IN THE run up to the match, Lee Sedol was feeling confident. He was about to play an artificial intelligence that had been trained to play the board game Go. But as one of the world’s best players of the game himself, Lee thought he would easily win. “I thought AlphaGo was based on probability calculation and it was merely a machine,” he said at the time. Even after he lost the first game of the match, Lee believed it was just because AlphaGo had made no mistakes. Then, in the now infamous move 37 of the second game, the AI seemed to be rewriting the rules of Go and played a move no human would have dreamed of. Lee, who ultimately lost the match 4-1, was dumbstruck: “This move was really creative and beautiful,” he said.
This contest marked a pivotal moment in the development of AI, and Demis Hassabis was one of the main people responsible for it. In 2010, he co-founded the research company DeepMind and began working on AIs that could play games better than people. At the time, Go was considered too hard for artificial intelligence to master. It has more possible moves than there are atoms in the universe. Yet the win over Lee in 2016 catapulted DeepMind – and Hassabis – to worldwide fame. AlphaGo’s victory was the biggest moment in AI since IBM’s Deep Blue defeated world chess champion Garry Kasparov in 1997.
Since then, Hassabis’s firm, which is now a subsidiary of Alphabet, Google’s parent company, has been honing its algorithms and looking beyond games. In November, its new AI, AlphaFold, cracked the devilishly hard problem of predicting protein structures. Proteins are strings of amino acids that fold themselves into incredibly intricate shapes and serve as the biochemical tools that make living things tick. Working out the shape of a protein usually requires years of experiments. AlphaFold does it without any. Hassabis believes this is just the beginning of what AI can do for science.
Advertisement
Timothy Revell: What are you trying to do at DeepMind?
Demis Hassabis: Our vision is to solve artificial intelligence. What we mean by that is to fundamentally understand intelligence and recreate it artificially, and then use that as a tool to help us understand the world and make a positive impact. Part of that is about advancing the pace of scientific discovery. The Hubble telescope allows us to see further into the cosmos, and that’s how I see AI, as a generalised Hubble telescope that can help us make a lot more progress in many areas of science.
How do you know what sort of intelligence to try to encode in AI?
We take a lot of inspiration from neuroscience. I’m both a neuroscientist and a computer scientist; my first degree was in computer science, my PhD was in neuroscience. We’ve tried to bring both together, being inspired by the capabilities the human mind has and how it operates.
What are the most exciting things you are working on?
There are so many projects going on. We try to identify problems in science that have a cross section of factors that we like. There has got to be a clear objective to optimise, and some training data. The other thing we look for is impact: is this a challenge that, if we solved it, would unlock whole new branches of scientific endeavour? Protein folding has all those characteristics. We’re also looking at quantum chemistry, pure mathematics and areas of physics and materials science.
You had a big success recently with proteins (see “The protein problem“). Why did you choose to work on them?
I think proteins are incredibly exquisite. They are like little bio-machines, transporting nutrients around your body. They are switches, they are motors, they are little factories. From a mathematical and computer science point of view, it’s intriguing to see that biology at that level starts looking digital and almost computational code-like. There are only about 100,000 proteins we know the structure of. That’s a tiny fraction of the 200 million or so proteins that are out there in nature.
If you could solve protein folding, then you could probably accelerate drug discovery. You might start unlocking some of the secrets about how life works much faster than you can with experiments, which are slow, painstaking and difficult. You could better understand disease. I think one of the big transformations in science over the next 10 to 20 years is more and more biology being understandable in this computational way.
What can AI help us find out about biology that traditional approaches can’t?
A lot of science’s low-hanging fruit that could be addressed by an Einstein-like lone genius was done in the early parts of last century. Now we are left with more complex systems that don’t yield themselves to simple laws. Up until now, biology has been regarded as too chaotic and too complex to apply physics approaches to it, but perhaps computer science and AI can help bridge that gap. That’s exactly where I think the AI sweet spot is.
You could think of it as having the world’s best research assistant on tap for you. They can find patterns, pull together disparate bits from different papers and find connections between them. It would still require the human scientist to come up with the goal and why that is an important goal. A lot of work that PhD students do is not the most fascinating or creative. If you had automated systems that take out some of the drudgery, you could free up the students to think about creative solutions.
One criticism of systems like AlphaFold is that they can become very good at one task but they can’t do anything else. Do we need a fundamental rethink of AI to make it more general?
No, but I think it’s a reasonable criticism. We developed AlphaZero, which can play any two-player game with no modification, but it has to learn from scratch each time. Its knowledge doesn’t transfer. Some people would say: that’s fine, you’ve got a general algorithm. Others find it unsatisfactory.
I would like that if an AI learns chess, it shouldn’t be a complete zero at Go; something should transfer over. Humans do this extremely well because obviously we cannot afford to start from scratch at every task. That’s called transfer learning and no one’s cracked it yet with an AI. It’s a super-active area of research and something I’m personally working on. Perhaps there are a couple more modules we need, like episodic memory and attention.
How would extra modules help?
It seems like the brain has at least two computational modes. You’ve got a cortex, which is quite slow at learning and needs quite a lot of examples, but it’s very stable. On the other hand, you’ve got a hippocampus, which is what I studied for my PhD. It’s a critical part of your brain and it is a superfast learner. That’s how you remember what you had for lunch yesterday; it’s your hippocampus doing that.
You might ask why don’t we just have one giant hippocampus. Fast seems better. Actually, the rat, which is pretty clever for its size, has a brain that is basically one ginormous hippocampus. But the problem is – this is my conjecture – if you have just a fast learner, and you learn something new, it’s very volatile. By learning how to ride a bike, you might wipe out your piano-playing expertise. Because it’s got to be fast, you don’t have much control over what gets overwritten.
That doesn’t sound ideal…
You don’t want your whole brain to be that volatile if you’re trying to build knowledge on top of knowledge. But on the other hand, you don’t want to have to encounter the lion at the watering hole at dusk more than once before you learn that’s really dangerous. So you need both systems and, presumably, some kind of transfer function between the two. That’s partially what happens when we sleep.
Does that mean AI needs a form of sleep?
Potentially. We do work on that. If you look at our early work getting AI to play Atari video games, memory replay is one of the key reasons it works. Not only does our AI play the game, it replays key aspects of what it did back to itself. You could call it sleep, but, obviously, it’s not actually sleeping. It’s in an offline mode.
On the progress bar of AI, where you have a robot vacuum cleaner at one end and human intelligence at the other, where are we now?
I would say somewhere about halfway. But there are different schools of thought. Some people think we as a community with deep learning and reinforced learning are not on the right track, so there’ll be some brick wall. Others on the pro side say, look, we have everything we need. It’s now just a question of scaling up: more computers, more data and we’ll eventually pop out a human-level general AI.
I think both camps are wrong. My view is that what we have invented so far is on the right track and is super-useful. But I don’t think we have all the ingredients. There will be further leaps, maybe only a handful, that are required.
Has DeepMind done any research on covid-19?
It came in a bit too early for AI to be of much real help for this particular outbreak. God forbid, if there’s another one in a few years’ time, I would expect AI to play a bigger part.
“AI is like a generalised Hubble telescope that helps us see further in many areas of science”
The genetic sequences of the proteins in the virus were sequenced early on by Chinese researchers and some of the structures had been experimentally determined. But for a dozen or so of the proteins involved, we did not know the structures. At the time, we already thought our AlphaFold system was pretty accurate but it hadn’t been battle-tested, so we didn’t want to make big claims about the structures of those proteins. But we also thought we might be sitting on something that could be useful for the scientific community. So we immediately
Talking of existential threats, can AI help us with climate change?
One of our most famous pieces of work we did for Google was to control the cooling systems in the data centres that everybody uses when you watch YouTube and so on. We managed to save 30 per cent of the power that those data centres use by more efficiently switching on and off the fans and the pumps and all the other amazing equipment that is in these data centres. So that’s obviously huge for cost-saving, but also for the environment.
You should be able to apply the same types of optimisation to other buildings or even a national grid. That would save a lot of energy. In the future, AI might contribute to things like material design or protein design, so that we have proteins that can break down waste plastics or create renewable biofuel.
Where will AI have the most impact in the next 10 years?
I hope to see dozens of breakthroughs on grand scientific challenges, and then many products, services and advances that are enabled because of that. I would be very surprised if that doesn’t happen in the next 10 years. We hope to be a big part of all that in a lot of different branches of science, from renewable energy – things like fusion – to quantum chemistry, material design and finding cures for diseases.
Nuclear fusion you say?
I probably shouldn’t have mentioned that. But yes, I feel like there are some interesting areas of fusion that I think are amenable to AI, to do with the control of the plasma for instance. I have probably said too much, but in future we may have something to discuss about that.
That sounds rather tantalising. With so much going on at DeepMind, how do you spend your day?
I have a rather strange schedule, because my work and my life are all intermeshed. I have a two-day set-up. So during the day, I manage the company and do meetings. In the evenings, I start again. I’m a very nocturnal person, so I sit down again around 10 pm and go into the small hours of the morning. And that’s when I write my research papers, read about and look into different areas, research them. That’s my quiet time, and I love it. And I do that seven days a week.
“What we have invented so far is on the right track. But I don’t think we have all the ingredients”
It sounds like you don’t have much time for sleep, which is important for humans as well as AIs.
Yes, I know, which is bad. I don’t know how many more years I can work like this. And I know very well how bad it is for your brain, so I do try to get 6 hours at least, but I sometimes fail. On the weekends, I try to catch up by lying in, although that’s not as good as getting a regular amount of sleep. I’m working on that, the sleep problem.
The protein problem
Proteins are essential building blocks of life. Inside cells, they form everything from molecular machines for generating energy to tools for sensing the environment, digesting food and carrying out repairs. Each protein is made up of long strings of 20 or so different varieties of amino acid, which are easy to identify. The resulting proteins fold themselves up into complex 3D shapes, full of spirals and loops.
Determining the shape that the protein assumes based on its constituent amino acids is mind-bogglingly difficult. According to one estimate, there are 10300 possible configurations for a single protein. Working out the shape often requires years of laborious experiments and analysis. As a result, we only know the structure of a tiny fraction of the proteins that arise in nature.
Since 1994, a competition called the Critical Assessment of protein Structure Prediction (CASP) has been held every two years to encourage progress. Teams that enter are given the sequence of amino acids that make up different proteins, whose 3D shape has been experimentally determined but not yet published. The teams compete to predict the protein shape from the amino acids. But little headway has been made until recently.
In 2018, DeepMind entered the competition for the first time with an AI called AlphaFold. It won. But it did so with a score of just 60 out of 100 for accuracy in the hardest category. A score above 90 is considered as good as experiments performed by humans.
However, in 2020, a redesigned version of AlphaFold entered and won the competition again. This time, it fared much better with a median score of 87 in the hardest category. It scored above 90 for two-thirds of the proteins across all of the different categories. “I was really wowed when I saw it,” said John Moult at the University of Maryland, one of the competition’s organisers, at the time.
AlphaFold uses many of the same principles as DeepMind’s other AIs, such as AlphaGo (see interview). Rather than being given instructions on how to predict a protein’s shape, the AI learns using a technique called deep reinforcement learning. Through huge amounts of trial and error, using both simulations and real examples, it learns how to make predictions for itself. This means that AlphaFold doesn’t use hardcoded human knowledge to make its predictions. That is probably just as well, because many iterations of the CASP competition have shown that we really aren’t particularly good at making them.
A brief history of DeepMind
The company made a name for itself training AI to play video games. Now it is helping diagnose disease and solve intractable scientific problems
2010
Demis Hassabis founds DeepMind along with Shane Legg and Mustafa Suleyman. The company was initially backed by a range of wealthy investors.
2014
DeepMind is bought by tech giant Google.
February 2015
DeepMind reports its first major success. It has trained an AI to play 49 different video games on the Atari 2600, a console released in the late 1970s. The AI learned how to play the games from scratch, just by watching pixels on the screen as it was played by humans. It went on to beat the highest scores of human players in 23 of the games.
March 2016
In an episode that made DeepMind world famous, its AI AlphaGo triumphed over Lee Sedol, then a world-leading player of the board game Go (see interview).
April 2016
DeepMind had signed an agreement with the Royal Free NHS Trust in the UK to develop an AI service for analysing patient data and automatically spotting signs of kidney problems. In 2016, 91av revealed that the agreement gave DeepMind access to a wealth of medical data from 1.6 million patients, including who visited them in hospital.
August 2018
A DeepMind system shows that it can spot the early signs of eye disease in retinal scans better than human doctors. The system had been trained on anonymised scans from 15,000 patients. It has the potential to flag up people who need treatment before their eye disease becomes irreversible.
December 2018
The company unveils AlphaZero, a version of AlphaGo that can learn to play other board games such as chess and a similar Japanese game called shogi. It was a stride towards a more general form of AI, which has long been considered an important goal.
October 2019
DeepMind’s AI AlphaStar achieves grandmaster status in the video game StarCraft, ranking in the top 0.2 per cent of players. StarCraft, a real-time strategy game set in space, is difficult for AI because there is incomplete information about the virtual world – you can’t see everything your opponent is doing – and there are a huge number of possible moves and tactics.
December 2020
DeepMind cracks protein folding. Its AI AlphaFold is so accurate at predicting a protein’s shape from its amino acid structure that it is on par with experiments performed by humans (see “The protein problem”).
