Zoubin Ghahramani is an Iranian-British computer scientist and statistician who has been a leading figure in probabilistic machine learning for three decades. As a professor at the University of Cambridge and later as vice president of research at Uber, Ghahramani has advanced Gaussian process methods, Bayesian nonparametric models, and approximate inference algorithms, consistently advocating for the position that machine learning systems should quantify their uncertainty rather than produce only point predictions.
Life and Career
Born in Tehran, Iran. Moves to the United Kingdom as a child and later studies computer science and cognitive science.
Earns his Ph.D. from MIT, working on unsupervised learning and mixture models under the supervision of Michael I. Jordan.
Develops influential work on sparse Gaussian processes, Bayesian optimization, and automatic model discovery at the Gatsby Computational Neuroscience Unit and later at Cambridge.
Co-develops the infinite latent feature model (Indian Buffet Process) with Thomas Griffiths, extending Bayesian nonparametrics beyond clustering to feature allocation problems.
Launches the Automatic Statistician project, which uses Bayesian model selection to automatically discover and describe statistical structure in data.
Elected Fellow of the Royal Society in recognition of his contributions to machine learning and statistics.
Gaussian Processes
Ghahramani has been one of the most influential advocates and developers of Gaussian process (GP) methods in machine learning. A Gaussian process defines a distribution over functions, where any finite collection of function values has a joint Gaussian distribution characterized by a mean function and a covariance (kernel) function. GPs provide a fully Bayesian approach to regression and classification, producing not just predictions but calibrated uncertainty estimates.
Predictive Distribution f* | X, y, x* ~ N(μ*, σ*²)
μ* = k(x*, X)[k(X, X) + σ²I]⁻¹ y
σ*² = k(x*, x*) − k(x*, X)[k(X, X) + σ²I]⁻¹ k(X, x*)
Ghahramani's contributions to GP methods include work on sparse approximations that reduce the cubic computational cost to make GPs practical for larger datasets, multi-output Gaussian processes for modeling correlated outputs, and the use of GPs for Bayesian optimization, where the goal is to find the maximum of an expensive-to-evaluate function with as few evaluations as possible.
One of Ghahramani's most ambitious projects is the Automatic Statistician, which uses Bayesian model selection over a compositional grammar of Gaussian process kernels to automatically discover structure in time series and other data. The system searches over kernel compositions (sums and products of base kernels) using marginal likelihood as the model selection criterion, and generates natural-language descriptions of the discovered patterns. This project represents a vision of Bayesian inference not just as a tool for fitting specified models, but as a framework for automated scientific discovery.
Bayesian Nonparametrics
Ghahramani has made fundamental contributions to Bayesian nonparametrics beyond Gaussian processes. His work with Thomas Griffiths on the Indian Buffet Process (IBP) introduced a Bayesian nonparametric prior for latent feature models, analogous to how the Chinese Restaurant Process provides a prior for clustering. The IBP allows each data point to possess an unknown number of latent binary features, with the total number of features growing logarithmically with the dataset size. This flexibility makes it suitable for problems where the dimensionality of the latent representation is itself unknown.
Legacy
Ghahramani's career embodies the principle that uncertainty is not a nuisance to be minimized but information to be exploited. His work on Gaussian processes, Bayesian nonparametrics, and approximate inference has provided the probabilistic machine learning community with both theoretical foundations and practical tools. His advocacy for probabilistic approaches has helped ensure that uncertainty quantification remains central to machine learning research, even as the field has increasingly focused on deep learning.
"Machine learning is fundamentally about making predictions under uncertainty. A system that cannot tell you how uncertain it is cannot be trusted." — Zoubin Ghahramani