Issue #94 Theory Edwin Dong, Anurag Shinde
Conformal Prediction: Reliable Uncertainty for Machine Learning
Machine learning systems are now increasingly deployed in high-stakes domains such as healthcare, autonomous vehicles, finance, and security.
In these environments, incorrect predictions can have serious consequences. Yet most machine learning models provide only a single prediction such as a class label or numerical value without conveying how confident the model actually is.
This limitation has motivated the development of uncertainty quantification techniques, which aim to measure how reliable a model’s predictions are.
In this issue, we will look at one of the most powerful and widely applicable uncertainty techniques called conformal prediction, a statistical framework that produces prediction sets with guaranteed coverage.
Unlike the previous methods, which rely on the distribution of the output data, conformal prediction works as a “distribution-free” method.
The paper “A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification” (Angelopoulos and Bates, 2021) provides a practical and accessible introduction to this idea.
It explains how conformal prediction can be applied to basically any machine learning model while providing rigorous statistical guarantees.
The Problem with Standard Predictions
Traditional machine learning models output a single prediction.
For example:
A classifier outputs a label: “cat”
A regression model outputs a number: house price = $450,000
However, these predictions do not express uncertainty or how confident the model is in its predictions.
Consider a medical diagnosis model. A prediction of “disease present” might be correct 95% of the time or only 55% of the time. Without a measure of uncertainty, it is difficult to determine how much trust should be placed in the model.
Standard approaches attempt to address this using:
Bayesian models
Bootstrapping
While useful, these methods often rely on assumptions about the data distribution or the model itself. When those assumptions fail, the uncertainty estimates may become unreliable.
