Quantifying Uncertainty in Deep Learning: Methods, Challenges, and Future Directions

Uncertainty quantification is essential in deep learning to ensure reliability, robustness, and interpretability in safetycritical applications such as healthcare, autonomous systems, and defense. While deep learning has achieved success across various applications, its black-box nature raises concerns about trust in its predictions, especially in safety-critical systems. This paper categorizes the sources of uncertainty in deep learning as aleatoric and epistemic uncertainties and classifies existing uncertainty evaluation methods into probability-based and calibrationbased methods. We also review the state-of-the-art techniques, which are divided into Bayesian, non-Bayesian, and ensemblebased methods, and discuss their advantages, challenges, and practicality. The main challenges in quantifying uncertainty, including computational efficiency, robustness to out-of-distribution data, and the need for standardized evaluation metrics, are also presented. Finally, we discuss future research directions to enhance the trustworthiness of deep learning models through improved uncertainty quantification.