Statistical decision making on non-euclidean spaces

背景：随着大数据时代的到来，数据从未像现在这样以如此大的数量和速度获得。与此同时，机器学习模型现在正在进入对成本和安全敏感的应用，从药物发现、自动驾驶或医疗决策等。在做出决策时，对不确定性和风险进行良好的衡量，对于机器学习工具在这些环境中的成功和采用至关重要。目的：DPhil项目的目的是利用大数据集诱导的几何特性来改进安全关键任务中机器学习模型中的不确定性量化。与EPSRC战略保持一致：这款DPhil与EPSCR研究的人工智能技术领域非常一致。研究方法的新颖性：这项研究将在微分几何、随机流形和贝叶斯机器学习之间建立跨学科联系，以获得新的见解。参与的公司或合作者：此DPhil项目由AWS学生与牛津-新加坡人机协作(HMC)计划合作提供支持。

Context: With the advent of the big data era, data has never been available in such volume and velocity. In parallel, machine learning models are now entering cost sensitive and safety sensitive applications ranging from drug discovery, autonomous driving or medical decision making among others. Having well calibrated measures of uncertainty and risk while decisions are made is paramount for the success and adoption of machine learning tools in these settings. AIMS: The aim of this DPhil project is to exploit the geometry induced by large datasets to improve uncertainty quantification in machine learning models in safety critical tasks. Alignment to EPSRC strategies: This DPhil is well aligned with the Artificial intelligence technologies area of EPSCR's research. Novelty of the research methodology:This research will establish cross-disciplinary links between differential geometry, random manifolds and bayesian machine learning for novel insights. Companies or collaborators involved: This DPhil project is supported by the AWS Studentship in collaboration with the Oxford-Singapore Human Machine Collaboration (HMC) initiative.