What is uncertainty calibration under covariate shift in deep learning models?

It evaluates whether predicted probabilities remain meaningful when input distributions change.

Uncertainty calibration under covariate shift studies whether a model’s predicted probabilities (e.g., confidence scores) still match empirical frequencies when P(X) changes. Even if P(Y|X) is stable, neural networks often become miscalibrated due to representation shift and feature scaling changes. Evaluation involves reliability diagrams across domains, Expected Calibration Error (ECE) under shift, and temperature scaling per environment. Advanced approaches also use domain-specific calibration layers or Bayesian posteriors over logits.