A Literature Survey on Estimating Uncertainty in Deep Learning Models: Ensuring safety in Intelligent Systems

Abstract

Popular Deep learning models suffer many drawbacks such as making wrong predictions with great confidence, lack of uncertainty estimation capability, and failure in real-time scenarios. The main reason for the uncertainty is due to the large gap between how neural networks are trained in practice and how they are evaluated in deployment. When it comes to safety-critical applications, it is very important to build confidence in the output that is obtained. A well-calibrated uncertainty quantification method can tell whether a model is confident in its predictions or not. This survey focuses on techniques used for uncertainty quantification in deep learning.