Practical uncertainty in neural networks
Ethan Goan · Queensland University of Technology · 2023
The adoption of machine learning technologies has grown considerably as the predictive performance of deep learning models continues to improve. Application of these systems for real-world scenarios requires not only raw predictive power, but also informative uncertainty information. Quantifying uncertainty comes at the expense of increased computation and time, and as a result most models do not aim to communicate any such information. This thesis addresses this by proposing practical means to quantify uncertainty in offline scenarios, real-time scenarios, and within existing neural networks not designed within a probabilistic framework.