{"url": "https://www.youtube.com/watch?v=GF1yCmxOgiQ", "page_url": "https://www.youtube.com/watch?v=GF1yCmxOgiQ", "title": "2015 IDE Inaugural Conference 7 Suleyman", "show_or_channel": "MIT Initiative on the Digital Economy", "date": "2016-10-26", "duration_s": 2371, "type": "talk", "speakers": ["Mustafa Suleyman"], "confidence": "medium", "notes": "MIT Initiative on the Digital Economy inaugural conference (2015), DeepMind applied AI talk; posted 2016. Content not checked beyond title.", "people": ["mustafa-suleyman"], "id": "2016-10-26_mit-initiative-on-the-dig-2015-ide-inaugural-conference-7-suleyman", "audio_duration_s": 2370.4555, "worker": "jc-w14", "transcribe_s": 31.5, "model": "large-v3", "segments": [{"start": 7.86, "end": 24.45, "text": "So let me introduce our next speaker. I first met him only five months ago down in Puerto Rico at a conference on artificial intelligence that the Future of Life Institute was pulling together. I was so blown away by what he and his company were doing, I right on the spot said, hey."}, {"start": 24.45, "end": 54.06, "text": "You've got to come up to MIT and share what you're doing with the folks here in this room today. And I'm thankful that he was able to join us, flew in from London. Mustafa Suleiman is the co-founder of DeepMinds. Artificial intelligence in the area of deep learning and machine learning is, as you know, taking the world by storm. There's been amazing changes that are happening in a lot of companies. Google recognized that, bought the company last year for $650 million. It's already having some big effects at that."}, {"start": 54.06, "end": 92.8, "text": "At Google, in areas in image recognition, voice, fraud detection, I'm sure Mustafa will talk a little bit about some of the things they're doing there. And increasingly in companies and for people all over the world. He's got a lot of interesting things to say about it, so I don't want to take up any more of the time and just invite Mustafa Suleiman up to join us now. Thank you. Hello. Hi. So..."}, {"start": 92.8, "end": 122.42, "text": "My name is Mustafa. I'm the co-founder of DeepMind. And we're building artificial general intelligence. So before I begin, I want to say that what I'm about to talk about is a long, long way away from what we've been discussing this morning. And I guess the invitation would be to take a moment to step back for the next half an hour and come with me to 40,000 feet where we look at what"}, {"start": 122.42, "end": 150.93, "text": "Life might be like in the next 20, 30 or 40 years. So first of all, what is artificial general intelligence? So I'm specifically talking about systems which learn automatically from raw pixel inputs. We talked a little bit earlier about the difference between horses and humans. And somebody mentioned that the defining characteristic of humans compared to everything else that we've developed in the world is that they have this capacity to learn."}, {"start": 150.93, "end": 166.9, "text": "They don't need to be pre-programmed. They don't need to rely on heuristics. We can learn and improve. Well, this is specifically what we're trying to imbue in our artificial systems. And one of the key characteristics of our systems is this focus on generality."}, {"start": 166.9, "end": 195.81, "text": "So we're specifically trying to develop systems which on their own, without any pre-programming, without any heuristics, with very minimal training, can perform well across a very wide range of tasks. And in fact, that's actually the key definition of intelligence. So we're focused on the most simple thing that an agent or a system can do, provided that they can do multiple things in parallel simultaneously. And this is..."}, {"start": 195.81, "end": 222.94, "text": "As distinct from narrow AI, which is pre-programmed for a very specific task. Now, we see narrow AI all around us in the world every day. You'll probably all be familiar with Deep Blue, the system that beat Garry Kasparov in 1997 at chess. An incredible supercomputer, a very impressive achievement in engineering terms and in narrow artificial intelligence. But when you ask it to play..."}, {"start": 222.94, "end": 252.18, "text": "Checkers or tic-tac-toe, it's completely incapable. And that's actually one of the key things that we're focusing on. How can we develop systems that can do both effectively extremely well? This is artificial general intelligence. It's flexible, adaptive, and it learns. And we use the reinforcement learning architecture to inform the kind of philosophy of our design and engineering. The key characteristics here are that we have an agent."}, {"start": 252.18, "end": 274.02, "text": "or some system, some set of algorithms that operates in some environment and the agent has a goal or has a policy of behavior, some model of the world and some notion of what it thinks is going to be successful in that environment. Now the agent can take actions in the environment and the environment also passes back"}, {"start": 274.02, "end": 301.42, "text": "observations and through this cycle of reinforcement the agent learns to improve its policy or its set of behaviors for how it interacts with the real world. This is obviously a very intuitive concept and it's the way that in fact humans learn and many other animals learn through direct interaction with their environment. But so far in the history of artificial intelligence over the last 40 years our systems haven't been capable enough of learning in this way."}, {"start": 302.86, "end": 318.0, "text": "So the environment that we've chosen to experiment with over the last three or four years is the old school Atari 2600. And this is 100 plus games from the 70s and 80s."}, {"start": 318.0, "end": 347.86, "text": "The advantage of using a system like this is that it's been independently developed, obviously, by another team. There's a very specific reward, which we can use as signal to train and improve our agents, and also that we don't imbue any of our own intelligence as human designers and biases by creating our own synthetic training or test environment. In some sense, this is a crude and small-scale approximation of interacting in the real world, which is where we're headed to."}, {"start": 348.66, "end": 369.79, "text": "So agents just get the very raw pixel inputs, around 30,000 inputs per frame, and they're wired up to the action buttons, but they're not told what those actions do. The goal is simply to maximize score, and everything is learned purely from scratch. So there's absolutely zero pre-programmed knowledge. We don't tell the agent that this is fire or this is move left or right."}, {"start": 369.79, "end": 395.2, "text": "Or that this is an alien and it might shoot you. Or that you need to collect keys in this level. We don't tell the agent anything whatsoever. It gets a raw feed of RGB pixel inputs. This is a single agent that learns to play all of the different games in Atari, or 57 that we've experimented with. And it's a single set of parameters. So we've designed an architecture which in itself is inherently general."}, {"start": 395.2, "end": 423.28, "text": "And which we believe in the long term has the potential to scale to very complicated real world problems and in the limit human-like intelligence. So the way to think about this is think of a robot in an arcade watching exactly the screen and being able to control the joystick. So there's no hacking involved in this whatsoever. Now I'm going to try and show you some videos. Hopefully I'll be luckier than some of the other guys. Yes, excellent, okay."}, {"start": 423.28, "end": 442.75, "text": "So this is Space Invaders, and this is before training. You can see that the agent at the bottom is controlling the agent right down at the bottom of the screen. Not really having any success pre-training, just trying to figure out what's going on in the world, randomly firing. After training..."}, {"start": 442.75, "end": 470.46, "text": "The agent has learned to control the spaceship so that it hides behind the obstacles as much as possible. Every single shot that it fires hits something. It deliberately goes for the enemies right at the top of the screen because they're more valuable. It has an incredible ability to predict shots, as you see there, that it hits the purple mothership right at the top of the screen. And as you can see, it's doing a very good job of basically avoiding all obstacles."}, {"start": 470.46, "end": 490.58, "text": "Whilst also, as you'll remember for those of you who played this game, it sort of speeds up towards the end and it does an amazing shot to get the final enemy and clears the whole board. We could leave it like this for a very long time. So if we can go to the next slide. Perfect. So this is the game of Breakout, if you want to run the video."}, {"start": 493.46, "end": 522.91, "text": "So in Breakout, you control the paddle at the bottom and the aim is to try and get clear as many blocks as you can in as little time. This is before training or after just 100 games of training. The agent really doesn't know what it's doing. It's randomly moving the paddle around at the bottom. After just 300 games of training, the agent is now pretty good. It basically doesn't really miss, which is about as good as some of the best people in our office can do. This is a really difficult game and you'll see on occasion the ball comes down."}, {"start": 522.91, "end": 553.7, "text": "Vertically. After 500 games, the agent develops a really interesting strategy of tunneling right to the back and firing the ball up the back to get maximum points with minimum effort. It's essentially learned this exploit. I mean, it's very difficult to learn this exploit. We actually had one of the best Atari players in the world come through the office and benchmark all of our games. And obviously, we end up beating him. So just to give you a sense of how general this player is, here's a medley of other games that we've played."}, {"start": 553.7, "end": 584.53, "text": "This is Enduro, a racing game where the agent has to dodge all of these obstacles. This is River Raid. The agent has to fire things that are in its path and collect little tools along the way. Battle Zone, a completely different physics, a very different structure to the environment. Pong, you'll be familiar with. Our agent here is in green on the right-hand side. Sequest is a really cool game. It has to collect divers, go to the top to get air and shoot the enemy fish. The boxing game here, you can see our agent spends a little while"}, {"start": 584.53, "end": 607.44, "text": "Pretending that it's even, and then manages to panel its opponent right in the corner and basically doesn't lose a point from there on. So that gives you a sense of how general some of these are. Can we go to the next slide? And we were recently recognized with a really exciting Nature front cover just a few months ago, which was a really great achievement that we were super proud of in the office."}, {"start": 608.37, "end": 634.9, "text": "We've since moved on to completely new environments where the physics of the simulation is very, very different. This is the open racing car simulator game where, again, the agent gets nothing apart from learning completely randomly. This is after a good deal of training. The agent basically learns to drive pretty smoothly. I mean, we're not going to be in the self-driving car anytime soon, but essentially the agent has learned to steer, to accelerate optimally, to brake."}, {"start": 634.9, "end": 668.22, "text": "And occasionally use bouncing off the edges to accelerate its speed and take over its opponents, which again was an exploit that you wouldn't really naturally think of as a human, but it's kind of a bug in the game. Can we skip? Thanks. Just a sec. So we're also, I should just explain what's happening here. So we're using exactly the same system with some modifications and tweaks in the last six months to learn to control a robotic arm."}, {"start": 668.22, "end": 697.39, "text": "And so what you'll see here is a video of the system learning purely from raw pixel inputs and from 2D vision. So just as you and I get to look at the screen now when we watch the video, the agent sees exactly the same thing, but it gets to control a moving arm. And the objective here is to try and catch the ball in the cup. After very few attempts, the agent is able to learn this problem purely from 2D vision."}, {"start": 697.39, "end": 726.42, "text": "And you'll see that this is a classic game in robotics control. Back in 1996, this was solved through a million trials of experimentation by an older agent. 2005, they'd got it around to around 300 trials. Today, it takes us 10 minutes and 10 trials. The reason why it takes 10 trials is because in between each experiment in the real world, the agent does a simulation of what it's learned from that, and without interacting in the real world, it learns to imagine"}, {"start": 726.42, "end": 751.7, "text": "What would happen if it tried lots of different control parameters? Somebody mentioned imagination and creativity earlier as being the kind of preserve of us intelligent humans. In some sense, you could think of imagination as this capacity to simulate without interacting with the real world. Offline, in our own minds, what would happen if I threw this into the audience? I don't have to do it to learn what will happen there. And in some respects, our agents are learning to do that."}, {"start": 751.82, "end": 768.43, "text": "We've also obviously applied it to many real-world problems in Google, and the Google Brain team, our partners in Google, have done an incredible job with some very impressive applications. I'll quickly run through a couple of them. You can see here that we're using the same deep nets that we use to improve..."}, {"start": 768.43, "end": 798.02, "text": "On the existing error rate of image recognition across Google products by a phenomenal amount. So in 2012, it was 16%. Today, it's 6.6%. This is better than most humans are able to recognize. Now bear in mind, on the ImageNet challenge, there are 14 million images taken from Flickr. So these are really realistic, natural photographs that real human beings have taken. And it's often very difficult to say what is actually in the screen in most situations."}, {"start": 798.02, "end": 819.66, "text": "And this is now in production in G Plus Image Search as of 2014. We also do text transcription. So look at this noisy street view scene here. There's all sorts of text and numbers of different sizes, different fonts, different languages. It's not clear whether it's a logo or whether it's real text. We learned to recognize this with..."}, {"start": 819.66, "end": 846.26, "text": "In near perfection, so here you can see our tool now deployed in Street View that is recognizing numbers and doing translation and even reading menus on Street View signs, which is how you'll see that Google Local Search and Google Maps Local Search has got so effective recently, both in identifying where places are but what services they provide and other characteristics that are learned purely from the raw pixel inputs of the Street View image."}, {"start": 847.15, "end": 876.56, "text": "This is a really cool development in the last few months. So here is obviously a natural photograph. And what we'd really like to do is to add a caption to describe this image, to say what's in the image. Again, purely from raw pixel inputs. We don't tell it this is an image with a baby girl holding a teddy bear. So the human might say it's a young girl asleep on the sofa cuddling a teddy bear. Our system says a close-up of a child holding a stuffed animal."}, {"start": 876.56, "end": 892.85, "text": "That was one example. Or a baby asleep next to a teddy bear. So this isn't in production yet. There are lots of false positives, but you can see what the general trajectory of very inherently general, flexible learning systems are."}, {"start": 896.5, "end": 925.26, "text": "Speech recognition. So remember, this is the same system that's also been deployed in production for Google speech recognition. So sometime around 2012, deep neural networks were used for acoustic modeling where they take an individual phoneme and they try to label it to a particular word that it's been trained on. It trains in about five days on all the speech that Google has in its recorded database, and it reduced the error rate by 30%. This was the greatest advance in research."}, {"start": 925.26, "end": 952.35, "text": "on speech recognition in 20 years. It's now launched on Android. And it's also used for 90 languages across Google Translate. So Tool AI is becoming a ubiquitous and general tool for us to be able to make the world a better place and improve our products and services. At Google it's now used for fraud detection, for spam detection, for image search, for handwriting recognition."}, {"start": 952.35, "end": 981.42, "text": "For speech detection, for landmark definition, and for street view images, as well as translation. It's now been replaced and is now in production in 60 systems across the company. So over the past sort of 20 or 30 years in computer science, we've been developing handcrafted, rule-based systems that essentially encode the knowledge and intelligence of clever human programmers into machine systems."}, {"start": 981.42, "end": 1012.27, "text": "And these rule-based systems have now been replaced by our general flexible learning-based systems. Now, of course, these kind of advances, particularly over the last two or three years in deep learning, have elicited lots of attention from various people, not least, of course, Elon Musk, who happens to be a good friend of ours and was one of the early investors in DeepMind. He said that it's worth reading Superintelligence, a book by Professor Nick Bostrom."}, {"start": 1012.27, "end": 1046.4, "text": "Because we need to be careful with AI because it's potentially more dangerous than nukes. So you might imagine I think this is somewhat hyperbolic. There have been various other texts on the subject suggesting that this might be our final invention. And also Bill Gates, I think we might be running out of battery here, Bill Gates and Stephen Hawking have both said these are things that we really need to be concerned about."}, {"start": 1046.4, "end": 1077.01, "text": "From our perspective, I think these are definitely things that we take seriously and that we think are important. But unfortunately, messages like these have been very much overblown, I think, in the media. And if you don't really pay any attention to what's happening in the field and the good that we are generating, then you might be forgiven for interpreting the situation a little bit more like this. Whether it's terminators looking to blow us up"}, {"start": 1077.01, "end": 1103.2, "text": "Female robots looking to seduce us or mad scientists looking to transcend. It's obviously clear that the sci-fi element to artificial intelligence has done a good job of attracting lots of eyeballs to web pages and generating lots of marketing and ad revenue through click-throughs. You won't be surprised to hear that AI has arrived, but we would argue that"}, {"start": 1103.2, "end": 1123.6, "text": "We mustn't jump to conclusions, and it's a shame that we sort of adapt so quickly in some respects and don't acknowledge the magic of the advances of the systems that we're developing. Peter Norvig, who is one of the founding fathers of artificial intelligence,"}, {"start": 1123.6, "end": 1141.49, "text": "Recently, he made a really interesting statement about this. He said that the narrative has changed. It's switched from, isn't it terrible that artificial intelligence is a failure, to isn't it terrible that artificial intelligence is a success. In some respects, there's a great truth in that."}, {"start": 1142.7, "end": 1166.82, "text": "As Eric said, we've put a lot of energy into organizing the community of ethics and safety researchers. We co-convened the conference in Puerto Rico that Eric talked about. And we've supported the Future of Life's open letter, which I think is really important. I mean, if you think about it, this is all of the researchers."}, {"start": 1166.82, "end": 1184.72, "text": "from industry and from academia that are involved in developing artificial general intelligence have signed an open letter saying how important it is that we respect ethical and safety considerations and establishing a research program and in many labs around the world this has now been taken on."}, {"start": 1184.72, "end": 1206.54, "text": "To look at things like verification. So can I be sure that the system that I've built is going to do the thing that I've designed it to do? I mean, this is a sort of response to the fear that these self-learning, self-improving systems have this capacity to edit their own code somehow or recursively improve in some uncontrollable way."}, {"start": 1206.54, "end": 1219.14, "text": "Validity, so can I convince myself formally using mathematical methods of the state and structure of these systems and that they won't demonstrate some kind of emergent properties?"}, {"start": 1219.14, "end": 1246.86, "text": "And then security, of course, can I ensure that they're robust and stable, that they can't escape, and that they're not subject to external influence from the rest of the environment? And of course, control. Even if we knew what we wanted, can we ensure that we have a mechanism for communicating with these systems that allows us to direct them in a stable manner? This, of course, kicked off Elon's donation to the research community of $10 million."}, {"start": 1246.86, "end": 1259.65, "text": "We've also been supporting the Cambridge Center for the Study of Existential Risk as well as the Oxford Future Humanity Institute. We've established our own ethics and safety board as part of the acquisition. This, I believe, is..."}, {"start": 1259.65, "end": 1289.01, "text": "Unprecedented in corporate history, it was a condition of our acquisition that we were able to establish a contractually binding, legally protected, perpetual representation of us as founders on a board with the most senior executives and decision makers at Google so that any of the technology that we develop before it goes into production is overseen by this board to ensure that it both respects the highest ethical standards but also the safety standards which we"}, {"start": 1289.01, "end": 1302.66, "text": "As engineers and developers of the technology can most readily see. And of this board, we also have an independently, a number of independent members on the board whose decision"}, {"start": 1302.66, "end": 1330.4, "text": "is binding where there is a disagreement or some kind of stalemate. And I think this is a really important step to, I think, a new model of governance within corporations that have a really significant role in the future of our civilization. And it's a step towards us inviting more accountability and more oversight and more collaboration with important social policymakers, other members of the industry."}, {"start": 1330.4, "end": 1352.56, "text": "As we have been engaging across the industry with our direct competitors, all of whom we know very well and we're in discussions with. So as you can probably tell, our take on this is that the existential risk framing has been a bit of a distraction from the core debate on ethics and safety. I think it's overshadowed the debate. The way that we perceive..."}, {"start": 1352.56, "end": 1379.82, "text": "Artificial intelligence is as a hugely powerful tool that we will ultimately control. And clearly we are building it to empower ourselves and to empower humanity, not of course to destroy us. And I think this technology has the potential to positively transform our world. And of course there are many, many urgent concerns that we should be interested in. There are many motivations that drive us directly. This is why we founded the company and why we've accelerated our progress by joining forces with Google."}, {"start": 1380.02, "end": 1402.03, "text": "So let me touch on a few of those for the remainder of my talk. I mean, the first, of course, is to say that we are potentially engaged in the acceleration of automation, which itself might induce what we fear to be long-term underemployment, structural underemployment over the next 20 or 30 years. This was really great work done by Carl Frey and Michael Osborne at Oxford."}, {"start": 1402.13, "end": 1430.58, "text": "They looked at the US Bureau of Labor Statistics, and by their estimate, over the next 20 years, around 47% of all jobs in the US are at threat from some kind of what they call computerization. So you can see this as digitization in the conventional definition, or you can see it as being turbocharged by artificial intelligence and machine learning, automating and, in some sense, replacing many of the technologies and tools that we're very familiar with."}, {"start": 1431.25, "end": 1453.9, "text": "Employment in the new world is something that's worth thinking about. In 1990, Chrysler, GM, and Ford, the big three, employed something on, had revenues of something on the order of 250 billion, which is roughly comparative to Google, Facebook, and Apple's revenues today. But remarkably,"}, {"start": 1453.97, "end": 1480.72, "text": "They had something like 1 million, 1.2 million employees. Today we have an order of magnitude less generating the same revenue annually. And of course a market cap that is 30 times greater than it was back in 1990. And I think this probably reflects the long-term vision of what companies like Google, Facebook, and Apple can potentially produce in the future."}, {"start": 1481.2, "end": 1510.77, "text": "But of course, these are really important issues, and we do need to discuss the question of structural underemployment. But there are also many, many other issues that I think we need to focus on and that are at stake in the world today. So it's probably worth taking a moment to reflect on some of these questions. 800 million people don't have any access to clean water, and I think what is most remarkable is that this trend is set to rise and not fall over the coming decade alone to 1.8 billion people."}, {"start": 1511.06, "end": 1531.76, "text": "Our rate of consumption in terms of our extraction of raw materials and the stress that we put on our ecological environment is utterly unsustainable. If we were to consume, if the rest of the world were to consume at the rate that we consume in the UK, then it's estimated that we would require something on the order of 3.1 planets worth of resources to sustain us."}, {"start": 1532.53, "end": 1557.3, "text": "800 million people are malnourished, but a third of food produced every year goes completely wasted. And so this raises long-term questions, I think, around what is the right mechanism for governance and decision making to ensure that we distribute the value of what we create on this planet in a way that is more consistent, perhaps, with the ethics that we might, in our personal lives, believe to be very important."}, {"start": 1558.86, "end": 1567.92, "text": "Concentrations of greenhouse gases are now higher than at any point in the last 800,000 years. I mean, this is an ecological crisis that I think"}, {"start": 1567.92, "end": 1593.04, "text": "It's very difficult for us to engage with directly, personally, and emotionally, but is one that we're not going to have any choice about engaging with in the 20, 30, and 40-year time frame. And it looks increasingly like the two-degree temperature rise is now baked in. It's very difficult to imagine how that would be irreversible. And that has very long-term structural impacts both on our industry, but also on our personal everyday lives."}, {"start": 1598.13, "end": 1619.06, "text": "The other thing to point out here is a graph that shows us in blue here of US pre-tax profits generated by the financial sector. In purple, the continuous steady line, we've got total GDP. And in green, we've got non-financial sector profit, which is roughly tracking it. And so what we see is that such an enormous amount of our..."}, {"start": 1619.06, "end": 1642.59, "text": "of profits in the US over the last 40 years have been generated by the financial sector, which has created enormous amount of volatility and lopsidedness that we might want to be cautious about as we move into the coming decades. The same situation I think is true of where value sits in our modern system."}, {"start": 1642.59, "end": 1667.7, "text": "In 2006, something on the order of 40% of GDP was held in the top five banks in the US. After the financial crisis, something like that increased to 56%. And so this raises questions of how we've responded to one of the largest financial crises in a generation. I'd suggest that our systems are both bigger now and more fragile than they were in 2008."}, {"start": 1668.62, "end": 1690.29, "text": "You'll be familiar with this meme, no doubt, from the last year of Piketty's work, but I think it's worth pointing out anyway. Incomes of the top 1% have risen by 300% in the last 30 years, and yet the incomes of the bottom two-thirds of our population in the world have fallen."}, {"start": 1690.29, "end": 1711.04, "text": "This is really a very cursory, high-level feel for some of the things that I think are at stake in the world today and some of the motivations and drivers for why we think developing general-purpose, flexible learning machines is something that we should be pushing for very, very aggressively. Potentially, they may help us to tackle some of our most..."}, {"start": 1711.04, "end": 1741.2, "text": "Intractable and stuck social problems. Everything we've ever created in our civilization has been a product of us as intelligent human beings creating more and creating beyond us. And AGI will massively amplify our ability to shape the world around us. Another framing of our mission, I might suggest, would be to create artificial intellectual horsepower, a cheap and abundant resource that we could use to help us solve some of our toughest global problems. Thank you."}, {"start": 1749.39, "end": 1784.94, "text": "So I guess I'm taking questions. So I appreciate you sharing a set of global problems of inequality, climate change, et cetera. What are you at Google and you as this broader community doing to bake in structures and incentives so that, because what we've found often in the past is new technology just exacerbates those things rather than reverses them."}, {"start": 1784.94, "end": 1814.99, "text": "So what are you baking in so that the new technology reverses rather than exacerbates? Right. So you could think about any data set or any information that we have in the world as high dimensional and time series based. So everything that we use to make our decisions as intelligent human beings seems to have this structure that it's very high dimensional and it goes back a very, very long way."}, {"start": 1814.99, "end": 1838.32, "text": "So if we can learn some of that structure so that we're able to make predictions from that structure and also to generate novel instantiations of how that structure might roll out, then you can imagine this to be an inherently very general tool that we can deploy at all sorts of problems. So for example, it may allow us to make breakthroughs in some of the..."}, {"start": 1838.32, "end": 1865.62, "text": "Really tough medical diagnoses problems that we have in cancer, in Alzheimer's. It may allow us to better simulate how we might make progress in material sciences. Now these are in themselves general tools which can go on to have massive structural effects in terms of the way that we organize our society and the way that we distribute our resources. So I think the way that we think about it is that"}, {"start": 1865.87, "end": 1881.52, "text": "We're developing tools that are available for public consumption. And we take a very, I guess, googly approach to the way that we share in the benefits of those. We won't be licensing our technology. We won't be"}, {"start": 1881.52, "end": 1913.14, "text": "Keeping it as proprietary. Everything that we produce, we publish, as you've seen from all of our papers. There's 26 or so papers out there in the public domain for other people to build on. And when we do learn in new application domains, so we have a very big effort in healthcare, for example, at the moment, our commitment is to make that information and the insight that we extract from learning in those domains free and publicly available."}, {"start": 1913.14, "end": 1941.1, "text": "Both Andy and Eric, what a surprise. I was at a seminar less than a week ago with a bunch of AI researchers, and the message I got from it was that the enthusiasm around the techniques that you all are using, particularly deep learning, convolutional networks and reinforcement learning, is a little bit over the top."}, {"start": 1941.1, "end": 1969.94, "text": "And if I heard the critique right, it was that, okay, these things are winning the benchmarks right now. They're not that much better. They're not getting better that much more quickly. And they're still kind of brittle out there in the real world and therefore not that useful out there in the real world. What's your response to that? Can you think of another technology which has replaced 60 production systems at one of the biggest technology companies in the world?"}, {"start": 1969.94, "end": 1999.15, "text": "I mean, thousands and thousands of people have spent decades handcrafting rule-based features to do completely different tasks, whether it's fraud detection, which is a very, very sparse signal. The fraudulent event happens one in 100,000, or whether it's image recognition, where the signal is very dense and compact because we're learning from a little window of pixels in an image. These are completely different domains, and the same single tool"}, {"start": 1999.15, "end": 2020.78, "text": "Solving all of those problems. Now, to me, that justifies attention. I mean, I never like hype, but I think it's remarkable progress and it justifies attention. The other thing I would say is, they're classification problems, right? So we've basically solved many of the large classification problems in the world. The next step..."}, {"start": 2020.78, "end": 2049.54, "text": "is being able to do classification well enough that we can learn much more complex environments. So environments where there isn't that kind of local structure and we can sample from that to generate predictions for what's going to happen in the future. Just that alone, once we have good generation, so not just saying this historical set of data looks like a cat or looks like the word dog or whatever, but actually to say I'm now going to imagine"}, {"start": 2049.54, "end": 2077.58, "text": "A new image that has never been seen before. So for example, some of our very most recent work just published was a paper called Draw, which learns over all of the numbers in the Street View image database. So these are raw images from the open world Street View. And what we ask it to do is, having learnt the structure of any random number, is to generate a five for us that's not in the data set."}, {"start": 2077.58, "end": 2108.19, "text": "A completely novel image of a five. And we do the same thing with toy cars and other small data sets. Now imagine when we can do that for a generic product. We can say generate for us a completely novel product in the space of fashion or something like this. So I think that's the next thing to come and that's what we've got all of our energy focused on. Just kind of a natural follow up to that."}, {"start": 2108.19, "end": 2136.0, "text": "One of the things that was particularly evocative in your presentation was being all those Atari video games. But if I remember correctly, there's some games where the humans are still significantly better and the machine hasn't been able to advance as rapidly. Can you give us some sort of sense of the kinds of problems that you think that your approach is particularly good at solving quickly, the ones that are still beyond yours or other AI approaches? And in particular, what does that imply for the kinds of things where machines will be"}, {"start": 2136.0, "end": 2155.41, "text": "So the kinds of problems that we're not very good at are those problems which require conceptual knowledge. So deep learning is a tool for solving perceptual recognition problems."}, {"start": 2155.41, "end": 2184.98, "text": "Whether it's auditory signals or visual information, we're good at solving problems where there is lots and lots of labeled training data of visual perception. What humans are particularly good at is aggregating all of that visual perception over your lifetime to build the idea of, say, a chair. So in some sense, I could use this as a chair. It looks nothing like a chair, but I have some conceptual representation in my mind of what it means to be a chair."}, {"start": 2184.98, "end": 2203.18, "text": "What's its function? What's its use? What's my historical record of interactions with chairs and chair-like things? And so anything that requires very abstract or sort of more abstract thought and reasoning and judgment is definitely going to be a safe domain, at least for the next decade."}, {"start": 2203.25, "end": 2232.85, "text": "But there are other issues. So we're not very good at what's called the temporal credit assignment problem. So we're very good where there's a local structure in a space. So the word dog, for example, I can learn from tens of thousands of examples of speaking the word dog. But where there is a very sparse and disparate signal that appears very occasionally in the training space, humans are still going to be very good at picking out that single variable"}, {"start": 2232.85, "end": 2262.51, "text": "Very revealing, salient insight. So for example, if you wanted to be an intelligent recommendation engine, you could look through the history of all of my interactions on the various products across Google, and I bet you you could pull out the one thing that was really going to blow my mind from a new domain that I was going to find really stimulating and interesting. And that's because the signal is very sparse. So if we could solve that problem, then I think we're encroaching on a whole new range of expertise and a whole new set of jobs."}, {"start": 2263.86, "end": 2292.22, "text": "Sure. You made a very interesting claim about having general purpose AI for which we in air quotes will still get control. You're also viewing the system with a high level of creativity and also at some point the ability to take its own tasks. How do you design the medical control mechanism which doesn't simply solve the subordinate problem of circumventing the control system? You raise a separate question. How do you do that? And a second question of..."}, {"start": 2292.22, "end": 2297.54, "text": "Is that a good thing or a bad thing?"}, {"start": 2297.54, "end": 2326.93, "text": "Well, I think in many domains, and that's why I sort of frame the context of what's at stake in the world today. In many domains, we need tools that will help us to learn in these novel environments much more efficiently. So we need creativity. We need artificial intellectual horsepower to help us to progress in some of these really tough domains. The question is, how do we restrain and control systems like this? And that, I think, has to do with secure virtual boxing."}, {"start": 2326.93, "end": 2345.65, "text": "It has to do with virtual brain analytics. So how do we look inside the network? Where in the network sits the representations for various ideas? So for example, one thing that we've been trying very hard on is to identify where the representation for learning the tunneling exploit sits in the brain."}, {"start": 2345.65, "end": 2362.22, "text": "And so we're developing tools that allow us as humans to both visualize that representation but also allow us to control the emergence of new novel representations that we want to ensure don't sort of pop up in places that they shouldn't be."}]}