Using AI to Accelerate Scientific Discovery | Campus Lecture with Demis Hassabis

you know when you have a rock star in the room when you see the room filled and busy and excited and definitely Dennis is a rock star as you all know I made a few talking points it's always a lesson in humility when reading a CV likes of one of them is he started he was a child prodigy in Chess at age four he started by age 13 he was at a master standard okay so that's a good start in life I guess he completed a levels and scholarship level exams two years early age 15 and 16. that's in the British system but it's also telling and Cambridge told him you're too young to come to Cambridge go do some video gaming okay now that's you can see how forward-looking Cambridge is I'm not sure I'm not sure Cambridge told them is to go into video gaming but anyway he took a sabbatical before starting to study that's another message please remember so he in the video gaming company bullfrog Productions then he went to Queen's College Cambridge completed computer science triples and graduated 1997 with a double first another specialty from Dennis and University of Cambridge Zen okay you would expect you would have stayed there you know worked with you know some uh uh fundamental scientists but no he went back into video gaming okay now that's that's a carrier you know that I'm sort of thinking why did I do only one thing with my life uh so and not only did you go back to video gaming he founded a video gaming company at xcr Studios and then he actually returned to Academia to obtain a PhD in cognitive neuroscience at UCL one of the top plays for neurosciences in the world supervised by enero Maguire and he was trying to find inspiration in the human brain for new AI algorithms I think this is extremely interesting because there is still something to be learned from the human brain and he wrote a landmark paper it's the first one he wrote which established a link between the constructive process of imagination and the reconstructive process of episodic memory Recall now this is very interesting because the process of imagination is I think what chat GPT doesn't quite do creativity is not there yet but maybe Dennis will show it to us today after his PhD founded deepmind with two colleagues deepmind which is uh you know the top AI company in the world and it started very quickly making headlines in 2016 after alphago the program they developed the Deep mines at Dennis was a Chief Architect beat a human professional goal player uh in a five-game match Now goal was considered to be much harder than chess chess was done you know years ago as you probably know and people didn't believe that actually AI based algorithms would be able to beat go and that means sure enough proved the rest of the community wrong um then he went on to several other developments of that sort I think the one that probably we are most interested in as scientist is Alpha fold Alpha fold protein folding prediction uh this is I quote Demis this is a lighthouse project our first major investment in terms of people and resources into a fundamental very important real world scientific problem and that's one of the reasons I'm so glad that epfl is giving uh doctoral noise goes out to Dennis because he's true calling in life is to solve fundamental scientific problems so mother of the other companies you can think of that are active in AI mostly want to generate more clicks on ads or something else of that sort even though they pay your bills right uh but I think if I may say but I think what I find extremely exciting at Deep Mind and also the next company that he founded isomorphic Labs is that Dennis is really interested to solve fundamental science problems and that's I think what we are all here to listen to um he has so many prizes I tried to go through your Wikipedia page you know it took several of my screens so I decided not to list all these prices but simply to say that um his work and the work of deepmind got four times the Breakthrough of the year uh mentioned in Science magazine and that's interesting because essentially it's computer science applied to science problems that essentially transform science and I think that's also something very important I think computer science has a key role to play in the 21st century to really attack fundamental science problems outside of computer science okay those of you at epfl and who know me I have said this many times computer science plays a central role it's potentially the physics of the 21st century and the demonstration is done by Dennis hassabis and his team and other people working at the Forefront of artificial intelligence now I'm going to finish by something that is very unusual okay which is uh the British novelist Ian mcaven in a book that you may have read which is called machines like me which is an alternative history so it rewrites history from the 60s to now and you have Alan Turing who is alive and works with his colleague Denis hasabis to really change the face of the World by using Ai and computer science for the good of humanity I could make a court here I I went up and looked a book but in the interest of time I suggest you read the book so very good so this being said as I said there is a small formal part here which is the delivery of the Dr Hal Roy scosa of epfl to Denisa sibis so I'll read the lodesho I'll read it in English so equal Polytechnic Federal confers the doctoral noise causal in to Dennis hasebis in recognition of his Visionary leadership in artificial intelligence and Neuroscience advancing the frontiers of science through his groundbreaking Innovations in AI driven scientific research with this I would like to hand over and congratulate Dennis thank you very much and we are going to thank you thank you very much thank you very much thank you and with this I let you take over using AI to accelerate scientific discovery I have to excuse myself unfortunately but I will watch the video thank you all right enjoy the show Thanks Martin thanks says I'm fantastic to be here and it's amazing honor to to receive the honorary doctorate and uh I've always um it's it's epfl I've always admired epfl and we have many collaborations uh with the fusion lab here and also such an intellectual environment here I've visited CERN nearby and it's an amazing place to do science so it's a real pleasure to be here to to accept this award and give this talk so I'm going to talk today about how I see using AI to accelerate scientific discovery itself and I'll also give just a quick preview of how we got to the point we are now with working on things like Alpha fold and then I'm going to talk a little bit about how I see things going into the future uh using AI to accelerate science so we found it uh deepmind back in 2010 it's almost like another age now it feels in terms of AI time uh but it was only you know of course 13 years ago and back in 2010 nobody was using um uh no one was really thinking about working in AI certainly in in Industry um and it's kind of amazing to see what's happened uh just in a few years uh going from 2010 where it was very difficult to raise any money to work on AI uh in Industry to to now with billions of dollars going into AI projects and it's been an incredible uh process to watch and be part of and we found a deep mind at the end 2010 because we thought a lot of progress could be made if we brought together Hardware advancements algorithm advancements and some knowledge about how the brain worked and get some inspiration from from the brain for certain types of algorithms and systems and bring that all together with amazing engineering and we felt back in 2010 that some very fast progress could be made the kind of progress that we've seen although even that I think has been surprising to even those of us in the industry so we tried to kind of form this Apollo program like effort a deepmind to make bring together multi-disciplinary groups of amazing researchers and Engineers to make the most progress we could as fast as possible our mission statement was to try and solve intelligence and then use it to advance science and benefit Humanity by solving intelligence what we meant was trying to understand the nature of intelligence and then recreate that uh in an artificial construct that I think is the science of AI so we've started off our first big result was with um classic Atari games from from the 1970s and uh our first system called dqn Atari dqn was our first sort of successful big system really back in around 2013 we got this working and what I was able to do is demonstrated I think the power of learning systems it was the first uh really big system that was an end-to-end learn system so um The Innovation here was that we didn't give the rules of the game or the criteria of how to get points or anything to the system all we gave the system was the raw pixels on the screen and it had to figure out everything else from first principles and being told the goal was to maximize the score and that's it had to figure out the controls and the and how to get points and so on and so at the time uh this was pretty astounding because no one had ever sort of done something at this kind of scale uh you know 100 by 200 pixels so the inputs was 20 000 pixel tools and that was pretty huge for any kind of neural net or or Learning System this dqn system went on to kind of learn any to play any Atari game to better than the best humans can play and get them to sort of maximize the score so for us this was a watershed moment I think quite a watershed moment in the in in the field uh in terms of what what these Learning Systems might be able to do if we train them in the right way so then as multi mentioned um probably one of our biggest results starting the biggest one we had in gaming uh was alphago which was our system in around 2015 2016 that we built to learn how to play go uh through a self-learning process that I'll describe and kind of famously managed to beat uh the world champion at go the legendary Lisa doll the South Korean uh Grand Master and why is go so much tougher than chess and and other games to play well partly it's the complexity of go um we've played humans have played go for thousands of years now has invented two three thousand years ago still as popular as ever today over 40 million active players and one measure of the complexity of it is that there are around 10 to the power 170 possible board positions uh in go Which is far more than there are atoms in the observable universe so there's no way to kind of uh Brute Force search through this incredibly complex uh combinatorial space in order to figure out uh what the Right Moves are and the other extra thing that's hard about go is then you'll see a go board on the right there is that um it's quite a esoteric aesthetic game it's one of the things that makes it beautiful and it's very difficult to write heuristics or rules down about how to evaluate how a machine might evaluate a position and and evaluate who's winning um so-called evaluation function which is critical to machines playing these types of games and in chess it's a lot it's um it's because it's a material game it's a lot easier and a calculation game it's a lot easier to actually specify what these heuristics are um but in go that's kind of impossible so people have spent 20 years prior to alphago trying to hand craft these heuristic systems to kind of describe um uh the aspects of a go position and one of the issues is that because go is such an intuitive game even the top human players Grand Masters like Lisa doll they don't really think about it in terms of calculations and rules it's much more about feeling uh and intuition which are traditionally two things that you would say um are difficult for computers to do so um alphago went on to win the match 4-1 in Seoul in 2016. um pretty famous match it became over 200 million people watched the match around the world uh and there's a fantastic uh YouTube documentary on it award-winning documentary on if you're interested to to see um how that played out um but the interesting thing was not only the fact that it won the match of course but it was also how it won that was pretty surprising and actually alphago created new strategies in go that had never been seen before even though we've played this game for thousands of years and um and the most famous one of that was move 37 uh uh in game two of the five game match which astounded the go world and still being sort of studied today and move 37 is that um that Blackstone uh circled by the red there on the board and um and this early in the game it's sort of Unthinkable to play uh that near to the center of the board it's sort of not considered to be uh a a good strategy at all and in fact when the professional commentators were commentating live on this game um they they thought there was an error in the computer system uh in order for it to play there so there was such a surprising move um and it's interesting actually just to touch on and maybe we can touch on this in the Q a is what does um this mean about creativity uh and intuition these sorts of um aspects that we think are very cognitive um and I think when I reflect back on this I I would say that there are at least three levels of creativity and perhaps alphago and systems like that the current systems we have exhibit the first two so if we think about interpolation um that's you know if you imagine you ask us one of these genitive systems to imagine you know show a picture of a cat um maybe uh it will have seen millions of pictures of cats before and then it does some kind of interpolation or averaging of all the cats that it's seen and then it produces a sort of prototypical cat and outputs that as an image um so you know that's an original cat it doesn't exist in the training set but it's a simple sort of matter relatively simple matter of kind of averaging across the exemplars that it's seen the second type of creatives and the higher level is extrapolation where you you sort of sample the distribution uh and so in alphago's case you know maybe you look at all the human games that have been played online and you start building up patterns of what are reasonable things to do and go reasonable positions but actually you then not just simply average uh what you've seen but you actually extrapolate these motifs into new parts of the distribution that perhaps um uh that are not in the training set beyond the training set and I think this is what move 37 is an alphago has clearly and systems like this have clearly demonstrated and I think that's very exciting if you think about the potential for that in science uh to extrapolate Beyond uh make you know models of the current distribution we can see and then extrapolate to come to New conclusions and new ideas that are beyond the boundaries of what we have data for and then finally and I think today systems don't do this you know maybe we can call it invention or out of the box thinking which I think is an even higher level of maybe abstract creativity that of course we as humans do effortlessly um but so far I would say machines don't do and I think the easiest way I can give you an example of that is of course alphago was able to make the these amazing creative strategies in go but it couldn't invent go right or invent chess that's that to me would be a new level of uh invention and I don't think that's impossible I think that's just partly actually the maybe the limitations of the current systems but also the limitations of the way we're able to describe the goals to these systems um so you know I think if you were to get a human game designer like the sorts of things I used to do to try and come up with go from scratch you know you might say something like come up with a board game that is beautiful aesthetically that takes five minutes to learn the rules there's only a couple of rules for Go but takes you know maybe multiple lifetimes to master and you complete a game in you know maybe six hours so within a human uh life afternoon so to be convenient so maybe you would specify these kinds of goals and then go would be you know an answer to to satisfy those goals but you you can see that would be very high level specification if one was to do that which so far our systems wouldn't I don't think be able to cope with that so just to explain a little bit about how alphago works and I'm actually combining several pieces of work here together when I describe this this is sort of simplified description of how all of these types of systems work this includes alphago and its successor programs alphago zero and Alpha zero and Alpha zero actually uh is a kind of General games playing system that can play any two-player board game uh a perfect information game uh better than uh world champion level and so the process is actually quite simple to describe so you start off with a neural network initialize neural network that basically plays randomly and that's you can think of that as version one of the system and it plays around a hundred thousand games of whatever board game is playing that creates a data set of these hundred thousand games and all the positions in those games and then you train a second uh uh version of the neural network on that data set and what you're trying to do here in the neural Network's got two jobs it's got to predict the probability of who will win from that particular position and How likely they are to win which whether it's the which of the two sides will win uh and also what are the what are the more probable moves that um might be tried in a certain position so that's the the sort of two prediction jobs of the of the neural network um then when you've trained version two on this version one data you have a little hundred game tournament uh and you match off version two against his predecessor uh and um and you have this hundred Match Game and then if it wins a significant amount of time uh we set the threshold to be 55 win rate then you replace the um the current uh system with with the new one so you replace version one with version two and then you you do this whole cycle again you play another hundred thousand games but this time version two playing against itself you generate a new data set slightly higher quality because version two is slightly better than version one and then you train a version three and and that that matches off against version two now if version 3 turns out not to be better than version two statistically better then you continue to generate another hundred thousand samples with the current version the current version two so you end up creating 200 000 games then for the new version three to be trained and eventually uh the new version will be the old version and if you do this um system in in in terms of go uh 17 times uh different games require different amounts of Cycles you go from playing randomly to uh uh better than any entity on the planet has ever played go and it's quite a thing to watch live especially in something like chess which I I I know well you know you start the process off in the morning at breakfast um you know around uh coffee time at 11AM you can still beat the system or you know and then by lunchtime it's better than world champion it's actually quite amazing to actually watch that live uh as as you're sort of you know uh going through your morning so it's quite a thing and it's it's very visceral if you play chess or play one of these games games to see that live in front of you that whole process so um what what's going on here then well effectively if one thinks of um a go tree as the tree of all possibilities and you imagine each node in this tree is a go position so you're currently in this you know the current position is the top of the tree and um you've got this enormous you know uh set of possibilities from each position you know uh the 10 to the 170 possible positions and you've got to find this sort of classic needle or Haystack how do we decide what the optimal strategy is the optimal move to play from this kind of position in a very finite amount of time you know maybe one minute of thinking time so Canadian is not tractable to do this in any kind of vaguely you know any kind of Brute Force way or exhaustive way so what we're basically doing is guiding the search with the model so the model is coming up with uh uh most probable moves uh the and the kind of most interesting moves to look at and therefore guiding the tree search to be very efficient in the in the in the budget that it uses the the sort of Monte Carlo tree search maybe it can only search around 10 000 moves per per decision and it's got to apply that very limited budget of search budget to most appropriately to the entire uh to the entire search tree and then when it runs out of time of course then it outputs um the best uh uh plan the best tree that it's uh uh at the best sort of path that it's found up to that point so we've been pretty lucky and using games and and we started with games I mean partly you've seen my background in games and I try to utilize every single strand of the things I've experienced in my career to bring them together to always make use of them uh and I always had in mind even though as Martin was describing all these different things I I did earlier in my career um from probably around 14 15 years old I I had already decided that AI was going to be my career and this is what I wanted to do with my life and then all the other things I picked to do from the phds to writing AI for computer games was um was was sort of in preparation if you like for doing something like deepmind and then we made use of this all this games knowledge and we have quite a lot of amazing games Engineers X Games engineers at deepmind who who work on these simulations and work on these programs we worked on games because we thought it would be the fastest way to make efficient progress with testing out algorithmic ideas um and uh and that turned out to be a case I'm really fortunate to make a whole bunch of big breakthroughs I've just talked about the eqn and alphago and Alpha zero and then we culminated in this program called Alpha star which um uh beat a gram the grand Masters in the most complex computer game strategy game called Starcraft 2 which has all sorts of other challenges from board games like partially observable um uh environment and it needs long-term planning and other other big challenges so with that that you know games has been phenomenal for uh for getting uh making progress over uh almost a decade worth of work to get the algorithms to a point where they were sophisticated enough to be applied to challenging real world problems and of course that's problems commercial problems and Industrial problems and pretty much every uh system that you use now at Google has a piece of deepmind technology in it from the the way the massive data centers are called to the voice on any of the devices that you speak to uh the Google Device you speak to are use wavenet which is the world's best text-to-speech most realistic test to speech system a piece of deepmind technology is is sort of in almost every service but for me in my passion my personal passion as Martin was saying is to eventually was always the plan was to use AI as this incredible tool almost like the ultimate tool for helping accelerate scientific discovery and and I think in the last few years maybe two three four years I think that's because come possible and I think it's a really exciting time now in AI uh to do this to bring over all of these Amazing Ideas that were developed in areas like games and other theoretical domains and bring that to the to real world important problems so for me I always had my eye on and I learned about the protein folding problem and I'll describe that very briefly for those of you don't know what it is uh in my undergraduate in the 90s at Cambridge I had a I had quite a lot of biologist friends there and one of them was particularly obsessed with this protein voting problem and um and still works on protein structure today in Cambridge as a as a as an academic and um and always sort of thought first of all that it would be tractable one day but also that it would be enormously important if one could solve this problem and I and I I really I've always loved this problem I've always thought it would have sort of filed it away in the back recesses of my mind as one day to come back to as we progress with AI because I thought it'd be very suitable for AI for a number of reasons so the protein folding problem is quite simple to explain proteins for those who don't know obviously the workhorses of biology they're essential to all life everything in your body is is pretty much supported by proteins and of course we'd like to know their function uh and their function uh is partially specified by their 3D structure so if you imagine the protein is um is is is specified by its amino acid sequence which you can Loosely think of as this genetic sequence and then um in the in the in nature it folds up into this 3D structure and that's what kind of governs its function now determining the structure of of these of these proteins often takes many many years if you're going to do this experimentally with a lot of painstaking work perhaps many of you in the in the in the audience work on these things and um you know the rule of thumb that again biologist friend have always told me is it can take a whole PhD uh uh you know four or five years to maybe even crystallize and figure out one protein structure so it's it's really a lot of work and um and Christian and finson very famously in his 1972 uh uh lecture for his Nobel lecture uh speculated that the 3D structure of protein should be formally determinable uh by the amino acid sequence in other words we should be able to predict this 3D structure uh immediately uh based just on solely on its sequence now of course that then triggered literally a 50-year kind of Grand Challenge in biology um and um and people have been trying to make progress on this uh ever since and it's obviously a whole field in itself so why is this such a hard problem to solve uh this structure prediction problem the structure prediction problem of the protein folding problem can it be solved computationally and why is this such a hard problem well again uh according to a contemporary of events and uh Leventhal uh leventh house Paradox you know he sort of back of an envelope calculated then maybe there's roughly 10 to the 300 possible uh confirmations that an average protein can take and yet obviously the Paradox is that in nature uh obviously spontaneously you know within milliseconds uh in many cases so how is this possible how does physics solve this then um you know is the big question and the other reason we picked this problem is that there's a fantastic uh really well run competition called Casp that runs every two years and has been running for nearly 30 years now and it's considered to be the gold standard Benchmark in in protein folding to test the best computational systems and it's a blind test so um experimentalists submit their their structures just prior to publishing uh for the competition and computationally after you have to submit your prediction and then later on they reveal the um the ground truth structure experimental structure and then you compare the ground truth against your prediction and uh usually uh for every cycle of the competition they have around 100 proteins that they've collected uh like that that they sort of withhold publication of as many of you know now or did achieve Atomic accuracy so this is accuracy of the computational prediction uh to within uh the width of an atom which is um what we were told was that the tolerance for it to be useful for uh real world biology and chemistry and uh and we we achieve this accuracy at cast 14 the 14th Edition of Casp in in November 2020. and if you look at um the the progress of um Casp and the winning team this is the score of the winning team uh in in each of the additions back the prior 10 years to we entered so from 2006 to 2016 there'd been almost no progress on the hardest category so-called free modeling category where there are no prior templates for these proteins and um and and you know this score on the on the y-axis is a you can think of it as roughly it's called gdt where you can think of it as roughly as the percentage of residues you've got accurately positioned uh to within a our tolerance a very tight tolerance and um and so they were stuck around 40 uh pretty much which is not useful uh in any way for experimentalists it's it's far too large in errors and um and in 2018 we we entered for the first time we started this work in 2016. the whole project is probably the most complex project we've ever worked on and we started it literally pretty much the day after we came back from from Seoul from Korea uh from the alphago match this was the next big project we took on and Alpha fold one we entered into cast 13 and for the first time ever we sort of entered a A system that had cutting-edge machine learning at the core of the system for the first time um and it improved accuracy immediately from the previous sort of winning scores by around 50 as you can see here and then we totally re-architected Alpha Fall 2 um with a whole bunch of new Innovations and taking the learnings from alpha fold one and then that finally reached Atomic accuracy which had led the organizers of Casp um to declare that the problem had been in inverted Commerce solved so how does this look well Alpha fold um takes many iterative steps towards solving a protein structure and you can see this enormously complex protein on the left here is just one example of one of the the structures that Alpha fold uh predicted um and you can see the ground truth in in in in uh in green and and in blue is the is the prediction and you can see on the right hand side how the different iterative steps are alpha fold sort of settles on a final prediction of the of the structure and it's still amazing to me when I watch these things and and see the complexity of then the beautifulness actually of of proteins uh how how this can you know be computationally predicted is still pretty astounding to me so then we went on to um obviously publish all the methods we open source the code and um and then we decided like how are we going to uh share this with the world so that the biology Community could make best use of it and we actually decided that the best way to do this was to create a database freely available database with free and unrestricted access to all the structures that Alpha 4 predicted and we did this in a great collaboration with emble ebi European bioinformatics Institute uh just up the road from us in Cambridge uh in the UK and normally when you do these types of systems you set up a server and then people submit uh experimentalists submit their structure they're interested in uh their sequence they're interested in and then you predict a structure and you send it back maybe a few days later that's the normal way people used to do uh uh uh was to provide access to these systems but because Alpha fold 2 was so fast as well as so accurate we realized that actually over the course of a year we could actually just fold every protein known to science so that's around 200 million depending how you count it obviously growing all the time and uh so we ended up over the course of a year folding every protein um there was and uh that we could get our hands on in the uniprot data set uh database and then we uh we put we made that available of course before we did that just as a as a as a side note and I'll come back to this at the end um we always have safety and ethics at the Forefront of our mind of every work that we do so before we release such a massive database we consulted actually with over 30 experts in many uh relevant fields of course biology but also biosecurity bioethics Pharma and and uh and also our human rights too and from that we actually made some slight modifications to what we were going to to put out there but the overwhelming feedback from from all of them was that the the benefits far outweighed any any perceived risk so since then so it's been out for you know Alpha folds been sort of available for a couple of years now or maybe getting on to three years and um already uh it's been incredibly uh uh um sort of gratifying to see how many uh things that uh Alpha fold has been applied to and the wide range of important problems it's being used for and here's just a small sample of those uh of those uh uh problems that have been tackled using Alpha full predictions and there are many more on our website unfolded.deepmind.com uh where you can see many more sort of testimonies and and examples case studies of of how alpha Falls being used and I'll just highlight a few here um it's been used by the University of Portsmouth to do design sort of plastic eating enzymes to tackle plastic pollution it's been used all over the place in big Pharma for drug Discovery from and also in Academia antibiotic resistance and malaria vaccines and also neglected diseases which is very important part of our mission uh neglect a tropical diseases is like leishmaniasis and Dengue that we think I often don't have a lot of investment in from Pharma that the ngos that work on these types of diseases uh it was great for them to get an immediate leg up by seeing the structures of the proteins that they were trying to Target and obviously do get straight into drug Discovery and then it's also being used in fundamental research too for for example helping with finding the structure of the nuclear poor complex and also designing things like molecular protein syringes for delivering payloads uh into the body so incredible diversity of of different things and this is always what we hope for for working on this problem of of protein structure prediction was that if it could be done it would unlock uh if we were in this new world of sort of protein structure abundance it could unlock that all these Downstream amazing pieces of research we've been very lucky as well in the terms of it's being recognized after fold by very nice uh accolades like Breakthrough of the Year from science method of the Year from nature but also the amount that is being used has been incredible so over a million researchers and biologists have used Alpha fold in the databases from all around the world pretty much every country in the world and the methods papers had over 10 000 citations already so just in terms of the higher bigger picture then if I step out of bed now and start looking at where is all this going um I like to use this phrase of science at digital speed and I think we're going to start seeing this more and more now and what do I mean by that well I mean science at digital speed in two different ways um first of all if you take Alpha fold um it's extremely fast solution of course so you can fold uh get the structure of a protein in a matter of seconds instead of potentially you know months or years if one will have to do that experimentally so the solution itself is very fast and that allows us to scale up to you know 200 million proteins but also the dissemination of that information is also at digital speed because it's a it's a it's a it's a technical solution so you know the moment you've you've kind of built the algorithm and you folded all the proteins we can then put it on the database and everyone in the world can then access it uh just with a simple keyword search so if you compare that to an experimental breakthrough you know amazing ones and they go they still made me take a decade it can take or more to propagate through training phds on those systems building the right Hardware uh equipment in order for those to work and training people to use those new techniques um and instead of you know something a science of digital speed is able to propagate almost immediately um just like any other piece of software or technology so I think this is maybe the first example of that but I think it's an interesting phenomenon that I think we'll start seeing a lot more of and just following up on that theme then if I I I like to think of that maybe we're entering a new era of what we like to call digital biology and um I think if you think about biology at its most fundamental level I think we can think of it as an information processing system so um I'll be at a very complex emergent one and um if we think of it in that way then in the same way that maths was the perfect description language for physics and physics phenomena perhaps um I think you know AI could end up being perfect for this type of complex emergent regime to describe and model uh what's going on in biology I think it may turn out to be the perfect fit for AI so I'm hoping that Alpha fold is a sort of proof of concept if you like and when we look back on it maybe 10 years from now uh it won't just be this isolated uh success story but actually we'll see that it heralded this new era uh of uh digital biology and we've actually followed up on this ourselves by starting a new spin out company called isomorphic labs to try and reimagine the whole drug Discovery process from first principles with AI um of course the pro to you know Alpha fold and the protein structure is just one very small aspect of doing drug Discovery right where one small piece of the whole puzzle an important piece but um nonetheless just one piece so we need to fill in all of these other modules that that are required actually to do drug Discovery from Target Discovery to uh designing chemical compounds and designing drugs so ifosomorphics approaches to sort of try and Tackle these most fundamental problems in drug design with this AI first or computational first approach and then we combine several advantages from different areas we have the new company has close collaborations with uh with deepmind and they are Ai and science teams there it also has huge resources and funding from alphabet the parent company of all of these companies uh to allow us to kind of think long term and ambitiously which um and so we can really try and go after the fundamental research to begin with to build this platform uh this general purpose platform a bit like we did I showed you with games uh at the early part of the talk and we also get to utilize of course all of Google's uh enormous sort of infrastructure which of course is incredibly important for computational and AI approaches and we get to do all of that so sort of take advantage of a big company backing but within the context of a fast-paced and Nimble startup environment we pulled together this world-class multidisciplinary team I'm a real believer in multi-disciplinary research work we've always had that a deep mind and we have that also isomorphic incredible team across AI chemistry bio physics and and engineering and we were lucky enough to have this amazing scientific Advisory Board that's been set up with um several Nobel laureates and uh excisely you know we have offices in London but we've also opened uh an office here in lausanne actually in the Innovation Park uh just locally and I have some of my colleagues here actually here today from from the lausanne office which is amazing so another great reason for me to be visiting so that's where we're going now and I swear I see uh these types of techniques and I think we can you know isomorphic in a deep mind we're trying to create maybe half a dozen more Alpha fold-like systems but in different parts of the of of the chain that's required to do drug discovery so maybe if I finish Now by just taking a further State back and looking at the holistically about all the work that we've done over the last decade plus and I think it goes back to that search tree Irish I was showing you um I think a lot of AI can be kind of thought of in the most General way in this way that you have these enormous combinatorial spaces um you know whatever the problem is that you're trying to solve and there's huge numbers of combinations that can be uh that need to be searched and what you do to make that tractable is you learn a model of that environment either directly from data or from simulations uh in in some cases or ideally you have both and then you use that model to guide the search in a tractable way according to some objective function and I think it's pretty simple to describe but actually if you think about AI in this way um I think there's and actually start looking for problems you can apply it to it turns out this is a very general solution and actually many problems can be uh modified to fit this pattern so here's the go tree again um with all the the search tree and and how it finds this optimal plan and the optimal moves um and now if we replace um this is obviously a cartoon but it just gives you the idea if we replace go positions now with chemical compounds you know one could imagine uh building up uh new drugs or or new chemical compounds that are useful for things using a some kind of search uh I like this but in this case the model will be modeling uh chemistry space and chemical properties we've actually had a pretty amazing couple of years now in Sciences um applying these types of systems and other systems to all sorts of scientific domains not just biology we've done some great work in quantum chemistry in pure mathematics solving some important conjectures in collaboration with some amazing uh field prize mathematicians Fusion here collaborating with the Fantastic Team here epfl to control Plasma in a fusion reactor and um and then work on genomics weather prediction all sorts of uh domains and I think it's just the beginning of what we're gonna see this sort of real flourishing of applying AI this kind of general purpose AI general purpose Learning Systems to pretty much every branch of science so just then a word of it's on on pining responsibly and and how is one uh undertake this work carefully and uh and ethically it's very very exciting time I think AI has the incredible potential to to help with almost all of Humanity's greatest challenges from um curing diseases to uh energy and sustainability in the environment um I think AI has the potential to help with all of these problems but um of course it's an incredibly powerful technology and um and it can be used uh it's a dual purpose technology depends on what we as society decide we're going to deploy and use it for and and how we build it and it needs to be built responsibly and safely and to be used for the benefit of everyone and as I mentioned from the beginning even back in 2010 when we were little tiny little office space in an attic in in London um we were planning for Success when we when we started in 2010 we imagined a world like we have we're in today and actually even Beyond where we're going to be in the next 10 years and uh we imagined that and we so therefore if one were to imagine this kind of world you would be thinking about ethics and safety from the beginning and that's what we were doing and it's been Central to our mission from the start we've always had an Ethics Charter um and uh an Ethics Committee and that's actually evolved now into the whole of Google having their published AI principles so now we're part of Google uh we we help across the board now uh in across the whole company uh with applying these principles uh correctly um everywhere where we deploy AI and we continue to try and provide this thought leadership on AI strategy globally risks and ethics and safety for the whole research community and Beyond so how should we Pioneer responsibly this is what obviously I'm sure you all agree this is what we should do with this type of Technology um and my my view is that we actually have a a way to approach this um and we call this artificial general intelligence you know when we get to human level uh uh cognitive capabilities across the board with these systems and I think we're approaching that pivotal moment in human history quite rapidly maybe it's a few years away um and my view is that we should not move fast and break things which is the kind of Silicon Valley Mantra that of course is very successful in building social media and huge companies and getting uh very uh uh great growth for applications and Technologies and it served us very well but I think when it comes to something as consequential as AGI I think instead we should be sort of looking towards the scientific method which of course you'll all be familiar with here and I think that's the approach that we should be taking talking with these kinds of systems so thoughtful deliberation hypothesis generation and rigorous testing of those hypotheses then carefully controlled environments and conditions detailed analysis updating rapidly on empirical data perhaps with external review and all with the aim to get a better understanding of these systems before we deploy them in the wild and I think the key thing here is sort of not so much not to move fast because um you know I think we need to move at PACE with these systems because of the enormous potential they have but ideally not to break things right or as far as possible not to do that with uh and and use as much foresight as possible so I think a transformative technology like AGI requires exceptional care and and our view at a deepminder Google is that we need to be bold and responsible with this technology and it really is an ad and I think I think although those I think those two words um there's some tension there I think it's a good tension to have so I'll just finish them by saying if we realize the full potential of AGI and and build these types of systems I think there'll be the ultimate general purpose tool to help us as scientists understand the universe and perhaps our place in it thank you very much [Applause] foreign [Applause] for this impressive demonstration of the power of the tools you've uh you've built we have sometimes for a few questions for the audience so the house rule is you need to come to the standing microphones that are situated here and there make a line if you have uh if many of you wish to to ask a question and we'll take them one after the other in the in the time that we have at our disposal I have maybe one first question yes white people can come into into the line you've shown how it worked your your system on go on on video games and then how what kind of mental shift did it take to someday say hey it would probably work on proteins as well well um for me it's you know maybe it looks like that if you um you know I think the missing piece of information if you if you look at my career is that I had already decided uh and actually through chess originally it was games that got me to AI because um most of you too young in the audience remember this but when I was playing chess in the 80s uh sort of semi-professionally when I was a kid I was using chess computers to train on and they were the originally they were these physical boards that you would actually press the squares on and they had little LED lights and uh and I was fascinated I must have been about eight nine years old when I got my first one that that uh how were they programmed you know this this machine this this bit of plastic in front of me you know how was it playing chess and helping me train and I think that that was I kind of determined at that point that I wanted to learn actually how this how to program machines to do that and in some ways that was more fascinating even than than the game itself was actually how these systems were were programmed and so then I went into professional computer games development but if you look at all the games that I made Park all of these things actually the Common Core theme was it had AI as the core part of the gameplay and again that just reinforced I saw something my probably most famous camera was theme park when I was around 17 years old and it was designing your own Disney World and then little people came into your Disney World theme park and played on your rides and designed the park and they all had AI controlling them and when I saw people enjoy that so much to how the game reacted to the player I sort of realized how much AI could bring to the world so then everything else I've done and sort of collected was with this in mind where where what we see today in mind thank you very much so we'll start on that side please make short questions hello uh thanks for an excellent talk uh your body of work and the journey has been very inspiration to me so the my question is the following when you are building for the first time these complex systems like Alpha zero I'm pretty sure many people have tried before and fed so for what for you what was like the conviction or the inspiration that somehow will build these systems that will turn out to be like a amazing breakthrough so can you give us a sneak peek into sure so so back in um perhaps when I was when when I was doing my undergrad um one of my really good friends at Cambridge was this guy called David Silva who ended up being he's one of the best reinforcement learning people in the world but also he um ended up being the lead Project Lead on alphago and we used to talk about and um he was also part of my first games company he was the CTO and um we always used to think about go actually which I learned at Cambridge there was a lot of um there's a really good go Club there and we both learned go there and we were always fascinated by how would we program that and we tried actually in in the early 2000s to do it in the way that the classical chess programs are being written with these heuristics and search and it kind of gets you nowhere you can't even you know get to strong amateur like that so we sort of shelved it for uh 10 years really and we decided then that actually you needed a learning approach that just confirmed our intuition that there's no way that you can we can enumerate out all the heuristics for something complex like that um even something like go let alone uh more General things than that um I think that reinforced our opinion that you needed learning systems and that were General thank you thanks thank you very much let's take a question from the other side of the room hello hello thanks for the great presentation um fascinating to see what's going on in this in this domain especially uh I'm glad to see that there's a progress in terms of drug Discovery and I can imagine that investors are queuing up and so um we also know that in this particular domain of Industry not all the well let's say some of the larger investors have proven that um sometimes colossal profits are more important than any ethical consideration so my question is how do you it is isomorphic Labs think about the monetization aspects and do you make any promises in terms of um um enlarging access to medication for the entire human population including the large part in the majority Who currently don't have it yeah well look I I I I can't make any promises but I can definitely tell you the the intention is and the idea is that if this all works out you know with this computational approaches which should be maybe even an order of magnitude faster to develop new drugs you know maybe instead of 10 years one year you could imagine and therefore maybe an auto magnitude cheaper to develop those things you know perhaps more perhaps even too old of magnitude as possible and then I think um if you have both those two things it's far easier to distribute that and work on many many diseases not just first world diseases in wealthy countries you know where I think farmer traditionally has to make its sort of calculations and actually you can do a a lot more and I think we've seen the beginnings of that with Alpha hold as I mentioned one of the places we collaborated with I think they're actually based in Switzerland was the dndi drugs for neglected diseases Institute it's part of I think who and um and uh we folded you know all of the proteins and all the the diseases that affect mostly the developing world and the poorer parts of the world but affect millions of people there hundreds of millions actually but are generally underserved by the current Pharma industry so I really hope that the work we're going to do is going to allow us to uh uh uh um broaden that access and then maybe also make uh you know act as a role model to Pharma to to do more of that too like open sourcing medication yes for sure exactly for these for these um uh for these you know these uh these these sort of diseases that affect the developing World more yeah that would be wonderful people are also queuing up to ask questions so thank you very much uh so quick question so uh what happened what turned out to be harder than your thoughts and would turn out to be easier than you thought over the last uh well 13 years yeah so um you know in in a way there were so many it's kind of hard to remember because there's so many different projects we've done and at all times there were difficult moments I think one of the hardest moments was after Alpha fold one um trying to push Alpha fold one to reach the atomic accuracy so that's the first thing we tried to do in roughly 2018 was push that first system and after around six months we realized it couldn't go any further and we had to go back to the drawing board and use what we learned but actually re-architect everything and for a long time probably another six months maybe even a year it would the new system Alpha Fall 2 was worse than Alpha fold one and you have to you have to be very confident and brave and persistent when you go through that those periods to you know just trust that it's going to come back and then it has more potential to go further because of the way you've re-architected it so that was um that was with Alpha Fall 2 and then maybe uh on what when what's gone easier I guess you could imagine language modeling has perhaps surprised everyone I think on the whole research Community including us of how um that may be relatively easy that was uh once Transformers and rlhf had been invented you know five six years ago thank you thank you we have five more minutes to go so I'm afraid we won't be able to take all questions time moves very fast but please next one okay so first off thank you very much for your work amazing and I want to ask if there are any plans to address any quality in regards with the uh alums or students from non-elite universities which usually have less funding less collaboration opportunities less visibility on their work so do you plan to address that because it usually is also an indicator of socio-economic uh circumstances and usually most of the inclusiveness work is regarding the individual circumstances and not so much the context of the person yes we we care about that very much and actually we have very big uh uh diversity and inclusion efforts at deepmind and um we we sponsor it's mostly in the UK currently but we're going to try and make it more International hundreds of Masters students scholarships from underrepresented backgrounds both socioeconomic and other types of underrepresented backgrounds and and um so and we've encouraged the UK government to do that so actually now they provide thousands of scholarships to the Masters level for machine learning so the idea I think we picked Masters level because that's a it's almost like a conversion course from stem undergrad you know science undergrad or math undergrad and then getting them into machine learning and then either phds or into industry so that's worked really well we've also sponsored um uh funded several chairs in machine learning at many of the universities uh in different parts of the regions of the UK and also internationally and then recently we did a five million dollar sponsorship to the African Institute mathematical site uh Sciences to kind of train the best students in Africa who are already in the mathematical Society uh in machine learning uh with uh you know and actually one of we've seconded one of our senior research scientists to be the director of that in in South Africa of the whole Institute uh for the next I think five years so we're trying to do a lot of things around that but all needs to be done thank you very much great thank you very much next question please yes thank you so much and um more than 10 years ago you bet on something that few people believed in and now ai is something that everybody is jumping in so um what do you imagine AI can do not for science but also for Humanity in general in the future how far it can go yeah I I think I've always imagined um there's almost no limits in some way to AI because if you look around us in modern civilization how did we create modern civilizers what we see all the amazing things we see around us um it was with intelligence right so um and I think if you think back I do you know if you do thought experiments like you imagine going back to uh uh pre-agricultural era and you you talk to a tribes person then about you know one day we'll build Manhattan in New York and fly over in a 747 uh in 10 hours you know it's it's would be Unthinkable and I think about intelligence human intelligence has done that and I think that there'll be almost as big another Revolution with these tools and and helping us solve all incredible problems uh that we think are quite intractable today maybe like energy sources material design I'm thinking as well as biology and Drug design curing hundreds of diseases um I think that's all within reach uh and then I hope we would have a much more um you know obviously you can call it radical abundance much more uh equality in the world and lots of these huge problems some of which we've created for ourselves as a society like climate uh we have we'll have really great solutions to um with AI helping the best experts in the world to solve these problems thank you thank you very much one here hello I'm Leon from open knowledge Association and I have a question linked to the use case of infrastructure do you intend to to use infrastructure AI in the infrastructure field for example for Waste Management or like can be simulation for the buildings or stuff like this yeah great question I'd love to do that I think there's huge potential in using AI to optimize existing infrastructure transport and power grids and so on I think to get maybe even like 30 40 more out of them that's what we found with the data centers we optimize these massive data centers that Google uses and Powers all search YouTube and so on we save 30 of the energy uh from the cooling systems by using AI to control them and they already were quite optimized systems but you can even imagine things like shipping shipping lanes I mean there's a huge amount of pollution is caused by that and just being more optimal on Logistics as well as inventing new Solutions like the plastic heating enzymes I mentioned or designing ways to deal with waste uh recycling I think all of that AI could help that you know that entire uh industry actually if we if we embed it in that thank you very inspiring thank you sorry to be the nasty timekeeper but this would be the last question first thank you for the very nice presentation really looking forward to the progress of Science of our domestic next decades my question is how do you see the role of Open Source in AI over the next decades and what do you see to be the most the best approach to AI in terms of balancing the the risks of of AI of course it's harder to maintain safety when everything is open but at the same time letting everyone contribute and using all of the brains in the world for this amazing development yeah now fantastic question and a complicated one right as you might expect you've actually said that you know you've explained that in your question so up till now uh you as you'll know like deepmind's pretty much published everything we've done uh uh all our nature science papers and open sourced Alpha fold and other things I think it's got to be on a case-by-case basis as the systems get more and more powerful more and more General the problem is if you just open source naively these systems um Bad actors can whether that's individuals or nation states even can obviously access it militaries and so on the same as scientists so it's a complicated problem because obviously one would like to do open science and and and of course things progress much more quickly like that and it's good to get external review of your systems um but how do you also restrict access to to Bad actors who might want to use it for um bad means so I don't really have a full solution for that maybe there will have to be more um access to like individual departments or universities or academics that are known to be working on the right things perhaps that's one way of doing it with the latest Frontier models or um but it's not a it's you know this is a current debate that's going on as you're probably well aware of how to balance the risks with um the benefits of doing that so um you know it's going to be a complicated question we're going to have to try and resolve going forwards the ideal case I think would be to develop and I would encourage any of you in Academia to be working on this is like more interpretability analysis of the current systems and we're not doing enough of that research on that work I would say relative to capability development so that we could develop evaluation tests for these systems so we can really rigorously know before we release them open source and other things that they're safe in certain ways what the properties they have so um so they can be perhaps have some guard rails around them that would be the best right and then you can put them out into the world for people to experiment but there's a constraint around what can be done with them maybe something like that but the problem is we don't know yet well enough what those robust evaluation systems should be yet thank you thank you thank you very much thank you all for your questions as well and your attention today I think we can give them as another round of applause thanks thank you fantastic

Channel: EPFL

Views: 24,260

Keywords: EPFL, science, technology, university, switzerland, Lausanne

Id: Ds132TzmLRQ

Length: 65min 6sec (3906 seconds)

Published: Thu Jun 01 2023