报告题目：Generic Inference for Gaussian Process Models

报告人： 陈泽汛 博士 （英国爱丁堡大学）

报告时间：2024年8月30日下午 15:00-16:00

报告地点：腾讯会议（215-121-098）

报告摘要：Gaussian Process (GP) modelling is pivotal in machine learning due to its inherent strengths, yet its principled and adaptable nature presents challenges in probabilistic inference, particularly regarding scalability with large datasets. This research talk delves into mainstream inference methods for GP models alongside advanced concepts. Firstly, Variational inference with inducing variables, grounded in latent GP models, facilitates multiple output and latent function support without demanding intricate conditional likelihood knowledge. By evaluating the likelihood as a black-box function and employing a Gaussian mixture as the variational distribution, efficient estimation of the evidence lower bound and gradients is achieved through sampling from univariate Gaussian distributions. Scalability to large datasets is realized through augmented priors via the inducing-variable approach, bolstered by parallel computation and stochastic optimization. Moreover, low-rank approximation via pivoted Cholesky decomposition expedites GP modelling inference, a technique embraced by prominent libraries like GPyTorch, demonstrated in their work on Blackbox Matrix-Matrix Gaussian Process Inference with GPU Acceleration. Finally, advanced topics such as variational inference for multivariate GP modelling with heterotopic sparse data and constrained GP models will be briefly touched upon.

报告人简介：陈泽汛，英国爱丁堡大学商学院助理教授，研究方向包括非参贝叶斯统计，概率化统计学习方法，时间序列分析，算法公平性与信用评级，多维度人口流动模型，机器学习方法在人口流动与城市规划上的应用，高斯过程机器学习在金融期权与市场风险上的应用等等。陈博士2013年本科毕业于山东大学数学与应用数学专业，2017年博士毕业于英国莱斯特大学数学系，研究方向是高斯过程机器学习算法及其推广与应用。博士毕业后曾在英国苏塞克斯大学作为博士后研究员参与英国EPSRC项目，提出高斯过程分类器的公平性改进。之后又在英国埃克塞特大学作为博士后参与英美联合US Army Research的项目，研究内容包括时间，空间与社交数据下人口流动问题的建模等。陈博士曾在各类期刊会议上发表过一系列高水平文章，其中包括Nature Communications，Neural Computation and Application，Measurement等等，目前是英国皇家统计协会会员, 英国皇家统计协会会员青年统计学家分会组委会秘书，英国数学与应该数学协会会员，国际统计协会会员，并担任各类数学统计，量化金融，机器学习，工程机械，人口流动，交通运输方面期刊的审稿人，如IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)，Association for the Advancement of Artificial Intelligence (AAAI)，Transportation Research Part C: Emerging Technologies，The European Journal of Finance， Journal of Forecasting等。