Your second-worst review is probably all you need
We analyze nine years of ICLR reviews and find that Reviewer 2 is not the problem.
The Upper Confidence Bound in Disguise
We show that UCB is a quantile of the predictive distribution in disguise, elicitable by the pinball loss even though no scalar loss can elicit the moment formula directly. The …
Gaussian Process Regression in 9,000 Google Sheets Cell Formulas
We implement a fully functional Gaussian Process regression pipeline — Cholesky decomposition, posterior predictions, and gradient-based hyperparameter optimization — in pure …
Empirical Gaussian Processes
We study Empirical GPs, a principled framework for constructing flexible, data-driven Gaussian process priors. By estimating mean and covariance directly from a corpus of …
Jihao Andreas Lin
Ax: A Platform for Adaptive Experimentation
We present Ax, an open-source platform for adaptive experimentation built on BoTorch. Off the shelf, Ax achieves state-of-the-art performance across a wide range of synthetic and …
Miles Olson
Probabilistic Machine Learning in the Age of Deep Learning: New Perspectives for Gaussian Processes, Bayesian Optimization and Beyond (PhD Thesis)
We explore the intersection of deep learning and probabilistic machine learning, addressing limitations of Gaussian processes in comparison to neural networks and proposing …
📄 One paper accepted to ICML 2023
Our paper "Spherical Inducing Features for Orthogonally-Decoupled Gaussian Processes" was accepted to ICML2023 as an Oral Presentation!
•
1 min read
Spherical Inducing Features for Orthogonally-Decoupled Gaussian Processes
We introduce spherical inter-domain inducing features that yield more flexible, data-dependent basis functions for orthogonally-decoupled GP approximations, narrowing the …
Efficient Cholesky decomposition of low-rank updates
We give a short and practical guide to efficiently computing the Cholesky decomposition of matrices perturbed by low-rank updates.
Batch Bayesian Optimisation via Density-ratio Estimation with Guarantees
We extend BORE to the batch setting and establish theoretical convergence guarantees for parallel Bayesian optimization.
rafael-oliveira