0:00:01.100,0:00:03.240 - Okay, so, good morning everyone. 0:00:03.240,0:00:04.913 I'll just get started. 0:00:04.913,0:00:08.840 My name is Shailesh and I give these talks 0:00:08.927,0:00:11.946 almost every year so this is a very deja-vu feeling for me. 0:00:11.994,0:00:13.330 The only thing different this time 0:00:13.330,0:00:15.732 is the stage is slightly thinner. 0:00:15.732,0:00:21.880 But great crowd, great list of talks so far. 0:00:21.880,0:00:25.870 So, Daniel called me a couple of weeks ago and said 0:00:25.870,0:00:27.703 "Why don't you give a keynote again?" 0:00:27.703,0:00:30.690 And I said, "You know, I'm running out of things to say now." 0:00:30.690,0:00:33.200 I've given four talks at different forums 0:00:33.200,0:00:36.880 with The Fifth Elephant and I wasn't sure 0:00:36.880,0:00:38.030 what I'm gonna talk about 0:00:38.030,0:00:41.320 So, then, one of these days I was talking 0:00:41.320,0:00:43.280 to one of my non-geek friends 0:00:43.280,0:00:46.225 and he was very excited about what I do 0:00:46.225,0:00:47.460 so he said, 'What do you do?' 0:00:47.460,0:00:49.330 and I, you know, it was on the phone 0:00:49.330,0:00:53.033 and I started talking to him about this, that, and the other. 0:00:53.033,0:00:55.200 And for about 45 minutes I was rambling 0:00:55.200,0:00:57.080 and this guy was very quiet. 0:00:57.080,0:01:01.887 I didn't realize he wasn't a techie 0:01:01.887,0:01:06.760 and I was going on and on and after 45 minutes I stopped 0:01:06.760,0:01:09.177 and said, "Are you still there? Are you listening?" 0:01:09.177,0:01:11.120 And he said, "Yeah, I'm listening. 0:01:12.020,0:01:15.781 "Can you tell me what do you do again?"[br](audience laughs) 0:01:15.781,0:01:20.280 And then I realized, how do I summarize this in two words? 0:01:20.280,0:01:24.194 So then I told him, "Hey, I'm building thinking machines." 0:01:24.194,0:01:26.660 And that's when he said, "Why didn't you say that before? 0:01:26.660,0:01:28.420 "It was so easy to say that, right?" 0:01:28.420,0:01:30.130 So that's how the title came by 0:01:30.130,0:01:33.520 and obviously we're not building thinking machines 0:01:33.520,0:01:36.590 but what I'm gonna talk about is towards thinking machines, right? 0:01:36.590,0:01:38.946 So, we have a long way to go. 0:01:38.946,0:01:42.326 So I added the word "towards" later. 0:01:42.326,0:01:45.820 So what I'm gonna talk about is all over the place. 0:01:45.820,0:01:48.705 I'm gonna talk about philosophy, science fiction. 0:01:48.705,0:01:51.436 I'll talk about algorithms 0:01:51.436,0:01:54.660 and I'm gonna talk about deep learning 0:01:54.660,0:01:57.885 and how to think about things beyond deep learning. 0:01:57.885,0:02:02.420 And let me give you a perspective and then we'll start. 0:02:02.420,0:02:04.581 So I'll take questions at the end. 0:02:04.581,0:02:05.660 It's not working. 0:02:09.479,0:02:11.020 It's not working, this. 0:02:39.380,0:02:40.842 That's fine. 0:02:40.842,0:02:45.330 All right, so, I ended my last year's talk on this quotation 0:02:45.330,0:02:47.850 So I thought I'll start on this quotation this time. 0:02:48.600,0:02:50.580 So I like this quotation because it puts a lot 0:02:50.580,0:02:53.719 of things into perspective of what we're doing, 0:02:53.719,0:02:56.061 how our civilization got here 0:02:56.061,0:02:57.490 and where we are headed. 0:02:58.230,0:03:02.886 So it says, "Our technology, our machines, is part of our humanity. 0:03:02.886,0:03:05.920 "We created them to extend ourselves 0:03:05.920,0:03:08.130 "and that is what is unique about human beings!" 0:03:08.130,0:03:11.520 And if you look at chairs, and dogs, and animals, and cats 0:03:11.520,0:03:15.134 they don't create machines to extend themselves. 0:03:15.134,0:03:17.330 They just have instincts and they follow their instincts. 0:03:17.330,0:03:20.020 Right, that's very unique about human civilization. 0:03:20.020,0:03:24.620 We've created Taj Mahal, and space flights, and internet. 0:03:24.620,0:03:26.950 So we've come a very long way. 0:03:26.950,0:03:29.120 So if you think about the tools, right? 0:03:29.120,0:03:33.247 The cavemen had tools and now we have 0:03:33.247,0:03:37.890 a completely robotic assembly line with no humans 0:03:37.890,0:03:40.220 and you could turn the lights off and nothing will happen 0:03:40.220,0:03:41.996 the cars will get produced, right? 0:03:41.996,0:03:44.210 If you look at our transportation 0:03:44.980,0:03:50.080 we have gone from just on-road, bullock carts, 0:03:50.080,0:03:54.180 to massive amounts of transportation that we can do now. 0:03:56.480,0:03:58.980 If you look at our ability to look further 0:03:58.980,0:04:02.170 into space, again... 0:04:06.480,0:04:08.680 Since Galileo, we have made a lot of progress. 0:04:08.680,0:04:12.210 Recently we saw the news of Pluto flyby 0:04:12.210,0:04:14.852 so now we're able to send satellites into space. 0:04:14.852,0:04:17.716 If you look at the first computer we built 0:04:17.716,0:04:19.300 and where we are today, right? 0:04:19.300,0:04:22.210 We have a huge data center, and really 0:04:22.210,0:04:24.160 if you look at the whole thing in perspective 0:04:24.160,0:04:26.498 we have made an enormous amount of progress 0:04:26.498,0:04:29.867 in the last so many centuries, right? 0:04:29.867,0:04:32.580 So if you look just at the technical part 0:04:32.580,0:04:35.300 the IT kind of intelligent machines 0:04:35.300,0:04:36.670 we're not talking about mixies 0:04:36.670,0:04:40.460 and other things, just look at what AI 0:04:40.460,0:04:43.622 and deep learning and all this stuff has produced. 0:04:43.622,0:04:45.900 Today's machines can play chess. 0:04:45.900,0:04:47.750 And there's no human on the planet 0:04:47.750,0:04:50.412 who can play chess better than the machine. 0:04:50.412,0:04:54.708 I want to take a pause and think about where we are. 0:04:54.708,0:04:56.741 There's no human on the planet 0:04:56.741,0:04:58.880 who can play chess better than a machine. 0:05:00.240,0:05:02.600 There's no human on the planet 0:05:02.600,0:05:05.020 who can play Jeopardy better than a machine. 0:05:06.960,0:05:10.970 And recently, Google came out with automatic cars 0:05:10.970,0:05:13.840 so the machines can drive cars and record show 0:05:13.840,0:05:18.896 that these cars are better than humans under ideal conditions 0:05:18.896,0:05:21.440 And they have much less accident rates 0:05:21.440,0:05:25.030 and all the accidents happened because of other humans drivers. 0:05:25.030,0:05:26.350 They're not because of cars. 0:05:27.910,0:05:29.998 And recently you also saw 0:05:29.998,0:05:32.730 how machines are able to create pictures, right? 0:05:32.730,0:05:34.420 So this is one of the things 0:05:34.420,0:05:37.753 that we saw what deep learning is internally doing. 0:05:37.753,0:05:39.300 And now think about all this. 0:05:39.300,0:05:44.060 Just think about where machines have gone today. 0:05:44.060,0:05:45.860 How many things they can do 0:05:45.860,0:05:48.259 which are way beyond our imagination 0:05:48.259,0:05:49.760 that machines could have done. 0:05:51.140,0:05:55.040 So obviously there's a lot they've done. 0:05:55.040,0:05:57.181 But can they do the following? 0:05:57.181,0:05:59.244 We would want to stress the limits 0:05:59.244,0:06:02.405 so one of the holy grails of AI 0:06:02.405,0:06:06.032 is to have a machine have a conversation with a human being. 0:06:06.032,0:06:07.858 We all know the Turing test 0:06:07.858,0:06:11.109 and the repercussions of this will be huge. 0:06:11.109,0:06:14.468 If you think about how we talk to the internet today 0:06:14.468,0:06:18.970 we carefully craft three-word, four-word queries, right? 0:06:18.970,0:06:21.950 And you know, we allow the internet to make mistakes. 0:06:21.950,0:06:25.110 We craft the queries again, we take the suggestions or not. 0:06:25.110,0:06:28.636 We talk to the internet like we're talking to a three-year-old. 0:06:28.636,0:06:33.200 Now in the day and age needs of massive data computers, NLP 0:06:33.200,0:06:37.210 and all this deep-learning stuff, imagine what a shameful thing it is 0:06:37.210,0:06:40.091 to talk to a computer like a 3-year-old. 0:06:40.091,0:06:43.990 So it's got the capacity of thousands of people 0:06:43.990,0:06:46.395 but it can't understand language. 0:06:46.395,0:06:48.200 So we need to change that. 0:06:48.200,0:06:50.650 Now imagine beyond keywords what can happen. 0:06:50.650,0:06:52.247 We can do question-answering 0:06:52.247,0:06:54.630 but how do we do question-answering today? 0:06:54.630,0:06:57.640 We have created Yahoo Answers, we have created Quora 0:06:57.640,0:07:01.140 and people who type questions, we do a match. 0:07:01.140,0:07:04.875 Between the questions and the answers 0:07:04.875,0:07:07.600 and then we again do retrieval. 0:07:07.600,0:07:09.640 We're still not answering questions. 0:07:10.520,0:07:11.980 Now think about conversations. 0:07:11.980,0:07:14.160 Conversation is an even more complex thing. 0:07:14.160,0:07:17.528 If it works out, what are the repercussions? 0:07:17.528,0:07:20.160 I don't want to study physics from my physics teacher. 0:07:20.160,0:07:22.840 I want to study it from Einstein or Feynman. 0:07:23.710,0:07:28.260 We already know all the language and the knowledge of these people. 0:07:28.260,0:07:32.020 Can we not have a persona of a person, Feynman or Einstein 0:07:32.020,0:07:34.780 and have a conversation with that person? 0:07:35.540,0:07:38.520 So, just imagine the future of what will happen 0:07:38.520,0:07:41.550 if we are able to just have conversations with the machines. 0:07:41.550,0:07:43.590 So, there's a long way to go between 0:07:43.590,0:07:46.679 keyword search and conversations. 0:07:46.679,0:07:48.987 Can we discover a cure for cancer? 0:07:48.987,0:07:51.000 There are a lot of diseases out there. 0:07:51.000,0:07:52.820 Now, obviously there is a lot 0:07:52.820,0:07:54.840 of research pharma companies are doing. 0:07:54.840,0:07:56.940 There's a lot of new initiatives in how 0:07:56.940,0:08:00.160 to use the high-end machine learning in pharma research. 0:08:00.990,0:08:06.140 But my contention is that I believe that the cure for a lot 0:08:06.140,0:08:08.647 of diseases is already out there. 0:08:08.647,0:08:11.470 In all the medical literature, if somebody 0:08:11.470,0:08:14.450 could actually read them, hold that knowledge 0:08:14.450,0:08:19.133 in the brain, in RAM, and do interconnections 0:08:19.133,0:08:21.739 we should be able to find a lot of things. 0:08:21.739,0:08:22.860 But what is the problem? 0:08:22.860,0:08:25.920 A single human expert, even in one field 0:08:25.920,0:08:28.730 cannot keep up with that quest of knowledge, right? 0:08:28.730,0:08:31.720 We'll forget some things, we won't read certain papers. 0:08:31.720,0:08:34.059 And therefore, it's the other problem. 0:08:34.059,0:08:37.190 We have too much knowledge and our individual brains 0:08:37.190,0:08:41.378 are not capable of forming those connections in the... 0:08:41.378,0:08:44.198 Because we can't even read that many documents, right? 0:08:44.198,0:08:46.040 But if a machine could do it 0:08:46.040,0:08:48.050 the way NLP has progressed 0:08:48.050,0:08:50.690 can we not find cures or new medicine? 0:08:52.600,0:08:54.940 Can I crack the next IIT Entrance Exam? 0:08:58.140,0:09:01.560 You're laughing today, but you never know. 0:09:01.560,0:09:03.776 Five years from now, what will happen? 0:09:03.776,0:09:08.740 And we should hope that if Watson is a test of intelligence 0:09:08.740,0:09:11.550 if Igloo is a test of intelligence 0:09:11.550,0:09:14.191 could this not be a test of intelligence? 0:09:14.191,0:09:17.560 The ability of AI system to be able 0:09:17.560,0:09:20.970 to actually solve an IIT paper and get a rank 1. 0:09:23.610,0:09:26.987 Can I search all the video scenes 0:09:26.987,0:09:28.610 which only have a goal shot 0:09:28.610,0:09:31.600 in the football videos and nothing else. 0:09:31.600,0:09:34.219 I don't want to watch the rest of it. 0:09:34.219,0:09:36.040 A lot of balls going here and there. 0:09:36.040,0:09:37.850 I just wanna see the goal shots. 0:09:37.850,0:09:39.965 Today I cannot do that. 0:09:39.965,0:09:42.140 Can my machines be intelligent enough 0:09:42.140,0:09:44.430 the vision part, that can actually find 0:09:44.430,0:09:46.330 this is a goal, this is a goal, this is a goal 0:09:46.330,0:09:48.563 the rest of it is something else. 0:09:48.563,0:09:50.920 So we can imagine the applications out there. 0:09:51.820,0:09:54.018 We were talking about sarcasm a lot 0:09:54.018,0:09:57.370 and we all understand sarcasm is a very hard thing to do. 0:09:57.370,0:10:01.542 And imagine if you could detect sarcasm, what else can you do? 0:10:01.542,0:10:03.440 You're writing an email to your boss 0:10:03.440,0:10:06.732 you're angry, you have written a sarcastic comment 0:10:06.732,0:10:09.260 and Gmail says, "Hey, are you sure about this?" 0:10:09.260,0:10:12.090 In the heat of the moment[br](audience laughs) 0:10:13.000,0:10:14.580 can I put it this way? 0:10:15.520,0:10:17.580 So, like, today we do attachments 0:10:17.580,0:10:19.820 can we detect sarcasm and things like that? 0:10:20.730,0:10:25.500 And to me the holy grail of AI is not really 0:10:25.500,0:10:28.265 all these big things, but a very simple thing. 0:10:28.265,0:10:31.080 Can a machine find a joke funny? 0:10:32.210,0:10:33.879 Now there are a lot of... 0:10:33.879,0:10:36.009 I don't know if you guys watch Star Trek 0:10:36.009,0:10:39.850 but Data, in 300 years, 400 years from now 0:10:39.850,0:10:40.990 is an android. 0:10:40.990,0:10:43.280 He is capable of all these other things. 0:10:43.280,0:10:46.840 He's a great supercomputer in a human form 0:10:46.840,0:10:48.700 but he's still struggling with humor. 0:10:48.700,0:10:50.620 That's how hard the problem is. 0:10:51.560,0:10:54.240 So obviously we have a long way to go. 0:10:54.240,0:10:56.380 We have come a long way and we have a long way to go. 0:10:56.380,0:10:59.290 So this talk is really about the way forward. 0:11:00.960,0:11:04.199 So, what do we imagine the future to be? 0:11:04.199,0:11:06.340 We want something like this. 0:11:07.230,0:11:09.350 Good and bad, hopefully good. 0:11:09.350,0:11:10.380 We want a Jarvis, right? 0:11:10.380,0:11:11.740 We all want a Jarvis 0:11:11.740,0:11:13.360 who'll takes care of the chores 0:11:13.360,0:11:16.870 and get rid of whatever 0:11:16.870,0:11:19.350 and we all want a Jarvis right? 0:11:19.350,0:11:22.531 So if you watch these movies again 0:11:22.531,0:11:24.556 after watching this talk 0:11:24.556,0:11:26.560 you'll have a very different perspective 0:11:26.560,0:11:29.708 on what we need to do to get here. 0:11:29.708,0:11:31.680 It's not gonna happen just because we're 0:11:31.680,0:11:35.040 gonna make more and more [br]Hollywood movies like this. 0:11:35.335,0:11:38.552 I mean, Asimov wrote [br]"I, Robot" in the 70s 0:11:38.705,0:11:40.531 and we're still not there. 0:11:41.397,0:11:44.098 It's not gonna happen because we keep [br]doing "data science" 0:11:45.297,0:11:47.840 And that's one of the reasons I wanted[br]to do this talk 0:11:47.840,0:11:49.533 'cause a lot of people keep 0:11:50.004,0:11:52.273 thinking "data science is the end [br]of the world" 0:11:52.273,0:11:54.024 but there's a lot more to data science 0:11:54.024,0:11:57.085 and I want to see how we can go beyond[br]data science 0:11:57.478,0:11:59.004 - and this is not data science. 0:11:59.014,0:12:00.558 This is artificial intelligence. 0:12:00.558,0:12:02.496 Right? So I want to draw the distinction 0:12:02.512,0:12:04.646 and say how we can move[br]beyond data science 0:12:05.058,0:12:06.328 - nothing wrong with it - 0:12:06.523,0:12:08.275 but it's, it's a done deal. 0:12:08.873,0:12:11.408 Right? We have software you can download, 0:12:11.408,0:12:13.166 you can put up whatever you want, 0:12:13.481,0:12:16.650 it's a done deal. Data science has been[br]packaged, already. 0:12:17.522,0:12:20.217 Right? If you look at Microsoft Azure, 0:12:20.217,0:12:22.276 or some of these other softwares, right? 0:12:22.301,0:12:23.712 It has already been packaged 0:12:24.053,0:12:25.963 All you have to do is download the right 0:12:25.963,0:12:28.361 software, put your data[br]in the right format, 0:12:28.361,0:12:32.875 and you're done. Right? So there's nothing[br]"great" about data science anymore. 0:12:33.840,0:12:36.289 Sorry about that, but, you know, 0:12:36.533,0:12:39.301 we need to jolt ourselves out of this[br]comfort zone, and say 0:12:39.301,0:12:40.941 "okay, we are all data scientists" 0:12:40.941,0:12:42.053 - that's not it, right? 0:12:42.102,0:12:44.557 How do we get here?[br]How will data science get here? 0:12:45.133,0:12:49.181 Alright. So, we'll get here by asking[br]a lot of deeper questions. 0:12:49.364,0:12:51.299 Right? Not the questions like 0:12:51.778,0:12:54.955 "Why is this customer[br]returning from Flipkart?", right? or 0:12:54.955,0:12:58.766 "Who's -- what is the next product to[br]recommend to somebody?", or 0:12:59.102,0:13:02.025 "Which movie you're going to ask?"[br]These are not the questions 0:13:02.457,0:13:04.639 that'll take us to the next stage. Right? 0:13:04.639,0:13:07.149 So the question that'll take us[br]to the next stage is 0:13:07.456,0:13:10.297 "what is learning?" Fundamentally, [br]philosophically. 0:13:10.540,0:13:13.867 "What is learning?" We see that we[br]are learning, children are learning, 0:13:13.867,0:13:16.244 everybody is going to school, [br]we all are learning. 0:13:16.244,0:13:18.356 We think that machine learning[br]is learning, 0:13:18.466,0:13:20.497 but what is learning really, right? 0:13:20.642,0:13:21.994 "What is understanding?" 0:13:22.444,0:13:26.539 What does that mean?[br]What does the word "mean" mean? 0:13:28.069,0:13:31.387 What is thinking? We keep saying 0:13:31.387,0:13:32.882 "Oh -- I'm thinking about this" 0:13:32.882,0:13:34.796 What are you doing when you're thinking? 0:13:34.967,0:13:37.945 So, today I'm going to show you an[br]equation of thinking. 0:13:37.977,0:13:40.842 Okay? So, it'll be fun -- I don't claim 0:13:40.842,0:13:42.764 this is - THE - equation of thinking, 0:13:42.764,0:13:45.072 but I'm trying to get to that plot point 0:13:45.072,0:13:46.928 where we start thinking about thinking, 0:13:47.865,0:13:49.172 and not just think. 0:13:50.296,0:13:54.483 "What is creativity?" Now, creativity is, 0:13:54.483,0:13:57.815 if you look at an artist, or a musician,[br]or even a scientist, 0:13:57.815,0:13:59.356 we create new inventions 0:13:59.356,0:14:01.417 out of the knowledge we have, 0:14:01.417,0:14:06.327 and innovation is a manifestation of[br]the knowledge in a certain form. 0:14:06.327,0:14:09.405 Right? A poet creates, [br]a musician creates -- 0:14:09.405,0:14:11.203 so what is creativity? 0:14:11.203,0:14:14.777 And the last question I have, here is 0:14:14.777,0:14:16.881 "What is consciousness?" Right? 0:14:16.881,0:14:20.210 So, ultimately, if you look at movies like[br]"I, Robot", 0:14:20.210,0:14:23.730 the word "I" from the robot is[br]not really about 0:14:23.730,0:14:27.110 the robot's great abilities at[br]mundane tasks, 0:14:27.110,0:14:29.092 but really it's about the "I" in it. 0:14:29.667,0:14:34.409 "I am a conscious being", and now what are[br]the consequences. 0:14:34.480,0:14:36.232 Right? So what is consciousness, and 0:14:36.249,0:14:39.188 can we have sentient machines at the[br]end of the day, right? 0:14:39.519,0:14:41.860 So, we won't go there today,[br]maybe we'll see 0:14:41.860,0:14:43.718 if we have time we'll watch a video, 0:14:44.247,0:14:46.949 but I'll try to cover the bottom three and 0:14:46.949,0:14:50.381 see if we can find something interesting. 0:14:50.381,0:14:53.827 So, learning. Learning is one of the[br]most basic things, 0:14:53.827,0:14:55.388 we all do learning all the time. 0:14:55.686,0:14:58.219 -- at least we all claim to[br]be learning all the time. 0:14:58.478,0:15:01.539 So, really, I'm going to use language and 0:15:01.539,0:15:03.662 not vision at first, but language as 0:15:03.662,0:15:07.122 the basis for all the examples. 0:15:07.729,0:15:11.053 So, learning really is many, many things: 0:15:11.473,0:15:13.710 the first thing we learn, so, you know, 0:15:13.710,0:15:17.447 the greatest example of a machine[br]learning system, or an A.I. system 0:15:17.660,0:15:18.775 is a human child. 0:15:19.488,0:15:21.136 And all you have to do is just observe 0:15:21.136,0:15:25.133 how a baby is growing up, how he's[br]picking language, how he's 0:15:25.133,0:15:29.333 picking walking, how he's picking [br]swimming, how he's picking tantrums, right? 0:15:29.733,0:15:34.193 And you learn so much about A.I. because[br]you're looking at the real A.I. 0:15:34.193,0:15:36.778 So what is learning? I want to use[br]that example 0:15:36.778,0:15:39.008 and see how we pick up language. 0:15:39.638,0:15:43.147 If I use the word -- if I start[br]-- imagine you're reading a novel, 0:15:43.433,0:15:45.763 or imagine words are coming at you one[br]at a time: 0:15:46.265,0:15:50.046 you see the word "united" - what do you[br]think the next word would be? 0:15:52.796,0:15:57.360 Right? "United States", [br]"United Something" , whatever. 0:15:57.360,0:16:00.644 then, [MIC CUTS], predicting. When we're[br]learning, 0:16:00.644,0:16:03.027 we are also simultaneously predicting. 0:16:03.027,0:16:05.917 And this is one of the flaws in current[br]machine learning: 0:16:05.917,0:16:10.896 that we keep thinking that learning is[br]separate, prediction is separate. 0:16:11.435,0:16:13.577 We'll learn first, then we'll score. 0:16:14.442,0:16:16.843 Right? But the human brain is not like that. 0:16:16.843,0:16:20.375 We don't learn for sixty years and then suddenly[br]we start behaving. 0:16:20.375,0:16:23.882 We're constantly learning and [br]we're constantly applying that learning, 0:16:23.882,0:16:26.432 and that is one of the fundamental[br]reasons why, 0:16:26.432,0:16:29.153 you know, I call the current model of[br]machine learning 0:16:29.153,0:16:33.271 like the [inaudible] which is never going[br]to become a data-flow architecture 0:16:33.271,0:16:34.806 ever, right? So that is one of the problems. 0:16:35.189,0:16:38.284 So imagine what we're doing now, we are[br]predicting what will come next.