Computational Aspects of Deep Gaussian Processes in Data Science

数据科学中深度高斯过程的计算方面

• 批准号: 2616566
• 金额: --
• 依托单位: Heriot-Watt University
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2021
• 资助国家: 英国
• 起止时间: 2021 至 无数据
• 项目状态: 未结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-2616566
• 关键词: Computational; Aspects; Deep; Gaussian; Processes

My project's area of research is the investigation of continuous-time/space models in data science. In particular, I study deep Gaussian processes. Applications for this modelling tool are abundant, thanks to its capacity of representing multi-scale behaviour. Examples where this property is of great importance include climate modelling and image analysis. Such applications give rise to the necessity of solving "standard" machine learning tasks such as regression and inference problems that are based on (deep) Gaussian processes. As of today, sampling from such models is computationally very expensive. My goal in this project is thus to improve existing sampling techniques and to develop new strategies for generating samples in a more efficient way. At a later stage, I would like to employ deep Gaussian process priors in Bayesian inference. To this end, I study efficient Markov chain Monte Carlo methods.

我的项目的研究领域是数据科学中的连续时间/空间模型的研究。特别是，我研究深高斯过程。这种建模工具的应用是丰富的，这要归功于其表示多尺度行为的能力。这一特性非常重要的例子包括气候建模和图像分析。这些应用程序需要解决“标准”机器学习任务，例如基于（深度）高斯过程的回归和推理问题。到目前为止，从这样的模型中采样在计算上是非常昂贵的。因此，我在这个项目中的目标是改进现有的采样技术，并开发新的策略，以更有效的方式生成样本。在稍后的阶段，我想在贝叶斯推理中使用深度高斯过程先验。为此，我研究有效的马尔可夫链蒙特卡罗方法。