Abstract
Conformal prediction (CP) is a distribution-free framework for achievingprobabilistic guarantees on black-box models. CP is generally applied to amodel post-training. Recent research efforts, on the other hand, have focusedon optimizing CP efficiency during training. We formalize this concept as theproblem of conformal risk minimization (CRM). In this direction, conformaltraining (ConfTr) by Stutz et al.(2022) is a technique that seeks to minimizethe expected prediction set size of a model by simulating CP in-betweentraining updates. Despite its potential, we identify a strong source of sampleinefficiency in ConfTr that leads to overly noisy estimated gradients,introducing training instability and limiting practical use. To address thischallenge, we propose variance-reduced conformal training (VR-ConfTr), a CRMmethod that incorporates a variance reduction technique in the gradientestimation of the ConfTr objective function. Through extensive experiments onvarious benchmark datasets, we demonstrate that VR-ConfTr consistently achievesfaster convergence and smaller prediction sets compared to baselines.