Speaker
Description
Gaussian Processes are the most popular and accurate modelling technique for estimating uncertainty in a system. However, they struggle to scale with high dimensional data and large datasets that are often required to train surrogate models of accelerator components. Typical alternatives include deep ensembles, Monte-Carlo Dropout and quantile regression. However, each of these methods has shortfalls, resulting in either a highly overconfident or underconfident estimate of the uncertainty in the model. This work will explore whether Hyper-GANs, generative models that learn distributions of neural network parameters, can be used to address these problems and provide accurate estimates of uncertainty for high and multi-dimensional outputs for surrogate models of accelerators. We will use a model of the ISIS MEBT as a case study and compare the quality of the uncertainty predictions against predictions using quantile regression.
