Brendan O'Donoghue

Brendan O'Donoghue

Director of Research, Google DeepMind

Ph.D., Stanford University · London, UK

Email Google Scholar GitHub Twitter

I lead a research team at Google DeepMind in London, where I've worked since 2015. My research sits at the intersection of generative AI, reinforcement learning, and optimization. Lately I've been focused on diffusion models for language, and I lead the team behind Google's experimental Gemini Diffusion model, which I spoke about in this interview. Previously I worked on reinforcement learning, especially Bayesian approaches to the exploration/exploitation trade-off, which resulted in algorithms like K-learning and the Uncertainty Bellman Equation.

Outside DeepMind I'm probably best known for my work in optimization, particularly SCS, the Splitting Conic Solver—an open-source solver for large-scale convex optimization that I wrote and still maintain, which gets downloaded a few million times a month.

Before DeepMind I led a machine learning team at Quantcast in San Francisco and London, working on large-scale ad targeting and real-time bidding. I did my Ph.D. at Stanford University with Stephen Boyd, on designing control policies via convex optimization. Before that, I read information and computer engineering at Cambridge.

Education

Stanford University

2007 – 2013 · Stanford, CA

M.S. & Ph.D., Electrical Engineering

Advisor: Prof. Stephen Boyd

Thesis: Suboptimal Control Policies via Convex Optimization, 2013

Gonville & Caius College, Cambridge

2003 – 2007 · Cambridge, UK

B.A., M.A. (1st Class Hons.), M.Eng. (Distinction)

Computer & Information Engineering

Cambridge–MIT Exchange Fellow, 2005/2006

Awards

Software

SCS — Splitting Conic Solver

Author & primary maintainer

  • Free, open-source numerical package for solving large-scale convex cone problems — LP, QP, SOCP, SDP, and more.
  • Written in C with Python, MATLAB, Julia, and R interfaces.
  • Downloaded almost 3 million times a month; in widespread use across academia and industry.

Other open-source numerical linear algebra, optimization, and control libraries:

QTQP “Cutie QP”—a quadratic programming solver FirstOrderLp.jl First-order methods for large-scale linear programming, in Julia AA Type-I and type-II Anderson acceleration for fixed-point iterations apg Accelerated proximal gradient solver with adaptive restart apgpy Accelerated proximal gradient method, in Python osc Operator splitting for control py-mkl-pardiso Python bindings for Intel MKL's PARDISO sparse linear solver macldlt Python bindings for the Apple Accelerate LDLᵀ factorization scs-python Python interface to the SCS solver scs-matlab MATLAB interface to the SCS solver

Publications

Journal Articles 13

Conference Articles 21

Other 8

Ph.D. Thesis