Abstract

Uncertainty quanti cation (UQ) in neural networks has become a critical component of trustworthy machine learning systems, particularly in high stakes regression applications such as healthcare, nance, and autonomous systems. This paper presents a comprehensive comparative study of ve prominent uncertainty quanti cation methods for neural network regression tasks: deterministic neural networks (baseline), Monte Carlo (MC) Dropout, Bayesian neural networks with variational inference (BNN-VI), deep ensembles, and conformal prediction. We evaluate these methods across multiple dimensions including predictive accuracy, calibration quality, out- of distribution robustness, computational eciency, and implementation complexity. Through systematic experiments on synthetic and real-world datasets with varying noise levels, we and that deep ensembles achieve the best overall predictive performance, while conformal prediction provides the most reliable coverage guarantees under distribution shift. MC Dropoutoers the best balance between computational cost and uncertainty quality, making it particularly suitable for resource-constrained applications. Our fndings provide practical guidance for practitioners in selecting appropriate UQ methods based on their specic requirements and constraints.