{"url": "https://www.youtube.com/watch?v=XtJVLOe4cfs", "page_url": "https://www.youtube.com/watch?v=XtJVLOe4cfs", "title": "Using AI to accelerate scientific discovery - Demis Hassabis (Crick Insight Lecture Series)", "show_or_channel": "Google DeepMind (Crick Insight Lecture)", "date": null, "duration_s": 5360, "type": "talk", "speakers": ["Demis Hassabis"], "confidence": "high", "notes": "Francis Crick Institute lecture; uploaded 2022-03-04, event date not found.", "people": ["demis-hassabis"], "id": "0000_google-deepmind-crick-in-using-ai-to-accelerate-scientific-discov", "audio_duration_s": 5359.0378125, "worker": "jc-w12", "transcribe_s": 140.3, "model": "large-v3", "segments": [{"start": 0.0, "end": 26.48, "text": "Hello. Well, welcome, everybody. It's almost normal. Almost, but when I look around the room and see all these masks, it isn't quite normal. But this is the start of normality, so let's celebrate it. And we've got a great start with Demis Hassabis, who all of you will have heard the name, and I'm going to put a little bit of colour to the name, because he has got rather a colourful CV."}, {"start": 26.48, "end": 55.63, "text": "Sorry, Demis. All right. And it is sort of interesting. It started at age 13 when he became a master at chess. I also play a little bit of chess, but I've never actually asked for a chess game with Demis. I wonder why. But I know why. Did his ANS levels, got an entrance into Cambridge at 15, 16, but they thought he should grow up a bit more before going, and he joined a gaming company."}, {"start": 55.63, "end": 81.52, "text": "And designed a game, Syndicate. Yes, isn't that right? Yes. And that sort of stayed with him, as you'll see in a bit later. Did computer science. Got a first at Cambridge. Famously said it was like being in a holiday camp at Cambridge. And then worked in another gaming..."}, {"start": 81.52, "end": 109.34, "text": "And set up Elixir Studios where he worked for a number of years before going to do, and listen to this, a PhD in neuroscience in UCL, in cognitive neuroscience, writing a paper on neuroscience of imagination, which did extremely well. Somewhat later, not very much later,"}, {"start": 109.34, "end": 136.96, "text": "He founded DeepMind, which all of you will be familiar with, bought by Google after some years, actually. And that resulted, and I think you will say a bit about it, Demis, won't you, to AlphaGo, which could beat the chess world champion, and then turned himself to something serious. And that was solving..."}, {"start": 136.96, "end": 166.13, "text": "The age-old problem that we've known about for many years from the birth of molecular biology, and I remember it being stated, of how you go from primary sequence through, of course, the simple problem into the primary amino acid sequence from DNA. How can you determine how proteins fold up? Use AI, particularly machine learning, to actually solve that problem, which he will talk about mainly today."}, {"start": 166.13, "end": 191.55, "text": "Now, all of this achievement, and that is an extraordinary achievement, which I won't say too much about, led to a lot of accolades. He's a fellow of the Academy of Engineering, a fellow of the Royal Society, and advises government, including on SAGE. Now, if this all sounds as if he's a bit of an overachieving nerd, I mean..."}, {"start": 191.55, "end": 201.97, "text": "Which is certainly true. He is also normal. He loves football and, believe it or not, is a fan of Liverpool Football Club."}, {"start": 204.18, "end": 228.35, "text": "Okay, so using AI to find out something, okay? Which he will tell you about. But I want to just say a few words about this because I feel really strongly about this. We try and understand life, usually we call it the mechanistic basis, in terms of primarily chemistry and, of course, also physics."}, {"start": 228.35, "end": 250.88, "text": "I mean, that is the mechanistic reduction approach that we generally use. But for me, that mechanistic explanation, critical as it is, only makes sense at the biological level when it's embedded in understanding how information is being managed in that complex chemical."}, {"start": 250.88, "end": 279.84, "text": "And I go further than that. I would actually say, as well as thinking of life as a chemical and physical activity, science activity, it is also an informational science. Now, that doesn't just apply to the obvious example of moving from DNA as encoding information, critical and central as that was, and of course we named the institute CRIC after that, but also..."}, {"start": 279.84, "end": 306.14, "text": "We associate that particularly with Watson and Crick. But also work from Jacob and Mono that looked at control systems, primarily in bacteria in the 50s and 60s, and continuing, of course, looking at the complexity of information management which makes up life. And for me, the future will be built on the combination of understanding the chemistry and physics of life."}, {"start": 306.14, "end": 316.4, "text": "Coupled with understanding the information management within life. And what we've seen up until now and Dennis will talk about is how machine learning can help."}, {"start": 316.4, "end": 345.9, "text": "Understand how to interpret living things. But I also think artificial intelligence thinking more generally and apply beyond the machine learning will actually lead to greater understanding of how life works, a somewhat different objective, coupled with physics and chemistry. And the reality is I don't think I know how that will work. I don't think Demis knows how that will work. But I am sure it will work and it will be central."}, {"start": 345.9, "end": 369.44, "text": "With that, I'm going to hand over to Dennis. Thanks, Paul, for that really lovely introduction. It's wonderful to be here. It's also wonderful to be in a mostly full lecture hall finally again after all these years."}, {"start": 369.44, "end": 398.16, "text": "It's a real pleasure to be at the Crick. It's always a pleasure to be there. Our offices are just around the corner. I really think the King's Cross area is a real hotbed for scientific discovery and intellectual discovery. It's fantastic to see how it's grown in the last decade or so. And of course, Paul and his team have done a wonderful job of setting up this amazing institute, which I think is a real jewel in the crown for UK science. And it's a real honor and pleasure to sort of be speaking on Crick's fifth anniversary with this lecture series."}, {"start": 398.16, "end": 415.95, "text": "So as Paul mentioned, what I'm going to talk about is using AI to accelerate scientific discovery. Our course will, given especially I'm at the Crick, I will be spending most of the lecture talking about AlphaFold and its impact. And I know many of you, it's really pleasing to hear you using AlphaFold."}, {"start": 415.95, "end": 446.13, "text": "And obviously, there will be lots of time in the Q&A to discuss that. I'm also going to cover a little bit about how we got there, the journey to getting to AlphaFold. We spent 10 years working really hard on the underlying machine learning and AI algorithms that allowed us to get to the point where we were able to tackle something like protein folding. So, first of all, with DeepMind, we founded DeepMind in 2010 in London, and the vision behind it was to create a kind of Apollo program effort."}, {"start": 446.13, "end": 465.87, "text": "for building what we call Artificial General Intelligence or AGI. And you may have come across this term AGI. And we use it internally to distinguish the type of work we work on from sort of normal, common, narrow AI that you might use every day, things like recommendation systems and other things."}, {"start": 465.87, "end": 487.25, "text": "So what we're really after at DeepMind is building this kind of general version of artificial intelligence. And there's not really a sort of set definition of this, but the working definition we use at DeepMind is a system that's capable of performing at human level or above across almost all cognitive tasks and domains."}, {"start": 487.25, "end": 510.66, "text": "So that's what we mean by artificial general intelligence. And you might ask, why use humans as a reference point? Well, for now, it's the only example of general intelligence we know of in the universe. And of course, it's been sufficient to create this amazing modern civilization that we all live in around us. So that's pretty general."}, {"start": 510.66, "end": 520.46, "text": "And pretty amazing. And so it would seem that if we can at least mimic those capabilities then we would have a very useful tool."}, {"start": 521.74, "end": 552.21, "text": "So our mission statement for DeepMind, so we think about the effort we're doing at DeepMind as an Apollo program. We're now over 1,000 research scientists and engineers, mostly based, as I said, in Kings Cross, just across the road. And our mission is to solve intelligence to advance science and humanity. And by solving intelligence, what we mean is fundamentally understand natural intelligence and then mimic that in an artificial construct. Now, why is..."}, {"start": 552.21, "end": 571.6, "text": "Are AGI important? Well, if we could solve it in this general way, then our vision for DeepMind and why I personally work on AI for my whole career is that one could potentially use it as a tool to apply to almost anything by its very nature. Now, historically..."}, {"start": 571.6, "end": 600.88, "text": "This, you know, working on AI obviously has been going on from at least the 1950s. People like Alan Turing, McCarthy, and these kinds of people started it off. And historically, there have been two, I would say, major approaches to trying to build artificial intelligence. The classical approach, one could say, is expert systems. So those are systems famously like Deep Blue that beat Garry Kasparov at chess in the 90s. And what you can think of is they rely on"}, {"start": 600.88, "end": 626.86, "text": "Hard-coded knowledge. Really, the engineers and the scientists designing the system, they have to understand what the solution is, and then they program that solution into the machine as a set of rules, logic rules, and then the machine executes those rules. But the problem with these systems is they can't deal with the unexpected, and they're basically limited to these pre-programmed solutions. So of course, they can't really solve anything."}, {"start": 626.86, "end": 651.71, "text": "They were born out of mathematics and inspired by logic systems. That's been mostly the mainstream tradition of AI for the first 50, 60 years."}, {"start": 651.71, "end": 680.14, "text": "On the other hand, there's a different approach, which one could call learning systems. And the idea here is that these AI systems learn the solutions for themselves directly from first principles, directly from data, directly from experience. And they can generalize, and the emphasis is on generalization, to new tasks and new situations that they haven't seen before. And of course, the promise of these systems is that they will be able to solve things we, as the designers, do not know how to solve yet."}, {"start": 680.14, "end": 697.26, "text": "So that's the potentially amazing thing about these types of learning systems. And, you know, there's a lot of where's the inspiration come for these kinds of systems, neural networks, reinforcement learning, things I'll talk about in a minute. Well, a lot of that was inspired by systems neuroscience, not copying."}, {"start": 697.26, "end": 727.22, "text": "Because obviously we don't actually understand the brain well enough to do that. But inspired by our neuroscience ideas and also validated by things we find out in neuroscience. And that's why I ended up doing my PhD in neuroscience, even though I was ostensibly doing that to get inspiration for new AI ideas. So the learning systems is, of course, this is how the human brain does it. And as I said, the emphasis on this approach is generality and learning."}, {"start": 728.34, "end": 743.6, "text": "So what's an example of a learning system? And there are many types of learning systems, but the one that we are known for at DeepMind and we specialize in specifically is reinforcement learning. And this is the idea of learning from first principles through a sophisticated form of trial and error."}, {"start": 743.6, "end": 767.31, "text": "And it's pretty simple to explain, and I'm sort of going to explain with this cartoon diagram. What happens is the AI system, which we usually call an agent on the left-hand side here, finds itself in an environment. And that environment could be the real world, in which case the agent's probably like a robot, physical robot, or it could be a virtual world, in which case, or simulation or a game, in which case the agent is some kind of avatar."}, {"start": 767.31, "end": 786.06, "text": "And the agent gets information about the environment through its observations. Usually we use vision, but you could use other modalities, hearing, touch, and so on. And based on these observations, which are always noisy and incomplete, the first task of the agent is to build a model of how the environment works."}, {"start": 786.06, "end": 812.43, "text": "So that's the first thing. Now, the agent's normally got a goal that it's trying to achieve in the environment. The goal is usually specified by the designers of the system. Maybe it's to win a game or maximize how many points it gets. And so the second job of the agent system is to, in the time it has available to it to think, it has to select the best action from its set of actions available to it at that moment in time that will best get it incrementally towards its goal."}, {"start": 813.1, "end": 839.22, "text": "And it does this by using the model it has of the environment to sort of predict what will happen if it takes that action. What will happen to the environment? What's the likely new observation? And is that going to help to get the agent towards its goal? And so once the thinking time's over, the agent outputs the action that gets executed. That may or may not make a change to the environment, and that drives a new observation."}, {"start": 839.41, "end": 864.86, "text": "So this is a dynamic system. And what this means is the agent is an active learner. It's not just passively looking at data. What actions it takes, it partly determines what the next experience, the next data that it's going to get is. So it's a kind of active learning system. Again, much like animals and us as humans. Now, this cartoon."}, {"start": 864.86, "end": 892.78, "text": "Diagrams obviously looks very simple. It is simple in some senses, but the complexity behind actually solving all the deep technical questions that arise from these systems are, of course, very complicated and something that's still ongoing. Now, I'm going to talk a little bit about AlphaGo, which was, I still think, probably the most famous example of a reinforcement learning system working at scale. So this combines neural networks, deep learning with reinforcement learning."}, {"start": 892.78, "end": 919.39, "text": "In our program AlphaGo, that was our program to play the ancient game of Go. Now, Go, for those of you who don't know, this is what a Go board looks like. I definitely recommend you trying it out if you don't know how to play it. But it's an amazing, absolutely incredible game. The rules are incredibly simple. You can place your stones. There's two players, black and white. They place their stones on the vertices of this 19x19 board. And the idea of the game is you're supposed to wall off..."}, {"start": 919.39, "end": 943.5, "text": "Empty wall of spaces of territory on the board. So you're kind of building walls. The other person's building walls as well. And the person that has the most territory at the end of the game wins. So this has been played for thousands of years and professionally hundreds of years in Asia. And it sort of occupies the intellectual echelon that chess does in the West, in China and Korea and Japan."}, {"start": 944.82, "end": 966.14, "text": "Now, why is Go so hard for computers to play? So after chess was sort of cracked in the mid-90s, the kind of Everest, if you like, or the holy grail for game AI was Go. And they used to try to initially, of course, what the AI people tried to do, researchers tried to do, was use the same expert system."}, {"start": 966.14, "end": 995.47, "text": "Approach that worked for chess and they tried to use that for Go. But the problem is it doesn't work for Go and for several reasons. One is that the search base is just much more massive than chess. So it's 10 to the 170 possible positions. Just to put that in context, there are only about 10 to the 80 atoms in the observable universe. So there's no way you can exhaustively search through and find the solution to what you should play in Go. So it really resists brute force."}, {"start": 995.47, "end": 1012.5, "text": "And then even more difficult is that it was thought to be impossible to write down an evaluation function. So an evaluation function is the key function to tell the system whether it's winning or the opponent's winning and by how much in a specific situation."}, {"start": 1012.5, "end": 1032.56, "text": "And for chess, this is relatively easy, straightforward to do because chess is a materialistic game. So at a minimum, you know, first all the approximation, you just count up the pieces and what they're worth. And that will tell you a little bit about who's winning. Obviously in Go, it's a lot more esoteric game. There's only one type of piece. So it's just way harder to specify."}, {"start": 1032.56, "end": 1062.03, "text": "What are the right rules are to evaluate a position? And in fact, one can see that if you interview top Go players, what they'll say is they, you know, basically this move felt right, right? You know, there's this sort of notion of relying on their intuition rather than calculation, which chess players tend to do. And so this, you know, this is why Go resisted AI approaches, classical AI approaches for 20 years post Deep Blue."}, {"start": 1062.03, "end": 1081.23, "text": "So we took a different approach, which is that let's let the AlphaGo system itself learn its own evaluation function and learn it through playing against itself hundreds of thousands of games. Let's not try and program in what its evaluation should be. Let's let it learn for itself."}, {"start": 1082.1, "end": 1105.65, "text": "So how did we do that? So this is a kind of little bit of a simplification of the whole system. And actually, after AlphaGo, there were future generations of the AlphaGo program. We've since improved it to AlphaGo Zero and then Alpha Zero, which can play any two-player game to world championship standard. Not just Go, but chess, Japanese chess, any two-player game with the same single system. But they all basically work in the same..."}, {"start": 1105.65, "end": 1134.75, "text": "And this is how they work. What we want to do is train a neural network that through self-play, playing against itself, is able to evaluate the position correctly, so this evaluation function, and also pick probable moves. What are the most likely moves that one should think about and think about playing in a particular position? So you start off with, imagine version one of this system, V01. It plays 100,000 games against itself. To begin with..."}, {"start": 1134.75, "end": 1163.39, "text": "It will be almost playing randomly, right? It doesn't know anything about Go other than the rules. So there's 100,000 games. And then you have this little database, basically, of 100,000 games, positions, and you know what happened at the end of the game. And then what you do is you train a version two of a new network, your network version two, and on that data. And you train it to do things like predict what moves were going to be played in a particular position by the version one system, and also who ended up winning the game."}, {"start": 1163.39, "end": 1188.45, "text": "So you can then kind of calibrate your evaluation function, right? So midway through the game, who's winning? Does that tally up with who actually ended up winning the game at the end? So then you have this version two. And what you do with version two is you then have a little mini tournament with version one. So you play 100 game match off. And if version two beats version one by some threshold amount, which we set to 55% win rate, then"}, {"start": 1188.45, "end": 1212.77, "text": "That was statistically significant, so you can be pretty sure v2 is stronger than v1. Now, you replace v1 with the v2 system as the master system, and you now do a new 100,000 games of v2 playing against itself, and you generate a new dataset of 100,000 games. Now, of course, this 100,000 games will be slightly higher quality."}, {"start": 1212.77, "end": 1237.12, "text": "So they'll be slightly better than the version 1 ones. So you're now learning from your own self-generated data, but it's a slightly higher quality data set. And so then you, of course, just repeat this. You train up a version 3 based on this version 2 data. You have the match-off. If, for some reason, the new version does not beat the old version, what you do is you just continue to generate another 100,000 gains with the old version."}, {"start": 1237.12, "end": 1262.64, "text": "And so then now you have 200,000 games to train the next new version. And eventually, the new version will be the old version. And it turns out, after about 17 or 18 cycles of this, you get to world championship level, starting from random. So it's kind of amazing to see the evolution. You can track it in real time. It's just really kind of an amazing thing to see when you see this for the first time."}, {"start": 1263.02, "end": 1275.44, "text": "So what does that do more pictorially? If you imagine that this sort of tree of all possibilities in a go position, so this is like a gaming tree, and you imagine that each of these little graphics is a go position."}, {"start": 1275.44, "end": 1305.04, "text": "And so, you know, this tree fans out in an incredible amount very quickly. And, you know, there's 10 to the power 170 positions. So how does one, you know, how can one find a good path through this enormous search space? Well, what happens is we use the neural network to guide the, we use a Monte Carlo tree search in this case for AlphaGo, to guide the search. So it doesn't need to look, it only looks at a tiny fraction of all the possibilities that are, you know, possible."}, {"start": 1305.04, "end": 1333.82, "text": "It only looks at the most plausible, most fruitful ones, the ones that I've colored in blue here. And then eventually when it's time to play a move after a few minutes, it will pick the best one from those blue options that it searched, in this case this path in pink. So that's how you can imagine the system works, is this model, this neural network."}, {"start": 1333.82, "end": 1359.23, "text": "It helps to reduce down massively by obviously many, many, many orders of magnitude the amount of searching that needs to be done. So we then took this AlphaGo system to a challenge match with Lee Sedol, 18 times world champion legend of the game Lee Sedol in Seoul in 2016. And we famously sort of won that game 4-1, over 200 million viewers. And people, researchers both in sort of..."}, {"start": 1359.23, "end": 1384.42, "text": "The AI field and also Go players, both worlds sort of said that this was a decade before they were expecting anything like this. So until AlphaGo came along, no program had even beaten a professional player, let alone a world championed, a sort of legend of the game. But the important thing was actually not so much, of course, it's really important that we won the match, but it was actually what was even more interesting was how AlphaGo won the match."}, {"start": 1384.42, "end": 1407.14, "text": "And I haven't got time to go into the details here, but it basically came up with its own original strategies. Because, of course, it wasn't constrained by what we as human players and programmers knew. So it sort of came up with its own ideas about what was a good go strategy and good go motifs. And most famously, this happened in game two of this match on Move 37."}, {"start": 1407.14, "end": 1436.43, "text": "And move 37 is this black stone here circled by the red. That's what AlphaGo played. And it was unthinkable. If you look at the commentators and what they said about this, everyone thought this was a terrible move. And the computer didn't know what it was doing. Because if you're a Go expert, you would never, at this early on in the game, play a stone on what's called the fifth line, this line in red here that I've outlined. And just suffice to say, that's just an unthinkable..."}, {"start": 1436.43, "end": 1461.07, "text": "And you would get told off by your Go teacher if you ever moved there this early on. And yet AlphaGo did that. And then what was really amazing was about 100 moves later, this stone happened to be in exactly the right place to decide the main fight that happened across the board. It was positioned perfectly to decide that fight, pivotally decide that fight, almost as if AlphaGo had sort of foreseen that 100 moves before."}, {"start": 1461.9, "end": 1487.94, "text": "So this game has now gone down in Go Legend. It's in all the books. Everyone studies it. And people now play on the fifth line. So that was kind of amazing to see. And if you're interested in AlphaGo, I can recommend the award-winning documentary of it, which is on YouTube now, and you can see all of the drama that was involved in that. So that was..."}, {"start": 1487.94, "end": 1512.24, "text": "You know, games. And we use games as a training ground for our algorithms because they're super efficient to use. You can run millions of them online in the cloud. It's also very easy to specify the goals because there's a clear metric, you know, winning a game or maximizing the points. So also, of course, it's partly due to my games background that I was thinking this would be..."}, {"start": 1512.24, "end": 1529.71, "text": "The right thing to use. And so for the past decade, since we started DMI from 2010, there's been amazing progress in AI and machine learning. And of course, all of you are aware of that. Really amazing. I just can't believe..."}, {"start": 1529.71, "end": 1548.42, "text": "Back in 2010, virtually nobody was talking about AI. We virtually couldn't get two pennies together to start DeepMind. We sort of had to beg, borrow, and scrape around to just get the initial small amounts of funding. And now, you know, everybody, 12 years later, 10, 12 years later, everyone seems to be talking about AI."}, {"start": 1548.42, "end": 1577.71, "text": "But for me, what's really exciting is I think we're on the cusp of the most exciting era of all. The last 10 years have been very exciting, things like AlphaGo, but I think we finally made enough progress now to tackle what we were really trying to do, which is, of course, we love games, and I love games, but the aim of what we were trying to do at DeepMind, this Apollo program, was not just to win at games. It was to apply AI, design these general algorithms that could be applied to important real-world problems. Of course, specifically..."}, {"start": 1577.71, "end": 1595.09, "text": "Specifically in scientific discovery, which is my personal passionate reason for building AI. But of course it can be applied to many other things industrially and commercially as well. So I think what's really exciting is we're at that start of that era now. And I'm going to talk a little bit about..."}, {"start": 1595.09, "end": 1624.37, "text": "What makes a suitable problem for the types of techniques we've currently developed? And of course, this itself is a moving feast because we're continuing to extend what these AI systems are capable of doing. But right now, the sorts of systems I've just described to you, what makes a suitable problem for them? Well, it's these three kind of principles that I look for when we're considering a new problem. One is, is the problem, can it be described or encapsulated as a sort of massive combinatorial search space?"}, {"start": 1624.37, "end": 1652.48, "text": "State space. We like that criteria because that shows you, if there's something like that, like Go, then brute force methods, expert systems and brute force methods will have a lot of difficulty in that area. So we look for that kind of criteria. Second is, is there a clear objective function or metric that we can optimize against, that the learning system can optimize against incrementally? So again, of course, in games that could be like points."}, {"start": 1652.48, "end": 1674.24, "text": "Then third, are there lots of data to learn from or, and ideally it's an and, an accurate and efficient simulator with which you can generate more data. So self-generated data and synthetic data is also fine if it's drawn from the right distribution. So those are the three criteria. It turns out that actually there's a lot of things that fit this."}, {"start": 1674.24, "end": 1703.65, "text": "So maybe this sounds quite restrictive to you, but actually there's a lot of things that can be couched in these terms. A couple of other ancillary things we look for is something that has an external benchmark or test. Games obviously are great for that because you can play world champions and other very impressive players, and that can calibrate for you externally how well your system's doing. And obviously we want to look for problems that have a big downstream impact. So it turns out..."}, {"start": 1703.65, "end": 1725.49, "text": "The protein folding ticks all of those boxes and then some. And obviously all of you in the room will know what this is, but I know there's people listening online who maybe are not in biology. So the protein folding problem is, of course, the problem of going from an amino acid sequence, sort of one-dimensional sequence of letters, to a 3D protein structure."}, {"start": 1725.49, "end": 1753.98, "text": "And this has been a kind of 50-year quest ever since Christian Effensen, I think, famously conjectured in his Nobel Prize lecture in 1972 that this should be possible. The three-dimensional structure of a native protein should be fully determined by the amino acid sequence. I do feel it's a bit like Fermi's last theorem of, like, yeah, this should be possible, but it's just a bit too small. You know, the margin's a bit too small. We'll leave you to figure it all out. And then there's 50 years of, obviously, attempts at trying to do this."}, {"start": 1753.98, "end": 1762.24, "text": "But this really does fit the criteria I just mentioned earlier on."}, {"start": 1762.24, "end": 1791.73, "text": "First thing is the state space. And famously, Leventhal, another big scientist of the same era, put forward this paradox of he calculated maybe for an average protein, there were 10 to the power 300 possible conformations. But then how is it that in nature, this protein spontaneously fault within milliseconds in many cases? And obviously, randomly sampling all the possible conformations would be totally..."}, {"start": 1791.73, "end": 1804.43, "text": "So this is one of those massive state spaces, even bigger than Go is. So the big question is, can the structured prediction problem be solved computationally?"}, {"start": 1805.04, "end": 1833.66, "text": "Now, for me personally, this has been a long road to this problem, to deciding when to tackle the problem. I think that's one of the key things in science, but also in startups, is you don't want to be too far ahead of your time. You want to be ahead of your time, say, five years, but you don't want to be 50 years ahead of your time. Otherwise, you're just going to be in for a lot of pain. And my sad story about that is always like Charles Babbage. I think he was maybe 100 years too ahead of his time. And it's a pretty tragic story if you read about his life, even though, obviously, he did amazing things."}, {"start": 1833.66, "end": 1862.99, "text": "So for me, I first came across the problem of protein folding actually as a student at Cambridge in the 90s because I had a friend who ended up doing his whole career in protein structure, still at the LMB now, would talk about protein folding as the most important problem to crack at any available opportunity, you know, in the bars, playing pool, whatever. This is what he would talk about. So it stuck in my mind as like a really interesting problem. And also even back then, obviously I was doing AI in my games at that point."}, {"start": 1862.99, "end": 1885.6, "text": "Like Theme Park and the games that Paul mentioned. But I kind of thought this might be a good thing for AI at some point. And then the second time I came across protein folding was with the game Fold It, some of you will know about, where I was doing my postdoc at MIT in the late 2000s, 2008, 2009. And I think it just came out."}, {"start": 1885.6, "end": 1911.71, "text": "Of course, I was fascinated about this idea of citizen science games. I still think there's a lot more actually that could be done there. Like, can you use all of the motivation and time gamers spend on their games? If we could couch our scientific problems in a game, maybe we could kind of harness all of that leisure time to do useful science. So it's a pretty cool idea. And I think Foldit's still one of the best examples of that. And what was intriguing was that..."}, {"start": 1911.71, "end": 1929.49, "text": "The best players, obviously were not biologists, they were just kind of gamers, discovered some real structures, I think for some quite important proteins. I think there were a couple of Nature papers actually published on structures that the Foldit players had found. And what that said to me, and this is a little..."}, {"start": 1929.49, "end": 1948.22, "text": "Screenshot of it is, you know, it's like a puzzle game, right? People bend the backbone and they try to minimize the energy. And of course, they would sometimes do counterintuitive moves. So they would make a move that would increase the energy temporarily and then only for them to do another fold and then that would reduce it overall."}, {"start": 1948.22, "end": 1976.13, "text": "And so if you think about what were these players doing as sort of amateur biologists, well, presumably as they were quite good gamers and they got quite into this game and they were using their intuition in a way, their pattern matching skills to figure out what the right moves were. And some of them had got really good at figuring out their intuition for protein structures. So that really then resonated with me when we did AlphaGo of like, what had we done there? Well, we just mimicked."}, {"start": 1976.13, "end": 2000.32, "text": "The intuition of these incredible Go masters who spent their entire lives playing Go. You could not find anyone more expert with their pattern matching skills than a Go player. And so I just felt like it must be then possible. That gave me some sort of hope that it would be possible to build a system to at least match the intuition of these Foldit players."}, {"start": 2000.32, "end": 2020.9, "text": "Pretty much the day after we got back from Seoul in, I think it was March 2016, we decided that it was the right time to start the AlphaFold project back then. So it's been five years. Now, the other key thing, the reason we decided to pick protein folding was the existence of CASP, which is the kind of..."}, {"start": 2020.9, "end": 2050.29, "text": "I would say the gold standard benchmark for protein structure prediction. It's run every two years since 1994. And it's run by this amazing team, John Malt and his team. John Malt founded it back then and has kept it running for 30 years, which is incredible. And the key thing about it is a blind prediction assessment, as I'm sure most of you in the room will know, which is that these are newly found experimental structures. They've not been published yet, but they're entered into the competition. So at the point where..."}, {"start": 2050.29, "end": 2076.29, "text": "The structure is not known. So there's no way you can overfit or do anything accidental that sometimes happens in machine learning. It's truly blind. So it's a fantastic test, if you like. And then, of course, later in the year, you do this in the summer. And then later in the year, they reveal the real structures and then score your predictions against that. So progress at CASP."}, {"start": 2076.29, "end": 2101.09, "text": "had been very slow historically. So really, for the previous decade before we started work on this area, these are the results as measured by this score, GDT score, which you can think of roughly as a percentage accuracy. And you can see the sort of CAS competitions from 2006 up to 2016. This is the score of the winning team."}, {"start": 2101.09, "end": 2126.29, "text": "On the hardest category, free modeling category. And you can see there's been sort of basically no progress. And then when we entered for the first time in 2018 in CAS 13, we suddenly jumped the winning score by nearly 50% by applying cutting edge machine learning as the core part of the system for the first time in a structured prediction system."}, {"start": 2127.09, "end": 2152.24, "text": "And then we re-architected. We sort of took those learnings and that experience. We tried to push that further. We realized those particular techniques were hitting a ceiling. So then we took the knowledge and the know-how, but we re-architected the system from the beginning with that knowledge to then try and reach atomic accuracy, which we managed to do in the next competition in CAS 14 in 2020."}, {"start": 2153.52, "end": 2168.22, "text": "So atomic accuracy was always our goal. We achieved that at CAS 14 in November 2020. This measured by the mean, the sort of median error was less than one angstrom."}, {"start": 2168.22, "end": 2186.29, "text": "So within an atom error, and therefore that was competitive. That's what we've been told by the CASP organizers. They would consider competitive or comparable with experimental methods, and more importantly, potentially useful in practice for experimentalists and others to use and rely on."}, {"start": 2186.29, "end": 2212.53, "text": "So AlphaFold2 got a 0.96 Angstrom error, and that was three times more accurate than the next best system in CAS14, even though by this point they had all been using our CAS13 and AlphaFold1 techniques that obviously we'd published by then. So now even the rest of the teams were using machine learning at this point, almost all of them, but AlphaFold2 was still three times better than that with this new system."}, {"start": 2213.52, "end": 2242.43, "text": "There was an amazing time where you get the first results back in September, October, and we were just sort of blown away by the beauty of the accuracy of the predictions. You can see here the ground truth is in green, the predictions are in blue. And even in some cases like this protein from SARS, from COVID, even the side chains are pretty good in some situations."}, {"start": 2242.43, "end": 2265.09, "text": "Those were the examples that we got back. We got this incredible score of 0.96 Angstroms. This is the Alpha Fold 2 architecture. It's an incredibly innovative machine learning architecture. I haven't got time to go into the gory technical details of this, but if you're interested in those, then obviously check out the full methods paper that we published last year."}, {"start": 2265.09, "end": 2294.38, "text": "But I just want to give you a flavor of the key technical advances, the key takeaways. So first of all, this is the most complex system we have ever built at DeepMind, I think, by quite some way. There was also AlphaStar, which we did for the StarCraft game, but this is the most complex system, I would say, which is saying something, because we build a lot of complex systems. And there's 32 component algorithms, and one measure of how complex it is is that there's 60 pages of supplementary..."}, {"start": 2294.38, "end": 2318.61, "text": "Quite dense supplemental information with the paper that we wrote. Describing all of the different aspects that were required. The key thing there is that there was no silver bullet. There wasn't just one algorithmic trick that would then solve this problem, but we needed a whole bunch of insights and innovations that needed to be carefully integrated and blended together."}, {"start": 2319.28, "end": 2334.03, "text": "But the key takeaways were, firstly, these are the key advances over the AlphaFold1 system that were all essential for getting this atomic accuracy performance. It's a full end-to-end system with including a recycling stage that sort of..."}, {"start": 2334.03, "end": 2359.33, "text": "Recycles the prediction and iteratively refines the structure over a few cycles. Now, end-to-end means we start off with the amino acid sequence and we give you back the 3D structure. With AlphaFold1, that wasn't the case. With AlphaFold1, we gave you back an intermediate thing, the distogram, so the distances between the residues, and then we had to use a different numerical method solution to actually give you back the 3D structure."}, {"start": 2359.33, "end": 2381.63, "text": "So there were sort of two parts to AlphaFold 1, and we managed to combine that into one end-to-end system. And in machine learning, one of the sort of things that's known is that the more end-to-end you can make it, the more you can allow the learning to flow through the whole system, the better it is for the system. So we were always planning to do this at some point, and we finally managed to do that with AlphaFold 2."}, {"start": 2381.63, "end": 2389.62, "text": "The other key thing is we switched out the kind of neural network we were using. We used this attention-based neural network, a very fancy one."}, {"start": 2389.62, "end": 2418.9, "text": "To infer the graph structure between the residues. So that actually kind of builds up the graph structure. Before that, we were using convolutional neural networks in AlphaFold1, which was borrowed from computer vision. But the problem with convolutions is that they make some certain assumptions, which are fine for computer vision. You know, neighboring pixels are sort of related to each other. But actually, that's not, if you think about it, that's not good for protein folding. Because often residues that are near each other or even residues that are far away could end up being close to each other in the final three."}, {"start": 2418.9, "end": 2445.36, "text": "So in fact, we were giving it the wrong biases by giving architectural biases by using convolutional neural networks. So that was key. And then perhaps the most vital thing of all was building in some evolutionary and physical constraints into the architecture, things like bond angles and things like that, without impacting the learning, which is another really hard thing in machine learning to do, is to mix."}, {"start": 2445.36, "end": 2456.37, "text": "Prior knowledge that you have about the system, but not impact the training and the learning. So I think it's a really great, AlphaFold2 is a really great example of how to blend those two types of knowledge."}, {"start": 2457.3, "end": 2478.9, "text": "So it's been a massive research effort for us, probably our longest project we've done. Five years, roughly 20 people, hugely multidisciplinary team, amazing team. Just want to give a shout out to them. But of course, it's not just machine learners. We also have biologists, chemists, and physicists on the team. So we needed all of those different types of interdisciplinary knowledge in order to crack this."}, {"start": 2478.9, "end": 2502.72, "text": "And then the final point that I think is interesting to make is that I mentioned at the start of the talk, our overall mission is to build AGI, this general system that can do anything. So normally, and with our games work, we're really interested in general systems. So we don't want domain-specific knowledge. So AlphaZero, as I mentioned, can play any two-player game from scratch without knowing anything about those games."}, {"start": 2502.72, "end": 2525.2, "text": "But when you're trying to apply those techniques to a specific problem that has important downstream consequences like alpha fold, then actually we don't worry about generality. We just throw the kitchen sink at it. We just want the performance. We want it to work and we want it to reach a performance threshold where it will be useful."}, {"start": 2525.2, "end": 2552.96, "text": "And, of course, with AlphaFold and protein structure prediction, that's absolutely the case. So that's why you see so many of these sort of general techniques, but also include some domain-specific knowledge in this case. And, of course, that's absolutely fine for something like AlphaFold, too. And then, of course, now, once we've built a system like that, we always go back to it and see, can some of the domain-specific aspects that we put in be removed?"}, {"start": 2552.96, "end": 2568.56, "text": "All replaced with a more general system. And that's how we went from AlphaGo that could only play Go to AlphaZero that could play any two-player game. We sort of pulled out the Go-specific things out of AlphaGo and just kept seeing if we could replace it with something more general."}, {"start": 2570.22, "end": 2599.82, "text": "So the speed and scale of this, this is another thing that we do is we often, again, in machine learning and our engineering stuff, we try to meet the performance we're trying to reach first. Then once we get there, we then spend time, go back to looking at the system and see if we can optimize it to make it more efficient. So by the end of doing AlphaVault 2, it's actually an incredibly fast system. Training the whole network only takes about two weeks on roughly eight TPUs, which is..."}, {"start": 2599.82, "end": 2630.0, "text": "Tenser processing units, which are roughly about 150 GPUs, which is a very modest amount of compute actually for modern day machine learning systems. So this is not a story about compute power. It's actually what was key was the architectural innovations. So this is really fast for training, but it's even faster for inference, for asking, getting the answer back for a particular structure, takes in the order of minutes or even seconds for an average protein on a single GPU machine."}, {"start": 2630.83, "end": 2644.24, "text": "So when sort of over Christmas, just after the competition around December 2020, we were thinking about, okay, how are we going to give broad access to all of you and biologists to the predictions?"}, {"start": 2644.24, "end": 2673.52, "text": "And normally the way you do it with these types of systems is you set up a server and then you can submit your structure to the server and it will give you back a structure. But we thought we could do something better than that, given the speed of the system, is we could actually just fold entire proteomes. And that was within reach. So there's a sort of moment of realization. And then in fact, we did do the human proteome over Christmas. So we set all our machines running, left them for Christmas, came back in January."}, {"start": 2673.52, "end": 2684.18, "text": "And there was the human proteome, which is one of the magical things about computational methods. They can carry on working while you have your Christmas lunch."}, {"start": 2684.18, "end": 2711.68, "text": "That's what we decided to do. We did our first version of human proteome. Of course, we had to check it and analyze it and build analysis tools for that. That took a few months. We had these really good predictions. Then we wanted to build a confidence measure of how good were these predictions. I'll explain that in a minute. But basically, experimental coverage of the human proteome is about 17% after."}, {"start": 2711.68, "end": 2730.82, "text": "You know, 30 years of painstaking and amazing experimental work. And we more than doubled that in one go to 36% with very high accuracy predictions. And around 58% of the human protein was covered by what we would call high accuracy predictions. And what was interesting is even the..."}, {"start": 2730.82, "end": 2759.04, "text": "The rest of the proteins that had a lower accuracy or confidence around them, they seem to map, and this is still an ongoing piece of research at the moment we're doing with some collaborators, to areas of the proteins that are thought to be unstructured in isolation. So in fact, one could maybe turn it on its head and think about alpha fold as a disorder predictor. If it can't predict something very well, then perhaps it is disordered in isolation."}, {"start": 2759.04, "end": 2772.02, "text": "So this is the most complete and accurate picture ever of the human proteome. So we published another paper on that with all the analyses of it. The key thing is we knew we wanted..."}, {"start": 2772.02, "end": 2801.36, "text": "Biologists to use this, and perhaps people, scientists and researchers who were not familiar with machine learning models. So we wanted to make sure, in this case, that confidence measures were really important. And normally they were nice to have for machine learning models, like how confident is your model about its prediction? But here we felt that they were critical. And so we came up with a nice sort of novel measure of confidence per residue confidence metric, which we call PLDDT."}, {"start": 2801.36, "end": 2832.74, "text": "Over 90 on that, one can think of it as very high accuracy with even the side chains would be good and sort of competitive with experimental methods. That's kind of like the dark blue. If it's above 70, we believe that's high accuracy, and you can think qualitatively that means that probably the backbone is pretty reliable, but maybe not the side chains and things. And then less than 50, as I mentioned earlier, could be a good predictor of disorder, but obviously more research is needed before we could assert that confidently."}, {"start": 2832.74, "end": 2854.67, "text": "Wide stop at the human proteome. Of course, we then decided to do further 20 proteomes for model organisms, all sorts of different important research species, but also some pathogens. And then we collaborated. We had this amazing collaboration with EMBL, EBI, the European Bioinformatics Institute based in Cambridge, and to..."}, {"start": 2854.67, "end": 2877.97, "text": "Build this into a standard database tool. And of course, the EBI hosts a lot of the world's most important databases, things like Uniprot. So they were perfect partners for this to leverage all their expertise in building databases and make it as simple as possible for biologists like all of you lot to use it as part of your standard process. And so we put all of these predictions up."}, {"start": 2878.0, "end": 2906.9, "text": "In the database and released it with the methods paper, free and unrestricted access to every structure for whatever use one wants to use it for. We felt in this case that was the best way to maximize the scientific impact and benefit to the research community. Of course, hopefully, what all the researchers do with it will downstream benefit humanity as a whole, which goes back to obviously our original and still mission."}, {"start": 2907.92, "end": 2937.52, "text": "We're now up to a million predictions in the database today, so we've expanded it to everything in SwissProt. Obviously, we've curated proteins of interest. We've also been very interested in helping out the WHO and various non-profit foundations who look at neglected tropical diseases, because often they don't have the structures for those things, and pharma companies don't generally invest in those diseases. So it's especially useful for them. So we've been working with the WHO very carefully on that and asking..."}, {"start": 2937.52, "end": 2956.72, "text": "And then what they want for global health priorities. So we provided them with a couple hundred thousand new structures on organisms for neglected diseases. And then the community has already done amazing things with AlphaFold just in six months of us releasing all of this and the open source code and the databases."}, {"start": 2956.72, "end": 2985.5, "text": "Colleagues of ours at EMBL have used it to help model their big complex biology, like the nuclear pore complex. I mentioned already people have shown AlphaFold is better than the state of the art in protein disorder predicting. We've worked with G7, WHO to give them back all the structures of the top 30 pathogens they've identified that could potentially cause a future pandemic. And has actually been pretty useful for experimentalists to help resolve"}, {"start": 2985.5, "end": 3007.74, "text": "They're experimental data on protein structures. And we expect there to be hundreds of other applications in the coming years. The impact of Alpha has been very good. 350,000 researchers have used the database already from 190 countries, millions of structures used, and so on. Obviously, we've had some nice accolades."}, {"start": 3007.74, "end": 3030.32, "text": "And what we plan to do from here is over the next year, we plan to fold every protein in Uniprot. So every protein sort of known to science, over 100 million proteins and counting. And obviously, we're continuing to collaborate with the EBI on improving the database and adding new features. And if you have any suggestions yourselves, we have lots of feedback forms in the database."}, {"start": 3030.32, "end": 3057.6, "text": "We do read all the emails. So please tell us if you have any suggestions or you found some structures that you think look strange. And then we will incorporate that in the next, you know, try and improve that in the next version of AlphaFold. So just to finish then, I want to obviously thank the PDB, who obviously without those initial experimental structures, we couldn't have trained AlphaFold and obviously the experimental biology community at large."}, {"start": 3057.6, "end": 3073.98, "text": "Thanks to everybody who made AlphaPhile possible, obviously the methods team and the human proteome team, our collaborators at EMBL, and also the CAS community. But just to finish, I'll just talk a little bit about the future-looking work. So we're now..."}, {"start": 3073.98, "end": 3090.46, "text": "Pressing forward on many other new dimensions. This is just the beginning for us, I would say, of what we can apply AI and computational methods to. We've been looking at protein complexes. Many of you have used our new AlphaFold system, AlphaFold Multimer, that can deal with multimers pretty well, disordered proteins."}, {"start": 3090.46, "end": 3117.58, "text": "Point mutations, ligand docking, protein interactions, protein design. This is just some of the things that we are currently looking at. And just to connect back with what Paul said at the beginning, and actually Paul and I have been discussing this on and off for like 25 years now or something. And maybe we've sort of joked, maybe finally is the time to do these things, is I think we're kind of the dawning of a new era, perhaps in digital biology. I think at its most fundamental level, biology is an information processing system."}, {"start": 3117.58, "end": 3144.66, "text": "But life is phenomenally complex, as all of you will know better than me, and it's an emergent process. So my suspicion is that it's too complex to ever be sort of describable in kind of neat mathematical equations. I know some people try that, and mathematicians are trying that, but I just don't think there'll be kind of a Newton's laws of motion for a cell. I just think it's too emergent and too interactive and too dynamic."}, {"start": 3144.66, "end": 3162.32, "text": "But potentially, it's the perfect type of regime for AI to work in. Sort of figuring out all the patterns in all these weak signals. So I actually think AI might be the right description language for biology, just as in the same way that math is the right description language for physics."}, {"start": 3162.38, "end": 3185.26, "text": "And I think and I hope that AlphaFold is a kind of proof of concept of that, a major proof of concept on that, and perhaps heralds the beginning of this kind of new era of what I would call digital biology. And the dream would one day be able to create something like a full virtual cell simulation that you could interrogate and do experiments in silico and then validate that experimentally."}, {"start": 3186.48, "end": 3207.07, "text": "I mentioned about scientific discovery. We've actually had an amazing sort of year. It's not just AlphaFold. We've had all these other big sort of high-profile papers and breakthroughs in many other areas of science, quantum chemistry, pure mathematics, fusion, genomics. You can read the papers if you're interested. I haven't got time, obviously, to go into all of them."}, {"start": 3207.07, "end": 3237.07, "text": "As you can imagine, a very fun time for us over the last year to tick off our dream list of problems in science and areas in science that are super fascinating and at the cutting edge themselves. I think there's so much more to come. I'll just end by saying the way I've always viewed AI is I wanted to build this tool, this ultimate general purpose tool to help scientists see further, a little bit like the Hubble telescope has allowed cosmologists to see the universe much more deeply and much further."}, {"start": 3237.07, "end": 3274.48, "text": "Thank you. Wonderful, Demis. Thank you very much. Really good. Okay, so we've got questions in the lecture theatre and questions also online. We'll start in the lecture theatre. And there's two microphones. And just come up and ask away. Number one. Hello."}, {"start": 3274.48, "end": 3293.97, "text": "I don't know if you can hear me, but fascinating talk. I was wondering if you could talk a little bit about how AI reinforces biases in society and how that might have an impact on health and disease and what DeepMind is doing to address that. Yes, great question. I think one always has to be very careful. Obviously, this is a very..."}, {"start": 3293.97, "end": 3323.84, "text": "Hot topic right now in research of addressing biases in the system. Part of that comes from the data sources. So that's a really important thing to look at and analyze itself, the data creation, where the data comes from. It's interesting, there's a lot of active research going on in things like generating synthetic data to perhaps rebalance the data bias that one might get from the real data. But the generation itself can also be biased. So you have to sort of be aware of this at every stage."}, {"start": 3323.84, "end": 3342.96, "text": "This is why we spend a lot of time broadening our diverse set of people that work on these systems because I think that even though these systems are learning systems, so they learn for themselves from data, there's still going to always be an echo in the design of the structure or the goal that you give it from the designers."}, {"start": 3342.96, "end": 3367.06, "text": "So even though it's not as direct as it used to be with the expert systems, there's still going to be an echo of the value systems of the people and the culture that they're in that designed it in order to set off the learning system and what it's going to learn. So I think one just has to have that at the forefront of your mind the whole time and get as diverse set of inputs as early as possible in the design of these systems and the engineering of these systems."}, {"start": 3367.06, "end": 3396.53, "text": "We're trying very hard to do that at DeepMind and sponsoring a lot of things in the ecosystem, diversity scholarships and other things to also bring up the next generation to get into this area. Next one at the back. So my question kind of has three parts, but it's one question. So you mentioned intuition several times in the talk, and I was just wondering how common it is that as humans we solve"}, {"start": 3396.53, "end": 3425.26, "text": "The kinds of problems that AI is suited for based on the three criteria you gave. How common is it for us to solve those problems using intuition? And then the second part of the question is, why is conscious computation in our sense slower than intuition is? And what does that tell us about consciousness? Well, these are the easy questions, yeah."}, {"start": 3425.36, "end": 3451.09, "text": "Great questions. So intuition, you might have to remind me of all of them as I answer them. Intuition, so I have a working definition of intuition. I don't want to start a huge row about, because obviously these are sort of in the realms of philosophy. What is intuition? What is consciousness? But my operational definition of intuition is knowledge. I don't think there's anything really mystical about it. It's knowledge that we've gained through experience, if you're talking about human intuition, but it's not consciously accessible."}, {"start": 3451.09, "end": 3478.51, "text": "So our conscious parts of the brain can't access it and we can't describe it, right? That's all it is. And I think in scientific endeavor, but also in our normal lives, if I tell you, you know, describe to me how you swim or ride a bike in a way that I could program into an expert system, incredibly difficult, right? Because it's below the level of what we're consciously able to access. And that's traditionally why those types of things have been very difficult for roboticists to do and so on."}, {"start": 3478.54, "end": 3492.54, "text": "So I think, of course, we use that as scientists, I think. And often we call it things like having good taste or good judgment. And I think the best scientists do have that. And I think that what that is is all these signals that we're..."}, {"start": 3492.54, "end": 3512.51, "text": "Bombarded with from our experiments or reading papers or listening to speakers. If you're a really good scientist, your intuitive part of your brain is going to synthesize some interesting new conjecture from that. So I do think we use our intuition a lot. But I don't think there's anything too mystical about it."}, {"start": 3512.51, "end": 3530.06, "text": "AI systems can also find those types of patterns and signals too. I think pattern matching or pattern recognition is a key part of intuition. That's the first question. Consciousness. I think you said conscious expression."}, {"start": 3530.06, "end": 3550.18, "text": "You know, I don't think that consciousness is necessary for intelligence. I think that if you were to push me on it, I would say that they're sort of doubly dissociable. So I could imagine intelligent systems that do intelligent things that don't feel conscious to us at all. And I would classify all the systems I've shown you today as that. I don't worry about any of them as being..."}, {"start": 3550.18, "end": 3571.98, "text": "Partially conscious. Some people do, some of my team do worry about not letting AlphaGo play Go games and whether that's, you know, it's sad because of that. But I don't think it is. Although there is something a little bit, because Go is considered to be an art form. So it is in Asia, right? So it is a little bit like, are we turning off Picasso that could make more Picasso's paintings, you know, Go games?"}, {"start": 3571.98, "end": 3596.45, "text": "That's not to be taken too seriously. I think the other way is true too. I think that if you look at the higher animals, dogs and dolphins and things like that, I think it's pretty clear, I would argue that they are conscious. They have consciousness, they seem to dream. If you look at a pet dog or something, they have complex social dynamics."}, {"start": 3596.45, "end": 3622.03, "text": "But they're not as intelligent as humans. And maybe in some cases, one wouldn't call them general intelligences, right? In the sense of they could do sort of generalize to anything. They're quite specific to their ecological niche. So it's a sort of interesting dissociation. But actually, I think that building AI in the way we're trying to build in this general way and analyzing that, it's going to be one of the best ways to find out more about what consciousness is."}, {"start": 3622.03, "end": 3651.97, "text": "Because I think if we compare these systems to the human brain and see what the differences are, I would say that we might see what the special essence is of perhaps human consciousness that makes us different. Peter. I was fascinated you brought in evolution, and I wonder if you could tell us how important evolution is in your models. Yes. What we think here as biologists is our lives are made immensely more difficult."}, {"start": 3651.97, "end": 3681.68, "text": "Because of the perversity of evolution, as Francois Jacob famously said, nature is a tinkerer, not an engineer. Whereas, in a sense, you're an engineer. Can you comment on whether nature could have done things in many different ways and achieved the same solution? In other words, the inverse of what you're doing. And how important is evolutionary constraint in your prediction?"}, {"start": 3681.68, "end": 3690.74, "text": "Well, I mean, the evolutionary constraint comes in pretty important with the multi-sequence alignment part of the system. So it is key. It's not..."}, {"start": 3690.74, "end": 3718.13, "text": "We did these ablation studies where you can remove all the different bits and see how well it still does. It's not completely essential, but it's important for the very high performance. That's where the evolutionary part comes in. In terms of your more general question, I think evolution is an incredibly powerful, simple algorithm. I think that's clear what that is. That's what's given rise to all of this."}, {"start": 3718.13, "end": 3745.25, "text": "And that is fascinating, but it isn't efficient, right? So it is sort of, as you say, a tinkering system, and it needed billions of years. And you might not, you know, who knows if you reran the experiment, would you get human intelligence or not? It's not clear. So perhaps there was, you know, a great, you know, people talk about these great filters that had to be got through, like going from, you know, going to eukaryotes and so on. So I think..."}, {"start": 3745.25, "end": 3774.75, "text": "We would like something a little bit more direct than evolution, although we do also have a team that does evolutionary programming. So actually mimics some of the aspects of evolution. So just to bring in that diversity element, because it is another way of exploring and finding novelty, is obviously to randomly try things. But it's a little bit of a brute force algorithm, I would say. So we're probably after something a little bit more efficient, if possible."}, {"start": 3774.75, "end": 3802.83, "text": "Online, what have we got coming? There's a lot online. Don't go away, we'll get through everybody. In many different directions. And one direction is really what's AlphaFold3. You already gave us some ideas of where you want to go. And one thing is dynamics. Obviously, at the moment, it's all static structures. So what about dynamics? How far do you think you can go? Could you go to really look at whole pathways?"}, {"start": 3802.83, "end": 3830.3, "text": "Predict how biological pathways might work, respond to inputs, and then also along what AlphaFall can do. Do you think you will be able, at some stage, to predict function from structure? And that's a summary of a lot of questions. Thank you. Yes, absolutely. We see this as the beginning of a big journey. You can think of it as moving up the hierarchy of complexity."}, {"start": 3830.3, "end": 3852.78, "text": "So, you know, obviously, we started off with this grand challenge and in the sort of static case, we would love to understand protein-protein interactions and protein ligand binding and, you know, all of these things, you know, the dynamic way these systems work. And we have some promising early work on that."}, {"start": 3852.78, "end": 3860.38, "text": "And also we're trying some fundamental things too, like can we understand protein folding from the physics aspect?"}, {"start": 3860.38, "end": 3884.08, "text": "So this is a sort of, it's almost like a translation machine we've built so far with AlphaFold2. You have a language which is the amino acid sequence, and then you want a different language, a 3D structure. So one can think of it almost like a translation model, even though it's not how it works. But can we actually model all the physics and the physics energies and so on? So we have some people who are pushing hard on that, and that might unlock things like..."}, {"start": 3884.08, "end": 3910.1, "text": "You know, activation energies of enzymes and things like that should be obviously super useful in things like protein design. So, and then we are starting to look actually potentially collaborating with Paul and others on pathways and building up slowly towards this dream of a virtual cell. The question is, that's the end goal. How does one get there? You know, alpha folds may be one of those components, but what's the next stages? What would be the interim stages if that was like a 10-year goal?"}, {"start": 3910.1, "end": 3939.23, "text": "And I think we've got some promising ideas for that now. I listed some of them at the end of my talk. And some of those will manifest themselves in Alpha Fold 3. Of course, there's also just the continual improving of the current system and making sure that we get that more and more accurate. And it's not perfect in all cases. And some of you may have some examples of where it doesn't work well. We'd love to understand what those are and improve that in a new system."}, {"start": 3939.23, "end": 3968.85, "text": "In the same way we did with AlphaGo. So one of the ways we do that is another version of AlphaGo, AlphaGo Zero. In that diagram I showed you about always insisting that the new version can beat the old version. We added something to that later on in AlphaGo, which was that not only did it have to beat the old version, it had to do better on this special set of problem positions that we knew the earlier system couldn't."}, {"start": 3968.85, "end": 3990.56, "text": "got confused in, let's say. We collected those automatically because we could detect when the AlphaGo system got confused. Very rarely, but we collected, initially it was about 100 positions, eventually it was about 1,000 positions. Initially, the AlphaGo system that beat Lee Sedol scored zero out of 1,000 on those positions."}, {"start": 3990.56, "end": 4016.22, "text": "Even though it's very strong in general. And then by the end of it, we had a second match against the Chinese world champion about a year later. And by then, it was getting about, not only was it much stronger at beating the older version, it was getting 950 out of 1,000 of those difficult positions right. Now I can imagine, I would like to create, with the help of all of you, a sort of database of problem proteins that we know AlphaFold2 doesn't do well on."}, {"start": 4016.22, "end": 4047.23, "text": "And then we would sort of train against those specifically as well as the overall accuracy. Okay, so to the back first. Hi, thanks a lot. This is kind of related to Peter Ratcliffe's question. I'm sure less articulate in the second iteration. But I've enjoyed doing a small amount of mutagenesis or e-mutagenesis with AlphaFold with my protein of interest. And as I add more amino acids,"}, {"start": 4047.23, "end": 4071.44, "text": "I start to wonder if it can predict a synthetic sequence, given it's only learnt on the proteome that exists. Yes. Well, it's an interesting question. It seems to do quite well on de novo proteins. So there were a couple actually in the CAS competition, and it did particularly well on those, surprisingly."}, {"start": 4071.44, "end": 4086.91, "text": "Somewhat surprisingly, given that obviously the evolutionary information is a key part of the inputs through the MSA. But one thing I would caution against is relying too much on the mutated predictions because"}, {"start": 4086.91, "end": 4116.37, "text": "AlphaFold2, one of the things it's weaker at is point mutation predictions that change the confirmation in a radical way, which obviously sometimes happens in disease. We're trying to figure out, that's one of the main things we're working on. It's one of the main things we get asked for, as you can imagine, useful for biologists. We're working really hard on it. It's a hard problem because in a way, one can think of AlphaFold as being trained to be robust to small changes. You want it to be, if it's"}, {"start": 4116.37, "end": 4144.24, "text": "Thank you very much."}, {"start": 4144.24, "end": 4168.32, "text": "In a way, we got to tell AlphaFold to pay special attention to that somehow and detect that it needs to, in a way. Say, okay, this isn't a normal thing that I should just be resilient to, robust to. It needs to actually take that really heavily, overweight it in some way, one could say, in its attention network. That's one way of thinking about it."}, {"start": 4168.32, "end": 4196.48, "text": "And then related to this is another question we've had is can we, in some situations, can you get multiple confirmations out of AlphaFold, not just one? And again, misfolded proteins and other things. That's another thing we'd like to do, sort of ask AlphaFold again, can you give me an alternate, the next best, the next most energy-efficient confirmation? And that's, you know, we can sometimes do that, but it's not reliable enough yet, I would say."}, {"start": 4196.48, "end": 4218.7, "text": "Yeah, hopefully in the next year or so, we'll have something a lot better for your e-mutagenesis experiments. Right, another one at the back, then we'll come to the front, then we'll go digital, okay? Thank you for that fantastic talk. So you mentioned during the talk, you know, this premise that in vivo proteins fold, you know, instantly, essentially."}, {"start": 4218.7, "end": 4247.95, "text": "However, they don't necessarily always fold just by themselves. So there are contributions from chaperones that will contribute to the folding. So to what extent do you think modeling the contribution of chaperones to stabilizing sort of intermediate conformations could contribute to refining alpha folds for problematic structures? Yeah, no, it's a great question. I think eventually,"}, {"start": 4247.95, "end": 4265.58, "text": "In the limit, we will have to be able to model the context, the chaperones and other things. I think this goes back to the earlier question from online about the dynamics of the situation. If we want to get up to virtual cell or anything, pathways even, we're going to have to take into account all these dynamic things that go on around it."}, {"start": 4265.78, "end": 4293.39, "text": "That was one of the reasons I think chaperones and things like that was also one of the reasons people were telling us alpha-fold would never work. It was basically essentially this conjecture was wrong from Anne Finson. That's the negative view of it, I guess, would be that you can't actually just go from the amino acid sequence to the structure without taking into account all of these things, which clearly happen in real biology. But networks are pretty clever at somehow understanding context."}, {"start": 4293.39, "end": 4313.73, "text": "Even if you don't explicitly give it to them, if it's there in the data, imagine that chaperones do certain systematic set of things with particular type of proteins. Even if you didn't put them directly into the inputs, the consequence of those things might be modelable. The best example we have of that is things like"}, {"start": 4313.73, "end": 4342.59, "text": "Proteins that have pockets and bind to certain things, heme-binding proteins and other things is the one we usually use as an example. Even if you don't give it the heme or what it's supposed to be binding, it makes the pocket just with the missing ion. It knows somehow that that shape, that protein should have that type of capacity, even though it obviously doesn't know anything at this point about the dynamics."}, {"start": 4342.59, "end": 4362.86, "text": "This is the idea of machine learning and AI picking up weak signals that we as human designers would never be able to program in, things like that. But somehow there are enough weak signals, let's say, in the data to allow you to approximate those influences. Thank you. Front mic, please."}, {"start": 4362.86, "end": 4384.56, "text": "Hi, that was an awesome talk. Thanks. I guess I have a more general question. So DeepMind's goal is artificial general intelligence, and I guess the main method that you use is reinforcement learning. Do you think that everything that encompasses intelligence comes down to maximizing points or minimizing risks?"}, {"start": 4384.56, "end": 4407.7, "text": "No, yeah, so I should just be clear. We use many techniques of DeepMind. So deep learning is key. That's why we're called DeepMind, because we bet on deep learning back in 2010 before sort of anyone else knew about it. Obviously, everyone uses that now. I guess what I was meaning is one of the more unique things about DeepMind is we put a lot of emphasis on reinforcement learning, this trial and error learning, this active learning."}, {"start": 4407.7, "end": 4432.9, "text": "I don't think everything is as easy to specify as absolutely, like the kind of points and minimizing risks. That's a whole research area in itself is how to specify goals. Another approach is the AI system could learn what the right value function is, maybe from a meta reward that you give it."}, {"start": 4432.9, "end": 4457.89, "text": "Clearly, as you get closer to human society things, it's very difficult for us to specify what really the objective functions are. Maybe another approach would be to get feedback from the human user or the human expert to train it in the way that's desirable for the task at hand. This is obviously a massive area of research."}, {"start": 4457.89, "end": 4484.85, "text": "Put up the RL diagram. That was one of the super complex things is that little box saying goal. Well, how do you specify it? How do you make sure it doesn't deviate from it in ways that you don't want it to, right? This is the risks part. You specify a goal. How can you make sure that that's the only thing it does? That's actually what you wanted. And what's a nuanced enough language in which to..."}, {"start": 4484.85, "end": 4513.23, "text": "Specify that goal. It's also going to be complicated the more real world we get with our applications. But right now, we're sticking to things that are quite easy to specify, I would say, like 3D structures of proteins or energy systems. All of those science things I showed you have quite, the reason we picked most of them is they have quite clear metrics that already the scientists in that area are trying to optimize for. So it's easy for us. But this is a big question we're going to have to overcome in the next few years."}, {"start": 4513.74, "end": 4541.95, "text": "Virtual. Okay, we have some questions about potential negative impacts of general purpose AI, be it by malicious designers that misuse them or people who badly design their systems. So what do you think should be done to safeguard against this or what can be done? Yeah, I worry about those things too. I think there are different sets of worries. There's two sorts of sets of worries."}, {"start": 4541.95, "end": 4571.78, "text": "One is what they might do themselves. I think we just touched on that with if you misspecify the goal and so on. But I think the obvious bigger worry is malicious humans using these systems in the wrong way. And I think part of that is making sure the general public and politicians and governments understand better what these systems are. What is AI, for example? I think there's a lot of misnomers out there about even the definition of what is AI."}, {"start": 4571.78, "end": 4598.46, "text": "And what isn't. And then beyond that, I think we're going to have to set up a bunch of principles, perhaps even regulations at some point with government about how these systems should be deployed or used once they hit applications. Because obviously that's when they have real world consequences. I think when you're doing research and games and so on, one of the reasons we picked games, it was a safe place to experiment in."}, {"start": 4598.46, "end": 4622.77, "text": "It's a sandbox, basically. So it's the perfect sandbox to which to investigate these ideas, and still is. By the way, I want to get clear, we still use games and simulations for all of our research, partially because of this, because it's a safe domain. But as you start doing more applications, self-driving cars, whatever that is, then obviously at that point..."}, {"start": 4622.77, "end": 4651.25, "text": "Rules and regulation and things that govern those areas already, whether that's transport or healthcare, whatever, need to be updated to deal with this new type of technology. Just in the same way they were updated to deal with mobile and internet and digital in the first place 20 years ago. I think there needs to be another overhaul, careful, thoughtful overhaul that doesn't stop useful, helpful innovation, but curtails this sort of bad usage cases."}, {"start": 4651.25, "end": 4674.82, "text": "Of course, one thing I worry about is because AI is so popular and buzzword now, and say in the startup world, people can raise millions of dollars, but everyone says they're working on AI. And it's ridiculous, like 99% of them, I would say, are not."}, {"start": 4674.82, "end": 4704.96, "text": "I don't even know what it is. Firstly, that's giving a bad name to the whole field, but also it can lead to cutting of corners of things that obviously shouldn't be done, but if one were to do it, you could make money on it, doing face recognition or something, but it would just be unethical. Okay, back, please. Thank you for the talk. How easy would you say this is to completely reverse engineer?"}, {"start": 4704.96, "end": 4725.84, "text": "I want a stick man or molecular motor. Spit me out an amino acid code. Well, this is, again, something we're really actively looking at, and I think others are as well. Can you put, say, alpha fold into the protein design loop? I think this is what you're talking about, a stick man protein, are you meaning? Okay."}, {"start": 4725.84, "end": 4750.19, "text": "So I think there are various ways to try this. Like you could sort of use it as an oracle. Like, okay, I'm going to change the amino acid sequence in this way. Tell me what the structure is. And you could imagine making a cycle out of that. There have been a few archived papers on that. We've tried that ourselves. We don't think that's quite the right way to do it. We think there's more that is needed to be this sort of true generative model, almost reverse model. This does get back actually to..."}, {"start": 4750.19, "end": 4779.04, "text": "Actually, I didn't answer that question on function from online about, well, the dream would be to say the function that you want, and then you just get the amino acid sequence that will produce the right protein structure for it. But I don't think, from what I understand, we don't understand enough yet how structure and function relate. Obviously, there is relation, but a lot of big biologists, scientists I talk to say there's much more to it than structure, of course, all the dynamics we've been discussing. So, thank you."}, {"start": 4779.04, "end": 4806.61, "text": "So I would say it doesn't quite work yet, but clearly this is something on our radar. When I say protein design, that's obviously the kinds of things we're looking at. It should be very useful in future for the type of thing you're talking about. Front mic. So all sorts of really interesting problems in biology that would be susceptible to AI. So practical question, how do we collaborate more between the Crick and DeepMind?"}, {"start": 4806.61, "end": 4826.43, "text": "Well, yeah, we collaborate quite a lot already. And we're actually potentially even setting up a small little wet lab here to help validate some of the predictions that we're doing. I'm just talking to Paul and Sam about that. So then that will also facilitate collaborations."}, {"start": 4826.43, "end": 4849.33, "text": "We've also started a new company that will take this further. I call it Isomorphic Labs. They will just take the biology part of this further beyond what DeepMind can do as just more of a general AI research organization. Once that gets settled and built up, I think that entity will also be able to collaborate more on specific biology issues."}, {"start": 4849.33, "end": 4878.8, "text": "Questions? I'm going to be around later, and then some of my team are here too. We're just across the road, so it's easy geographically to collaborate. The problem is, as you can imagine, since we've done AlphaFold and released it, we've been inundated with collaboration ideas, and we're still a relatively small team, especially working on these types of things. It's all about prioritizing what would actually make the most difference and stack ranking that."}, {"start": 4878.8, "end": 4908.0, "text": "Feel free to let us know what your idea is. Back, Mike, please. Hi, thanks for the talk. So all the deep mind AIs have been like changes that no one expected, big paradigm shifts and leaps ahead. Do you think that that is going to continue to be the case of the way forward? Do you think it's going to be more slow and iterative from now on? And if it is in those big leaps, what is the biological challenge that you think is most dependent on a kind of big..."}, {"start": 4908.0, "end": 4938.03, "text": "Leap or change in way of thinking like that. Do you mean in AI itself or AI applied to biology? AI applied to biology. Okay. Yeah. Well, I think there are going to be a lot more leaps actually. So, or it's going to feel like that, that there'll be leaps. So, I think clearly the big one is can we get to dynamic systems? I think that's the next. I mean, I guess I've answered that and a lot of other questions. I think that's the real. And can we get down to, are our systems really learning?"}, {"start": 4938.03, "end": 4967.6, "text": "Can they learn physics properties of what's going on? So I think those are the next sort of big challenges, and we'll see if that's possible. Obviously, no one knows if that's possible yet, but I suspect that it will be. Thank you. I think we probably are getting worried about Demis' health here. The last two questions here. Hi, so I wanted to ask you about the explainability of AI. Yes. Because as we know, deep learning is..."}, {"start": 4967.6, "end": 4994.3, "text": "To some extent, the black box. And I think explainability of that is quite important when we are getting close to the domains like healthcare and even folding the proteins. So I wanted to ask, to what extent do you actually understand decision-making in AlphaFold? And what should be, in the future, kind of like spread of the resources towards getting the great..."}, {"start": 4994.3, "end": 5011.94, "text": "Result versus actually understanding how decisions were made by the deep learning network. Yeah, great question. I'm glad you asked that actually. So I kind of have a multi-part answer to this. So first of all, absolutely, these systems are the kind of..."}, {"start": 5011.94, "end": 5040.61, "text": "The sort of cliche at the moment is that they are black box systems, which is true, right? You sort of put these inputs, you train them, and then they give you, what you're measuring them by is the output, not really how it got there. But we're also science, so we're engineers and scientists at DMI, so we would also like to understand how, you know, how is it doing this? And perhaps one day you could imagine reverse engineering out some principles that are sort of human explainable about what it's doing. Now, what I would say is initially,"}, {"start": 5040.61, "end": 5068.77, "text": "AI is an engineering science. I call it an engineering science. So how does that differ from a natural science, like what all of you do with biology or physics or chemistry, is that the phenomena you're studying already exists. It's out there. And it's phenomenally complex to study and break down and reduce down into its principles. With engineering science, we have to build the artifact first, and then we can analyze it. But the analysis part..."}, {"start": 5068.77, "end": 5098.1, "text": "There's no reason to think it would be any easier than for natural phenomena. I sort of just gave you a feel for the Alpha-False complexity. It's crazily complex. And so it's going to take the same amount of effort. Now, until now, I don't think enough effort and I don't think a lot of effort has gone into understanding these black boxes. But that's partly because until very recently, they didn't really do anything very interesting."}, {"start": 5098.1, "end": 5126.51, "text": "So if one's going to spend five years analyzing a black box, well, it better be the case that the black box is doing something worthwhile. So until things like AlphaGo, I don't think there was anything really worthy of the amount of effort that would take. But even AlphaGo, our game systems, the best work I have, an example of that so far, is AlphaZero, which plays chess. We just published an archive paper."}, {"start": 5126.51, "end": 5155.14, "text": "With some of the programmers on that system, but also with Vladimir Kramnik, who's the ex-world chess champion. He was fascinated by AlphaZero, how it played, and he wanted to help us analyze the AlphaZero system. Of course, as an ex-chess player, I was fascinated by that too. So we went ahead with that project because in chess, as I mentioned at the beginning, that's one of the most studied areas of domains of AI. Chess programs have been made since the days of Turing and Shannon, right? They even famously..."}, {"start": 5155.14, "end": 5183.3, "text": "And of course, we know what rules should look like for chess, these heuristics that the classical systems use. So we were wondering, could we extract out of our AlphaZero network those rules, things like king safety, queen mobility, the value of the pieces. And so I'd sort of point you at that point. We just released that recently as, yes, it turns out we could actually reverse engineer our..."}, {"start": 5183.3, "end": 5207.42, "text": "What AlphaZero thought about those situations and then compare that to what classical machines could. Do you remember the name of the paper, the title? I can't remember. If you look at me, there's only one paper we've done together. It's quite recent. I think we need to basically, there's a whole much more research to come."}, {"start": 5207.42, "end": 5225.78, "text": "We're a transitory moment in AI. I think in the next 10 years, you'll see loads more tools, visualization tools, analysis tools, for analyzing important systems. It's just only now have we got systems, I think, that are worth like an alpha fold too, that people will even be interested to understand the how of."}, {"start": 5225.78, "end": 5254.4, "text": "My suspicion is that it's not a fundamental problem with these systems. They're black boxes. They are currently, but I think they can be interrogated. My final example of that is the neuroscientist hat on. We have a whole neuroscience team at DeepMind as well. What I say to them is that we should understand these artificial networks at least as well as we understand brains. Currently, we understand them less well because brains are sort of..."}, {"start": 5254.4, "end": 5272.91, "text": "You know, human brain's a bit of a black box too. If I tell you, you know, how are you coming up with these questions? It's a bit of a black box, right? You can tell me some theories, but who knows, right? What's really happening. But of course we can, you know, do fMRI and single cell recording, all these things now, so we're finding out more. But in the artificial brain..."}, {"start": 5272.91, "end": 5286.38, "text": "We have a lot more freedom to do all sorts of things. We did a bunch of ablation studies on AlphaFold2. So that's the first step of analyzing something. So we spent a good few months."}, {"start": 5286.38, "end": 5305.5, "text": "Breaking down which parts of the components of the system we needed, turns out all of them. But why? So that's a little bit of understanding. But we should be, as a lower bound, we should understand as much about these artificial networks as we do about brains, you know, natural brains. And currently we're definitely..."}, {"start": 5305.5, "end": 5330.32, "text": "Less understanding than that. So there's a lot of headroom to go. So if people are interested in that question, we have a whole team looking at building those kinds of analysis tools. We call it virtual brain analytics. And I think it's a fascinating research area in itself. Great, thank you. Great, thank you. It's an excellent question to finish, I would say. So we will finish there. Tremendous talk."}, {"start": 5330.32, "end": 5347.79, "text": "Tremendous energy, still going at an hour and a half. Still some way to go. After it, Demis is going to be involved in one of our smaller discussion groups, so we're going to go straight there after it. But no, thanks, Demis. Wonderful performance, wonderful work. Thanks, everyone."}]}