere between 10 and 30 °C" is much more useful — it tells you how much to trust it.
Most ML models give a single best guess. Bayesian machine learning treats the model's parameters as uncertain, so the model can also say how sure it is. That matters whenever wrong, overconfident predictions are costly:
- medicine — "I'm not sure about this scan; ask a doctor";
- self-driving — slow down when perception is uncertain;
- finance and forecasting — plan for a range, not a single number;
- active learning — ask humans to label the examples the model is least sure about;
- small datasets — where uncertainty is naturally large.