The International Conference on Kernel-Based Methods and its Applications (ICKBM) is being held at Xi’an Jiaotong Liverpool University in Suzhou, PRC this year. From October 10-14, experts in kernels methods from approximation theory, statistics, machine learning and other fields came together to learn about progress from the past year. One of the goals of this conference was to identify how the different communities that use kernel methods can come together to share knowledge and collaborate.
SigOpt research engineer Michael McCourt gave an invited talk at this session and will be speaking on recent progress in kernel methods for multiobjective optimization. He last spoke at this conference series in 2017, and is excited to return to see old friends and meet new friends. The details of his talk are below.
Title: Kernel-based Bayesian Optimization of Multiple Objectives
Abstract: Kernels provide an effective tool for approximating scattered data in arbitrary dimensions; they appear in the functional analysis, spatial statistics and machine learning communities. In this talk, we will review how kernel approximations appear in the context of Gaussian processes/random fields, and how such approximations give rise to the Bayesian optimization framework: a sample-efficient strategy for black-box optimization. We will then introduce our adaptation of this strategy for identifying the Pareto frontier of multiple competing objectives, as well as variations which enable users to interactively focus on the most desired components of the frontier in a more sample-efficient fashion. This will be demonstrated on examples from deep learning and materials science.
Slides: Google slides link
After ICKBM concluded, Mike stopped by Tokyo to see visit colleagues at ANA and at RIKEN. While visiting Ha Quang Minh at RIKEN’s Advanced Intelligence Project, Mike gave a talk introducing Bayesian optimization and presenting different applications that SigOpt has considered over the previous year. The details of his talk are below, as well as a picture of him and Dr. Minh.
Title: Kernel-based Bayesian Optimization of Multiple Objectives
Abstract: Bayesian optimization is a valuable tool to efficiently search for the optimum of a noisy, black-box function. This talk will introduce the topic, using Gaussian processes to power the statistical modeling powering the sequential decision making process of optimization. Then, variations on the standard formulation will be discussed to handle complications arising in various applications. Examples from machine learning and materials sciences will be presented.
Slides: Google slides link
