Vanilla Bayesian Optimization Performs Great in High Dimensions

Watch a 35-minute AutoML Seminar presentation exploring how vanilla Bayesian optimization (BO) performs effectively in high-dimensional problems. Learn why the perceived limitations of BO in high dimensions may be primarily due to assumptions about objective complexity rather than inherent algorithmic limitations. Discover how a simple modification - scaling the Gaussian process lengthscale prior with dimensionality - enables standard BO to outperform specialized algorithms on tasks with thousands of dimensions. Follow along as speaker Carl Hvarfner challenges common beliefs about BO's capabilities and demonstrates how existing high-dimensional BO approaches can be viewed through the lens of model complexity. Gain insights into how adjusting complexity assumptions can dramatically improve optimization performance without requiring complex algorithmic modifications.