⚙️ Core UQ Methods
Now the methods themselves — six families of techniques for making a neural network say how sure it is, each with a hands-on implementation.
Part 1 — Foundations
You built the case for uncertainty quantification: why accuracy alone is insufficient in a clinical setting, why confident errors are more dangerous than cautious ones, and how the softmax layer creates an illusion of confidence even on inputs the model has never seen anything like. You also separated aleatoric uncertainty — irreducible noise in the data — from epistemic uncertainty, which more data can reduce. That distinction decides the clinical response, and it drives every method that follows.
About this part
This is the longest part of the course, and the heart of it. It opens with the Bayesian view of a neural network — treating weights as distributions rather than fixed numbers — and then works through the two practical approximations that make that idea tractable: variational inference and MC Dropout.
Deep Ensembles arrive next by a completely different route. They make no Bayesian claims and require no special layers or loss functions: you simply train the same architecture several times from different random initializations and look at how much the members disagree. It is the simplest idea in this part and, on most benchmarks, one of the strongest.
Two further methods take their own paths to the same goal. Evidential Deep Learning predicts a distribution over probabilities in a single forward pass, and Conformal Prediction steps outside all of these frames to give prediction sets with a coverage guarantee that holds without any assumption about your model.
Every concept session is paired with an implementation session that builds the method in PyTorch on real chest X-ray data, so you finish with working code, not just intuition. The part closes with a summary session comparing all five approaches side by side.