This tutorial is on Gaussian Processes for Regression: Models, Algorithms, and Inference. It is taking place online as a webinar run by the Committee on International Relations in Statistics (CIRS) within the American Statistical Association (ASA). It is running on April 9 from 12noon to 1:30pm in Eastern Time. See this link for the latest versions of all tutorials.
The instructor is Prof. Tamara Broderick. Contact information can be found here.
Abstract: In regression tasks, we aim to use observed data to predict a continuous outcome as a function of a set of inputs. For instance, we might want to understand how ocean current (a velocity vector field) varies as a function of space and time given GPS data from buoys---or how the accuracy of a machine learning algorithm varies as a function of its adjustable parameters given the results of some runs of the algorithm. It is common for the relationship between outputs and (continuous) inputs to be highly nonlinear but smooth; for observations to be sparse, noisy, and unevenly spaced; and for uncertainty in our predictions to be of interest. We will discuss how Gaussian processes offer a useful and popular framework to address these concerns. We will explore both benefits and pitfalls of Gaussian processes as commonly used in practice. Though we will review key facts, it will be helpful for audience members to have basic familiarity with (1) Bayesian data analysis and its goals and (2) both univariate and multivariate Gaussian distributions.
Slides:
Code for demos: