Uncertainty quantification for Bayesian Neural Networks

本计画的目的是设计有效的马尔可夫链蒙地卡罗方法来估计贝氏神经网路的未知参数。神经网络的参数数以千计，目前的MCMC方法与这种规模的问题斗争。除了新的MCMC方法外，我们还将验证实际NN学习问题的性能，例如强化学习和工程和应用数学的其他合适应用，并在理论上验证性能。鉴于NN目前在应用科学中的广泛使用，这项工作在许多领域都具有潜在的影响力。

The aim of this project is devise efficient Markov Chain Monte Carlo methods to estimate the unknown parameters of a Bayesian Neural Network (NN.) NNs have parameters that number many thousands and present MCMC methods struggle with problems of this size. In addition to new MCMC methodology, we will verify performance on practical NN learning problems, e.g. in Reinforcement Learning and other suitable applications from engineering and applied mathematics, and also verify the performance theoretically where we can. This work is potentially impactful in many areas given the current wide spread use of NN in the applied sciences.