Drew Purves

I am a research scientist at Google DeepMind, where I am drawing on my experience in computational ecology, simulation modelling, machine learning and software development to help create Artificial General Intelligence.

My interest in computers began in Christmas 1984 when, at age seven, I was lucky enough to receive the Commodore 64 that sits in my office at home today. Although the games that you could buy on cassette were fun, I preferred to program my own in BASIC. Worm-but-in-a-maze. Tetris-but-top-down. Sokoban-but-I-programmed-it-myself. Funnily enough my friends and family never loved these games as much as I did :)

My interest in ecology began later on when a sharp-eyed biology teacher, Richard Hepell, noticed the Artificial Life simulations I was running on the biology department’s 386s. Richard handed me a copy of the standard ecology undergraduate textbook (‘Ecology’ by Begon, Harper and Townsend), opening my eyes to the fascinating dynamics of ecosystems, and our attempts to understand them through theory. For this reason I applied to study natural sciences, with a focus on all things evolutionary and ecological, at Cambridge University. Nobody in my family had ever been to university, so I had no idea about how to navigate what was, genuinely, a whole new world for someone with my background. But I figured that, surely, it would help if I turned up at the interview with some research results? So I brought along some charts showing punctuated equilibrium in simulations of the evolution of foraging strategies. Looking back now, I think this probably did help me to get in ;)

Studying ecology at Cambridge was a privilege, which excellent teaching that was, crucially, combined with many field trips to woodlands, fens, the coast, islands, and even to a the tiny village in Andalucia which has since become a spiritual second home for my family. These trips reinforced how interesting nature was, but also reinforced how far we were from having the kind of rigorous, mathematical understanding of ecosystem that we have long taken for granted in other areas. So I opted to my PhD with a leading theoretical ecologist, Richard Law, at the University of York, who was one of the people trying to create a firm mathematical and computational foundation for ecology. During that PhD I learned to go beyond just creating interesting simulations, and instead use them rigorously to address scientific questions. I also learned a little about how to constrain ecological models against data, in order to create simulations that could actually capture real, observed dynamics -- something that still gives me a buzz to this day.

Next, I was lucky enough to land a five-year postdoc at Princeton University, with another leading ecologist, Steve Pacala. Steve’s group was at the very cutting edge in terms of constraining ecological models against data. During that postdoc I learned about that magical thing The Likelihood – suddenly, stats made sense! – and Bayesian approaches. I ended up developing my own adaptive MCMC algorithm (now available under the name Filzbach) which, over the following years, enabled myself and others to fit all kinds of really nasty models to all kinds of data, both at Princeton and later on at Microsoft Research. I focussed in particular on forests, which are an amazing type of ecosystem to study for many reasons, not least the amazing amount of public data available. In the years 2001-2006, I was routinely fitting highly non-linear simulation models to over a million measurements of real trees. On a more general level, I also learned that problems that were highly important from a societal perspective, like climate change, could also be intellectually and scientifically challenging. Obvious in retrospect, but novel at the time.

Just when it was starting to look like I might land a permanent job in academia, a truly unique opportunity arose at Microsoft Research, which wanted to build a Computational Ecology and Environmental Science group (CEES), under Stephen Emmott and Rich Williams. Strangely, they wanted to do this at their centre in Cambridge, UK. It was a no-brainer. Thanks in large part to a strong steer from Stephen and Rich, for the next eight years myself and the other members of the team pursued a two-part mission to carry out, and publish, research that pushed the boundaries of computational ecology; whilst creating new software tools to help others do the same. I became head of the group early on, when Rich moved on to pursue other challenges. Our team never grew above ten full-time staff (permanent scientists, postdocs, developers), but nonetheless we published well over 200 research papers (including at least five in Science and Nature), many of which represented significant advances in the field (again, see below). We also generated exciting early data science software prototypes that played significant roles within Microsoft as a whole. Looking back though, the thing that I am most proud of about CEES is that the CEES family – from the permanent scientists, to the developers, to the postdocs and PhD students that we mentored, to the many interns that we hosted – have landed amazing, exciting jobs in academia, NGOs, businesses, and government which are clearly related to their experiences within the group. On a daily basis they are drawing on that experience to address everything from forest ecology, to remote sensing, to biology and medicine, to conservation, to the internet of things, to data visualization, to national security, to applications of machine learning in the wider economy. This is the true legacy of CEES, and it is thrilling to watch it play out.

As for me, I recently found myself faced with another no-brainer – an invite from Demis Hassabis to help Google DeepMind create human-level Artificial General Intelligence. How could anyone turn down the chance to be involved in the most important technological transition in human history? But also, this kind of AI is needed to address major global challenges, as I know from personal experience thinking about them for many years (these problems are so complex that is not obvious we can solve them without Artificial General Intelligence). A year or more before I joined DeepMind I had realized that it was the most exciting company on the planet. What I didn’t know was that I might have a role to play within it: remember, I don't have a background in neuroscience, reinforcement learning, or deep learning! Luckily, it turned out there was a role. Indeed, a surprisingly ambitious and long-term one that will require me to draw on all my experience so far, and to learn about and move into several fields, all of which are new to me, many of which are new to the world.

Please note the mixture of luck, encouragement and patronage that has shaped my career to date. I have always tried to pass on some of the latter two wherever possible, for example by taking up external positions as Treasurer of the British Ecological Society, Affiliated Lecturer at Cambridge University, and Honorary Reader at UCL, where I help to mentor two PhD students.