A novel deep learning approach for one-step conformal prediction approximation

Deep Learning predictions with measurable confidence are increasingly desirable for real-world problems, especially in high-risk settings. The Conformal Prediction (CP) framework is a versatile solution that guarantees maximum error rate given minimal constraints [1]. In this paper, we propose novel conformal loss function approximates the traditionally two-step CP approach single step. By evaluating and penalising deviations from stringent expected output distribution, model may learn direct relationship between input data p-values. We carry out comprehensive empirical evaluation to show our function’s competitiveness seven binary multi-class prediction tasks on five benchmark datasets. On same datasets, achieves significant training time reductions up 86% compared Aggregated (ACP, [2]), while maintaining comparable approximate validity predictive efficiency.