这个专访是在阿法狗1:0李世石之后,第二盘开局之前: http://www.theverge.com/2016/3/10/11192774/demis-hassabis-interview-alphago-google-deepmind-ai 全部专访比较长,节选几段重要的: Sam Byford: So for someone who doesn’t know a lot about AI or Go, how would you characterize the cultural resonance of what happened yesterday? Demis Hassabis: There are several things I’d say about that. Go has always been the pinnacle of perfect information games. It’s way more complicated than chess in terms of possibility, so it’s always been a bit of a holy grail or grand challenge for AI research, especially since Deep Blue. And you know, we hadn’t got that far with it, even though there’d been a lot of efforts. Monte Carlo tree search was a big innovation ten years ago, but I think what we’ve done with AlphaGo is introduce with the neural networks this aspect of intuition, if you want to call it that, and that’s really the thing that separates out top Go players: their intuition. I was quite surprised that even on the live commentary Michael Redmond was having difficulty counting out the game, and he’s a 9-dan pro! And that just shows you how hard it is to write a valuation function for Go. A big reason for holding these matches in the first place was to evaluate AlphaGo’s capabilities, win or lose. What did you learn from last night? Well, I guess we learned that we’re further along the line than — well, not than we expected, but as far as we’d hoped, let’s say. We were telling people that we thought the match was 50-50. I think that’s still probably right; anything could still happen from here and I know Lee’s going to come back with a different strategy today. So I think it’s going to be really interesting to find out. Just talking about the significance for AI, to finish your first question, the other big thing you’ve heard me talk about is the difference between this and Deep Blue. So Deep Blue is a hand-crafted program where the programmers distilled the information from chess grandmasters into specific rules and heuristics, whereas we’ve imbued AlphaGo with the ability to learn and then it’s learnt it through practice and study, which is much more human-like. If the series continues this way with AlphaGo winning, what’s next — is there potential for another AI-vs-game showdown in the future? I think for perfect information games, Go is the pinnacle. Certainly there are still other top Go players to play. There are other games — no-limit poker is very difficult, multiplayer has its challenges because it’s an imperfect information game. And then there are obviously all sorts of video games that humans play way better than computers, like StarCraft is another big game in Korea as well. Strategy games require a high level of strategic capability in an imperfect information world — "partially observed," it’s called. The thing about Go is obviously you can see everything on the board, so that makes it slightly easier for computers. The main future uses of AI that you’ve brought up this week have been healthcare, smartphone assistants, and robotics. Let’s unpack some of those. To bring up healthcare, IBM with Watson has done some things with cancer diagnosis for example — what can DeepMind bring to the table? Well, it’s early days in that. We announced a partnership with the NHS a couple of weeks ago but that was really just to start building a platform that machine learning can be used in. I think Watson’s very different than what we do, from what I understand of it — it’s more like an expert system, so it’s a very different style of AI. I think the sort of things you’ll see this kind of AI do is medical diagnosis of images and then maybe longitudinal tracking of vital signs or quantified self over time, and helping people have healthier lifestyles. I think that’ll be quite suitable for reinforcement learning. So let’s move onto smartphone assistants. I saw you put up a slide from Her in your presentation on the opening day — is that really the endgame here? No, I mean Her is just an easy popular mainstream view of what that sort of thing is. I just think we would like these smartphone assistant things to actually be smart and contextual and have a deeper understanding of what you’re trying to do. At the moment most of these systems are extremely brittle — once you go off the templates that have been pre-programmed then they’re pretty useless. So it’s about making that actually adaptable and flexible and more robust. What’s the breakthrough that’s needed to improve these? Why couldn’t we work on it tomorrow? Well, we can — I just think you need a different approach. Again, it’s this dichotomy between pre-programmed and learnt. At the moment pretty much all smartphone assistants are special-cased and pre-programmed and that means they’re brittle because they can only do the things they were pre-programmed for. And the real world’s very messy and complicated and users do all sorts of unpredictable things that you can’t know ahead of time. Our belief at DeepMind, certainly this was the founding principle, is that the only way to do intelligence is to do learning from the ground up and be general. AlphaGo got off the ground by being taught a lot of game patterns — how is that applicable to smartphones where the input is so much more varied? Yeah, so there’s tons of data on that, you could learn from that. Actually, the AlphaGo algorithm, this is something we’re going to try in the next few months — we think we could get rid of the supervised learning starting point and just do it completely from self-play, literally starting from nothing. It’d take longer, because the trial and error when you’re playing randomly would take longer to train, maybe a few months. But we think it’s possible to ground it all the way to pure learning. Is that possible because of where the algorithm has reached now? No, no, we could have done that before. It wouldn’t have made the program stronger, it just would have been pure learning. so there would’ve been no supervised part. We think this algorithm can work without any supervision. The Atari games that we did last year, playing from the pixels — that didn’t bootstrap from any human knowledge, that started literally from doing random things on screen. Is it easier for that because the fail states are more obvious, and so on? It’s easier for that because the scores are more regular. In Go, you really only get one score, whether you’ve won or lost at the end of the game. It’s called the credit assignment problem; the problem is you’ve made a hundred actions or moves in Go, and you don’t know exactly which ones were responsible for winning or losing, so the signal’s quite weak. Whereas in most Atari games most of the things you’re doing give you some score, so you’ve got more breadcrumbs to follow. Could you give a timeframe for when some of these things might start making a noticeable difference to the phones that people use? I think in the next two to three years you’ll start seeing it. I mean, it’ll be quite subtle to begin with, certain aspects will just work better. Maybe looking four to five, five-plus years away you’ll start seeing a big step change in capabilities. How important was Google’s support to AlphaGo — could you have done it without them? It was very important. AlphaGo doesn’t actually use that much hardware in play, but we needed a lot of hardware to train it and do all the different versions and have them play each other in tournaments on the cloud. That takes quite a lot of hardware to do efficiently, so we couldn’t have done it in this time frame without those resources. So what are your far-off expectations for how humans, robots, and AIs will interact in the future? Obviously people’s heads go to pretty wild sci-fi places. I don’t think much about robotics myself personally. What I’m really excited to use this kind of AI for is science, and advancing that faster. I’d like to see AI-assisted science where you have effectively AI research assistants that do a lot of the drudgery work and surface interesting articles, find structure in vast amounts of data, and then surface that to the human experts and scientists who can make quicker breakthroughs. I was giving a talk at CERN a few months ago; obviously they create more data than pretty much anyone on the planet, and for all we know there could be new particles sitting on their massive hard drives somewhere and no-one’s got around to analyzing that because there’s just so much data. So I think it’d be cool if one day an AI was involved in finding a new particle.