Uncertainty Quantification via Conformal Prediction

Modern computer vision and machine learning systems are routinely deployed in high-stakes settings such as medical image diagnosis, autonomous perception, and remote sensing, where a single point prediction offers no guarantee about how much to trust it. Standard remedies like Bayesian posteriors, bootstrap intervals, softmax confidence scores, and quantile regression either rely on strong distributional assumptions, require expensive re-sampling, or provide no rigorous coverage guarantee at all. Conformal Prediction (CP) closes this gap: it is a model-agnostic, distribution-free framework that wraps around any pre-trained predictor and converts its point predictions into prediction sets or intervals that provably contain the true outcome with a user-specified probability, in finite samples, with no assumptions on the underlying data-generating distribution.

This tutorial gives a self-contained introduction to conformal prediction and a curated tour of selected recent topics of interest to the ICVGIP audience: conformalized image classification with adaptive prediction sets, conformal prediction under distribution shift, and conformal uncertainty quantification for sequential and dependent data. The session closes with a hands-on implementation block so that participants leave able to conformalize their own models.