Abstract:Black-box machine learning models are now routinely used in high-risk settings, like medical diagnostics, which demand uncertainty quantification to avoid consequential model failures. Conformal prediction (a.k.a. conformal inference) is a user-friendly paradigm for creating statistically rigorous uncertainty sets/intervals for the predictions of such models. Critically, the sets are valid in a distribution-free sense: they possess explicit, non-asymptotic guarantees even without distributional assumptions or model assumptions. One can use conformal prediction with any pre-trained model, such as a neural network, to produce sets that are guaranteed to contain the ground truth with a user-specified probability, such as 90%. It is easy-to-understand, easy-to-use, and general, applying naturally to problems arising in the fields of computer vision, natural language processing, deep reinforcement learning, and so on. This hands-on introduction is aimed to provide the reader a working understanding of conformal prediction and related distribution-free uncertainty quantification techniques with one self-contained document. We lead the reader through practical theory for and examples of conformal prediction and describe its extensions to complex machine learning tasks involving structured outputs, distribution shift, timeseries, outliers, models that abstain, and more. Throughout, there are many explanatory illustrations, examples, and code samples in Python. With each code sample comes a Jupyter notebook implementing the method on a real-data example; the notebooks can be accessed and easily run by following the code footnotes.

Conformal inference is (almost) free for neural networks trained with early stopping

Conformal Prediction: A Gentle Introduction

Onboard Out-of-Calibration Detection of Deep Learning Models using Conformal Prediction

Early stopping by correlating online indicators in neural networks

Conformal Validity Guarantees Exist for Any Data Distribution (and How to Find Them)

Early Stopping of Untrained Convolutional Neural Networks

Conformal Inference for Invariant Risk Minimization

CONFINE: Conformal Prediction for Interpretable Neural Networks

Conformalization of Sparse Generalized Linear Models

Nothing Conformal about Adaptive Conformal Inference

A novel Deep Learning approach for one-step Conformal Prediction approximation

Conformal prediction for robust deep nonparametric regression

Conformal Prediction using Conditional Histograms

Transductive conformal inference with adaptive scores

Self-Calibrating Conformal Prediction

A Large-Scale Study of Probabilistic Calibration in Neural Network Regression

Evaluating the Utility of Conformal Prediction Sets for AI-Advised Image Labeling

Making learning more transparent using conformalized performance prediction

A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification

Predictive Inference with Feature Conformal Prediction

Exact and Robust Conformal Inference Methods for Predictive Machine Learning With Dependent Data