Conformal Prediction
Conformal prediction is a model-agnostic framework for generating prediction intervals or sets with guaranteed coverage probabilities, addressing the crucial need for reliable uncertainty quantification in machine learning. Current research focuses on improving the efficiency and robustness of conformal prediction methods, particularly for non-i.i.d. data (e.g., time series, graphs) and biased models, exploring techniques like adaptive conformal prediction, weighted conformal prediction, and score refinement to achieve this. These advancements are significant because they enhance the trustworthiness and applicability of machine learning models in high-stakes domains such as healthcare, autonomous systems, and financial modeling, where reliable uncertainty estimates are paramount for safe and informed decision-making.
Papers
A novel Deep Learning approach for one-step Conformal Prediction approximation
Julia A. Meister, Khuong An Nguyen, Stelios Kapetanakis, Zhiyuan Luo
MAPIE: an open-source library for distribution-free uncertainty quantification
Vianney Taquet, Vincent Blot, Thomas Morzadec, Louis Lacombe, Nicolas Brunel