WEBVTT 00:00.000 --> 00:02.960 The following is a conversation with Tomaso Poggio. 00:02.960 --> 00:06.200 He's a professor at MIT and is a director of the Center 00:06.200 --> 00:08.360 for Brains, Minds, and Machines. 00:08.360 --> 00:11.640 Cited over 100,000 times, his work 00:11.640 --> 00:14.560 has had a profound impact on our understanding 00:14.560 --> 00:17.680 of the nature of intelligence in both biological 00:17.680 --> 00:19.880 and artificial neural networks. 00:19.880 --> 00:23.840 He has been an advisor to many highly impactful researchers 00:23.840 --> 00:26.120 and entrepreneurs in AI, including 00:26.120 --> 00:28.000 Demisus Habbis of DeepMind, 00:28.000 --> 00:31.200 Amnon Shashwa of Mobileye, and Christoph Koch 00:31.200 --> 00:34.120 of the Allen Institute for Brain Science. 00:34.120 --> 00:36.400 This conversation is part of the MIT course 00:36.400 --> 00:38.120 on artificial general intelligence 00:38.120 --> 00:40.240 and the artificial intelligence podcast. 00:40.240 --> 00:42.760 If you enjoy it, subscribe on YouTube, iTunes, 00:42.760 --> 00:44.600 or simply connect with me on Twitter 00:44.600 --> 00:47.960 at Lex Freedman, spelled F R I D. 00:47.960 --> 00:52.480 And now, here's my conversation with Tomaso Poggio. 00:52.480 --> 00:54.520 You've mentioned that in your childhood, 00:54.520 --> 00:56.960 you've developed a fascination with physics, 00:56.960 --> 00:59.720 especially the theory of relativity, 00:59.720 --> 01:03.600 and that Einstein was also a childhood hero to you. 01:04.520 --> 01:09.040 What aspect of Einstein's genius, the nature of his genius, 01:09.040 --> 01:10.200 do you think was essential 01:10.200 --> 01:12.960 for discovering the theory of relativity? 01:12.960 --> 01:15.960 You know, Einstein was a hero to me, 01:15.960 --> 01:17.200 and I'm sure to many people, 01:17.200 --> 01:21.680 because he was able to make, of course, 01:21.680 --> 01:25.200 a major, major contribution to physics 01:25.200 --> 01:28.520 with simplifying a bit, 01:28.520 --> 01:33.520 just a gedanken experiment, a thought experiment. 01:35.200 --> 01:38.880 You know, imagining communication with lights 01:38.880 --> 01:43.240 between a stationary observer and somebody on a train. 01:43.240 --> 01:48.240 And I thought, you know, the fact that just 01:48.560 --> 01:52.720 with the force of his thought, of his thinking, of his mind, 01:52.720 --> 01:55.640 it could get to something so deep 01:55.640 --> 01:57.520 in terms of physical reality, 01:57.520 --> 02:01.320 how time depends on space and speed. 02:01.320 --> 02:04.120 It was something absolutely fascinating. 02:04.120 --> 02:06.720 It was the power of intelligence, 02:06.720 --> 02:08.440 the power of the mind. 02:08.440 --> 02:11.120 Do you think the ability to imagine, 02:11.120 --> 02:15.200 to visualize as he did, as a lot of great physicists do, 02:15.200 --> 02:18.640 do you think that's in all of us human beings, 02:18.640 --> 02:20.600 or is there something special 02:20.600 --> 02:22.880 to that one particular human being? 02:22.880 --> 02:27.160 I think, you know, all of us can learn 02:27.160 --> 02:32.160 and have, in principle, similar breakthroughs. 02:33.240 --> 02:37.200 There is lesson to be learned from Einstein. 02:37.200 --> 02:42.200 He was one of five PhD students at ETA, 02:42.600 --> 02:47.600 the Eidgenossische Technische Hochschule in Zurich, in physics. 02:47.600 --> 02:49.840 And he was the worst of the five. 02:49.840 --> 02:53.600 The only one who did not get an academic position 02:53.600 --> 02:57.040 when he graduated, when he finished his PhD, 02:57.040 --> 03:00.000 and he went to work, as everybody knows, 03:00.000 --> 03:01.720 for the patent office. 03:01.720 --> 03:05.000 So it's not so much that he worked for the patent office, 03:05.000 --> 03:07.880 but the fact that obviously he was smart, 03:07.880 --> 03:10.240 but he was not the top student, 03:10.240 --> 03:12.640 obviously he was the anti conformist. 03:12.640 --> 03:15.720 He was not thinking in the traditional way 03:15.720 --> 03:18.760 that probably teachers and the other students were doing. 03:18.760 --> 03:23.760 So there is a lot to be said about trying to do the opposite 03:25.960 --> 03:29.800 or something quite different from what other people are doing. 03:29.800 --> 03:31.840 That's certainly true for the stock market. 03:31.840 --> 03:34.800 Never buy if everybody's buying it. 03:35.800 --> 03:37.440 And also true for science. 03:37.440 --> 03:38.440 Yes. 03:38.440 --> 03:42.440 So you've also mentioned staying on the theme of physics 03:42.440 --> 03:46.440 that you were excited at a young age 03:46.440 --> 03:50.440 by the mysteries of the universe that physics could uncover. 03:50.440 --> 03:54.440 Such, as I saw mentioned, the possibility of time travel. 03:56.440 --> 03:59.440 So out of the box question I think I'll get to ask today, 03:59.440 --> 04:01.440 do you think time travel is possible? 04:02.440 --> 04:05.440 Well, it would be nice if it were possible right now. 04:05.440 --> 04:11.440 In science you never say no. 04:11.440 --> 04:14.440 But your understanding of the nature of time. 04:14.440 --> 04:15.440 Yeah. 04:15.440 --> 04:20.440 It's very likely that it's not possible to travel in time. 04:20.440 --> 04:24.440 We may be able to travel forward in time. 04:24.440 --> 04:28.440 If we can, for instance, freeze ourselves 04:28.440 --> 04:34.440 or go on some spacecraft traveling close to the speed of light, 04:34.440 --> 04:39.440 but in terms of actively traveling, for instance, back in time, 04:39.440 --> 04:43.440 I find probably very unlikely. 04:43.440 --> 04:49.440 So do you still hold the underlying dream of the engineering intelligence 04:49.440 --> 04:54.440 that will build systems that are able to do such huge leaps 04:54.440 --> 04:58.440 like discovering the kind of mechanism 04:58.440 --> 05:00.440 that would be required to travel through time? 05:00.440 --> 05:02.440 Do you still hold that dream? 05:02.440 --> 05:05.440 Or echoes of it from your childhood? 05:05.440 --> 05:06.440 Yeah. 05:06.440 --> 05:10.440 I don't think there are certain problems 05:10.440 --> 05:13.440 that probably cannot be solved, 05:13.440 --> 05:17.440 depending on what you believe about the physical reality. 05:17.440 --> 05:23.440 Maybe it's totally impossible to create energy from nothing 05:23.440 --> 05:26.440 or to travel back in time. 05:26.440 --> 05:35.440 But about making machines that can think as well as we do or better, 05:35.440 --> 05:39.440 or more likely, especially in the short and mid term, 05:39.440 --> 05:41.440 help us think better, 05:41.440 --> 05:45.440 which in a sense is happening already with the computers we have, 05:45.440 --> 05:47.440 and it will happen more and more. 05:47.440 --> 05:49.440 But that I certainly believe, 05:49.440 --> 05:53.440 and I don't see in principle why computers at some point 05:53.440 --> 05:59.440 could not become more intelligent than we are, 05:59.440 --> 06:03.440 although the word intelligence is a tricky one, 06:03.440 --> 06:07.440 and one who should discuss what I mean with that. 06:07.440 --> 06:12.440 Intelligence, consciousness, words like love, 06:12.440 --> 06:16.440 all these need to be disentangled. 06:16.440 --> 06:20.440 So you've mentioned also that you believe the problem of intelligence 06:20.440 --> 06:23.440 is the greatest problem in science, 06:23.440 --> 06:26.440 greater than the origin of life and the origin of the universe. 06:26.440 --> 06:29.440 You've also, in the talk, 06:29.440 --> 06:34.440 I've said that you're open to arguments against you. 06:34.440 --> 06:40.440 So what do you think is the most captivating aspect 06:40.440 --> 06:43.440 of this problem of understanding the nature of intelligence? 06:43.440 --> 06:46.440 Why does it captivate you as it does? 06:46.440 --> 06:54.440 Well, originally, I think one of the motivations that I had as a teenager, 06:54.440 --> 06:58.440 when I was infatuated with the theory of relativity, 06:58.440 --> 07:05.440 was really that I found that there was the problem of time and space 07:05.440 --> 07:07.440 and general relativity, 07:07.440 --> 07:12.440 but there were so many other problems of the same level of difficulty 07:12.440 --> 07:16.440 and importance that I could, even if I were Einstein, 07:16.440 --> 07:19.440 it was difficult to hope to solve all of them. 07:19.440 --> 07:26.440 So what about solving a problem whose solution allowed me to solve all the problems? 07:26.440 --> 07:32.440 And this was what if we could find the key to an intelligence 07:32.440 --> 07:36.440 ten times better or faster than Einstein? 07:36.440 --> 07:39.440 So that's sort of seeing artificial intelligence 07:39.440 --> 07:42.440 as a tool to expand our capabilities. 07:42.440 --> 07:47.440 But is there just an inherent curiosity in you 07:47.440 --> 07:53.440 and just understanding what it is in here that makes it all work? 07:53.440 --> 07:55.440 Yes, absolutely. You're right. 07:55.440 --> 08:00.440 So I started saying this was the motivation when I was a teenager, 08:00.440 --> 08:06.440 but soon after, I think the problem of human intelligence 08:06.440 --> 08:14.440 became a real focus of my science and my research, 08:14.440 --> 08:27.440 because I think for me the most interesting problem is really asking who we are. 08:27.440 --> 08:31.440 It is asking not only a question about science, 08:31.440 --> 08:37.440 but even about the very tool we are using to do science, which is our brain. 08:37.440 --> 08:39.440 How does our brain work? 08:39.440 --> 08:41.440 From where does it come from? 08:41.440 --> 08:43.440 What are its limitations? 08:43.440 --> 08:45.440 Can we make it better? 08:45.440 --> 08:49.440 And that in many ways is the ultimate question 08:49.440 --> 08:53.440 that underlies this whole effort of science. 08:53.440 --> 08:58.440 So you've made significant contributions in both the science of intelligence 08:58.440 --> 09:01.440 and the engineering of intelligence. 09:01.440 --> 09:04.440 In a hypothetical way, let me ask, 09:04.440 --> 09:08.440 how far do you think we can get in creating intelligence systems 09:08.440 --> 09:11.440 without understanding the biological, 09:11.440 --> 09:15.440 the understanding how the human brain creates intelligence? 09:15.440 --> 09:18.440 Put another way, do you think we can build a strong ass system 09:18.440 --> 09:24.440 without really getting at the core, understanding the functional nature of the brain? 09:24.440 --> 09:28.440 Well, this is a real difficult question. 09:28.440 --> 09:34.440 We did solve problems like flying 09:34.440 --> 09:43.440 without really using too much our knowledge about how birds fly. 09:43.440 --> 09:51.440 It was important, I guess, to know that you could have things heavier than air 09:51.440 --> 09:55.440 being able to fly like birds. 09:55.440 --> 10:00.440 But beyond that, probably we did not learn very much. 10:00.440 --> 10:08.440 The brothers right did learn a lot of observation about birds 10:08.440 --> 10:12.440 and designing their aircraft, 10:12.440 --> 10:17.440 but you can argue we did not use much of biology in that particular case. 10:17.440 --> 10:28.440 Now, in the case of intelligence, I think that it's a bit of a bet right now. 10:28.440 --> 10:36.440 If you ask, okay, we all agree we'll get at some point, maybe soon, 10:36.440 --> 10:42.440 maybe later, to a machine that is indistinguishable from my secretary 10:42.440 --> 10:47.440 in terms of what I can ask the machine to do. 10:47.440 --> 10:50.440 I think we'll get there and now the question is, 10:50.440 --> 10:56.440 you can ask people, do you think we'll get there without any knowledge about the human brain 10:56.440 --> 11:02.440 or the best way to get there is to understand better the human brain? 11:02.440 --> 11:08.440 This is, I think, an educated bet that different people with different backgrounds 11:08.440 --> 11:11.440 will decide in different ways. 11:11.440 --> 11:17.440 The recent history of the progress in AI in the last, I would say, five years 11:17.440 --> 11:26.440 or ten years has been that the main breakthroughs, the main recent breakthroughs, 11:26.440 --> 11:31.440 really start from neuroscience. 11:31.440 --> 11:35.440 I can mention reinforcement learning as one, 11:35.440 --> 11:41.440 is one of the algorithms at the core of AlphaGo, 11:41.440 --> 11:46.440 which is the system that beat the kind of an official world champion of Go, 11:46.440 --> 11:52.440 Lee Siddle, two, three years ago in Seoul. 11:52.440 --> 12:00.440 That's one, and that started really with the work of Pavlov in 1900, 12:00.440 --> 12:07.440 Marvin Miski in the 60s and many other neuroscientists later on. 12:07.440 --> 12:13.440 And deep learning started, which is the core again of AlphaGo 12:13.440 --> 12:19.440 and systems like autonomous driving systems for cars, 12:19.440 --> 12:25.440 like the systems that Mobileye, which is a company started by one of my ex, 12:25.440 --> 12:30.440 Okamnon Shashua, so that is the core of those things. 12:30.440 --> 12:35.440 And deep learning, really the initial ideas in terms of the architecture 12:35.440 --> 12:42.440 of these layered hierarchical networks started with work of Thorston Wiesel 12:42.440 --> 12:47.440 and David Hubel at Harvard up the river in the 60s. 12:47.440 --> 12:54.440 So recent history suggests that neuroscience played a big role in these breakthroughs. 12:54.440 --> 12:59.440 My personal bet is that there is a good chance they continue to play a big role, 12:59.440 --> 13:03.440 maybe not in all the future breakthroughs, but in some of them. 13:03.440 --> 13:05.440 At least in inspiration. 13:05.440 --> 13:07.440 At least in inspiration, absolutely, yes. 13:07.440 --> 13:12.440 So you studied both artificial and biological neural networks, 13:12.440 --> 13:19.440 you said these mechanisms that underlie deep learning and reinforcement learning, 13:19.440 --> 13:25.440 but there is nevertheless significant differences between biological and artificial neural networks 13:25.440 --> 13:27.440 as they stand now. 13:27.440 --> 13:32.440 So between the two, what do you find is the most interesting, mysterious, 13:32.440 --> 13:37.440 maybe even beautiful difference as it currently stands in our understanding? 13:37.440 --> 13:44.440 I must confess that until recently I found that the artificial networks 13:44.440 --> 13:49.440 were too simplistic relative to real neural networks. 13:49.440 --> 13:54.440 But, you know, recently I've been started to think that, yes, 13:54.440 --> 13:59.440 there are very big simplification of what you find in the brain. 13:59.440 --> 14:07.440 But on the other hand, they are much closer in terms of the architecture to the brain 14:07.440 --> 14:13.440 than other models that we had, that computer science used as model of thinking, 14:13.440 --> 14:19.440 or mathematical logics, you know, LISP, Prologue, and those kind of things. 14:19.440 --> 14:23.440 So in comparison to those, they're much closer to the brain. 14:23.440 --> 14:28.440 You have networks of neurons, which is what the brain is about. 14:28.440 --> 14:35.440 The artificial neurons in the models are, as I said, caricature of the biological neurons, 14:35.440 --> 14:39.440 but they're still neurons, single units communicating with other units, 14:39.440 --> 14:50.440 something that is absent in the traditional computer type models of mathematics, reasoning, and so on. 14:50.440 --> 14:56.440 So what aspect would you like to see in artificial neural networks added over time 14:56.440 --> 14:59.440 as we try to figure out ways to improve them? 14:59.440 --> 15:10.440 So one of the main differences and, you know, problems in terms of deep learning today, 15:10.440 --> 15:17.440 and it's not only deep learning, and the brain is the need for deep learning techniques 15:17.440 --> 15:22.440 to have a lot of labeled examples. 15:22.440 --> 15:31.440 For instance, for ImageNet, you have a training set which is one million images, each one labeled by some human 15:31.440 --> 15:34.440 in terms of which object is there. 15:34.440 --> 15:46.440 And it's clear that in biology, a baby may be able to see a million images in the first years of life, 15:46.440 --> 15:56.440 but will not have a million of labels given to him or her by parents or caretakers. 15:56.440 --> 15:59.440 So how do you solve that? 15:59.440 --> 16:07.440 You know, I think there is this interesting challenge that today, deep learning and related techniques 16:07.440 --> 16:18.440 are all about big data, big data meaning a lot of examples labeled by humans, 16:18.440 --> 16:22.440 whereas in nature you have... 16:22.440 --> 16:29.440 So this big data is n going to infinity, that's the best, you know, n meaning labeled data. 16:29.440 --> 16:34.440 But I think the biological world is more n going to 1. 16:34.440 --> 16:42.440 A child can learn from a very small number of labeled examples. 16:42.440 --> 16:49.440 Like you tell a child, this is a car, you don't need to say like in ImageNet, you know, this is a car, this is a car, 16:49.440 --> 16:53.440 this is not a car, this is not a car, one million times. 16:53.440 --> 17:05.440 And of course with AlphaGo and AlphaZero variants, because the world of Go is so simplistic that you can actually learn by yourself 17:05.440 --> 17:08.440 through self play, you can play against each other. 17:08.440 --> 17:15.440 And the real world, the visual system that you've studied extensively is a lot more complicated than the game of Go. 17:15.440 --> 17:22.440 On the comment about children, which are fascinatingly good at learning new stuff, 17:22.440 --> 17:26.440 how much of it do you think is hardware and how much of it is software? 17:26.440 --> 17:32.440 Yeah, that's a good and deep question, in a sense is the old question of nurture and nature, 17:32.440 --> 17:40.440 how much is in the gene and how much is in the experience of an individual. 17:40.440 --> 17:55.440 Obviously, it's both that play a role and I believe that the way evolution gives put prior information, so to speak, hardwired, 17:55.440 --> 18:02.440 it's not really hardwired, but that's essentially an hypothesis. 18:02.440 --> 18:14.440 I think what's going on is that evolution is almost necessarily, if you believe in Darwin, it's very opportunistic. 18:14.440 --> 18:23.440 And think about our DNA and the DNA of Drosophila. 18:23.440 --> 18:28.440 Our DNA does not have many more genes than Drosophila. 18:28.440 --> 18:32.440 The fly, the fruit fly. 18:32.440 --> 18:39.440 Now, we know that the fruit fly does not learn very much during its individual existence. 18:39.440 --> 18:51.440 It looks like one of these machinery that it's really mostly, not 100%, but 95% hardcoded by the genes. 18:51.440 --> 19:02.440 But since we don't have many more genes than Drosophila, evolution could encode in us a kind of general learning machinery 19:02.440 --> 19:09.440 and then had to give very weak priors. 19:09.440 --> 19:20.440 Like, for instance, let me give a specific example, which is recent to work by a member of our Center for Brains, Mines and Machines. 19:20.440 --> 19:30.440 We know because of work of other people in our group and other groups that there are cells in a part of our brain, neurons, that are tuned to faces. 19:30.440 --> 19:33.440 They seem to be involved in face recognition. 19:33.440 --> 19:43.440 Now, this face area seems to be present in young children and adults. 19:43.440 --> 19:54.440 And one question is there from the beginning, is hardwired by evolution or somehow is learned very quickly. 19:54.440 --> 20:00.440 So what's your, by the way, a lot of the questions I'm asking, the answer is we don't really know, 20:00.440 --> 20:08.440 but as a person who has contributed some profound ideas in these fields, you're a good person to guess at some of these. 20:08.440 --> 20:14.440 So, of course, there's a caveat before a lot of the stuff we talk about, but what is your hunch? 20:14.440 --> 20:21.440 Is the face, the part of the brain that seems to be concentrated on face recognition, are you born with that? 20:21.440 --> 20:26.440 Or are you just designed to learn that quickly, like the face of the mother and son? 20:26.440 --> 20:42.440 My hunch, my bias was the second one, learned very quickly and turns out that Marge Livingstone at Harvard has done some amazing experiments in which she raised baby monkeys, 20:42.440 --> 20:47.440 depriving them of faces during the first weeks of life. 20:47.440 --> 20:52.440 So they see technicians, but the technicians have a mask. 20:52.440 --> 20:54.440 Yes. 20:54.440 --> 21:10.440 And so when they looked at the area in the brain of these monkeys that were usually you find faces, they found no face preference. 21:10.440 --> 21:26.440 So my guess is that what evolution does in this case is there is a plastic area, which is plastic, which is kind of predetermined to be imprinted very easily. 21:26.440 --> 21:31.440 But the command from the gene is not a detailed circuitry for a face template. 21:31.440 --> 21:33.440 Could be. 21:33.440 --> 21:35.440 But this will require probably a lot of bits. 21:35.440 --> 21:39.440 You had to specify a lot of connection of a lot of neurons. 21:39.440 --> 21:53.440 Instead, the command from the gene is something like imprint, memorize what you see most often in the first two weeks of life, especially in connection with food and maybe nipples. 21:53.440 --> 21:54.440 I don't know. 21:54.440 --> 21:55.440 Right. 21:55.440 --> 21:56.440 Well, source of food. 21:56.440 --> 22:00.440 And so in that area is very plastic at first and it solidifies. 22:00.440 --> 22:10.440 It'd be interesting if a variant of that experiment would show a different kind of pattern associated with food than a face pattern, whether that could stick. 22:10.440 --> 22:25.440 There are indications that during that experiment, what the monkeys saw quite often were the blue gloves of the technicians that were giving to the baby monkeys the milk. 22:25.440 --> 22:33.440 And some of the cells instead of being face sensitive in that area are hand sensitive. 22:33.440 --> 22:35.440 That's fascinating. 22:35.440 --> 22:45.440 Can you talk about what are the different parts of the brain and in your view sort of loosely and how do they contribute to intelligence? 22:45.440 --> 23:04.440 Do you see the brain as a bunch of different modules and they together come in the human brain to create intelligence or is it all one mush of the same kind of fundamental architecture? 23:04.440 --> 23:21.440 Yeah, that's an important question and there was a phase in neuroscience back in the 1950s or so in which it was believed for a while that the brain was equipotential. 23:21.440 --> 23:22.440 This was the term. 23:22.440 --> 23:31.440 You could cut out a piece and nothing special happened apart, a little bit less performance. 23:31.440 --> 23:50.440 There was a surgeon, Lashley, who did a lot of experiments of this type with mice and rats and concluded that every part of the brain was essentially equivalent to any other one. 23:50.440 --> 24:12.440 It turns out that that's really not true. There are very specific modules in the brain, as you said, and people may lose the ability to speak if you have a stroke in a certain region or may lose control of their legs in another region. 24:12.440 --> 24:33.440 So they're very specific. The brain is also quite flexible and redundant so often it can correct things and take over functions from one part of the brain to the other, but really there are specific modules. 24:33.440 --> 25:02.440 So the answer that we know from this old work, which was basically based on lesions, either on animals or very often there was a mine of very interesting data coming from the war, from different types of injuries that soldiers had in the brain. 25:02.440 --> 25:23.440 And more recently, functional MRI, which allow you to check which part of the brain are active when you're doing different tasks, as you can replace some of this. 25:23.440 --> 25:32.440 You can see that certain parts of the brain are involved, are active in certain tasks. 25:32.440 --> 26:01.440 But sort of taking a step back to that part of the brain that discovers that specializes in the face and how that might be learned, what's your intuition behind, you know, is it possible that the sort of from a physicist's perspective when you get lower and lower, that it's all the same stuff and it just, when you're born, it's plastic and it quickly figures out this part is going to be about vision, this is going to be about language, this is about common sense reasoning. 26:01.440 --> 26:09.440 Do you have an intuition that that kind of learning is going on really quickly or is it really kind of solidified in hardware? 26:09.440 --> 26:10.440 That's a great question. 26:10.440 --> 26:21.440 So there are parts of the brain like the cerebellum or the hippocampus that are quite different from each other. 26:21.440 --> 26:25.440 They clearly have different anatomy, different connectivity. 26:25.440 --> 26:35.440 Then there is the cortex, which is the most developed part of the brain in humans. 26:35.440 --> 26:47.440 And in the cortex, you have different regions of the cortex that are responsible for vision, for audition, for motor control, for language. 26:47.440 --> 27:07.440 Now, one of the big puzzles of this is that in the cortex, it looks like it is the same in terms of hardware, in terms of type of neurons and connectivity across these different modalities. 27:07.440 --> 27:17.440 So for the cortex, I think aside these other parts of the brain like spinal cord, hippocampus, cerebellum and so on. 27:17.440 --> 27:28.440 For the cortex, I think your question about hardware and software and learning and so on, I think is rather open. 27:28.440 --> 27:40.440 And I find it very interesting for us to think about an architecture, computer architecture that is good for vision and at the same time is good for language. 27:40.440 --> 27:48.440 It seems to be so different problem areas that you have to solve. 27:48.440 --> 27:54.440 But the underlying mechanism might be the same and that's really instructive for artificial neural networks. 27:54.440 --> 28:00.440 So we've done a lot of great work in vision and human vision, computer vision. 28:00.440 --> 28:07.440 And you mentioned the problem of human vision is really as difficult as the problem of general intelligence. 28:07.440 --> 28:10.440 And maybe that connects to the cortex discussion. 28:10.440 --> 28:21.440 Can you describe the human visual cortex and how the humans begin to understand the world through the raw sensory information? 28:21.440 --> 28:36.440 What's for folks who are not familiar, especially on the computer vision side, we don't often actually take a step back except saying with a sentence or two that one is inspired by the other. 28:36.440 --> 28:39.440 What is it that we know about the human visual cortex? 28:39.440 --> 28:40.440 That's interesting. 28:40.440 --> 28:53.440 So we know quite a bit at the same time, we don't know a lot, but the bit we know, in a sense, we know a lot of the details and many we don't know. 28:53.440 --> 29:05.440 And we know a lot of the top level, the answer to the top level question, but we don't know some basic ones, even in terms of general neuroscience forgetting vision. 29:05.440 --> 29:11.440 You know, why do we sleep? It's such a basic question. 29:11.440 --> 29:14.440 And we really don't have an answer to that. 29:14.440 --> 29:18.440 So taking a step back on that. So sleep, for example, is fascinating. 29:18.440 --> 29:21.440 Do you think that's a neuroscience question? 29:21.440 --> 29:30.440 Or if we talk about abstractions, what do you think is an interesting way to study intelligence or most effective on the levels of abstraction? 29:30.440 --> 29:37.440 Is it chemical, is it biological, is it electrophysical, mathematical as you've done a lot of excellent work on that side? 29:37.440 --> 29:42.440 Which psychology, sort of like at which level of abstraction do you think? 29:42.440 --> 29:48.440 Well, in terms of levels of abstraction, I think we need all of them. 29:48.440 --> 29:56.440 It's one, you know, it's like if you ask me, what does it mean to understand a computer? 29:56.440 --> 30:04.440 That's much simpler. But in a computer, I could say, well, understand how to use PowerPoint. 30:04.440 --> 30:13.440 That's my level of understanding a computer. It's, it has reasonable, you know, it gives me some power to produce slides and beautiful slides. 30:13.440 --> 30:28.440 And now somebody else says, well, I know how the transistor work that are inside the computer can write the equation for, you know, transistor and diodes and circuits, logical circuits. 30:28.440 --> 30:33.440 And I can ask this guy, do you know how to operate PowerPoint? No idea. 30:33.440 --> 30:49.440 So do you think if we discovered computers walking amongst us full of these transistors that are also operating under windows and have PowerPoint, do you think it's digging in a little bit more? 30:49.440 --> 31:00.440 How useful is it to understand the transistor in order to be able to understand PowerPoint in these higher level intelligence processes? 31:00.440 --> 31:12.440 So I think in the case of computers, because they were made by engineers by us, these different level of understanding are rather separate on purpose. 31:12.440 --> 31:23.440 You know, they are separate modules so that the engineer that designed the circuit for the chips does not need to know what is inside PowerPoint. 31:23.440 --> 31:30.440 And somebody can write the software translating from one to the other. 31:30.440 --> 31:40.440 So in that case, I don't think understanding the transistor help you understand PowerPoint or very little. 31:40.440 --> 31:51.440 If you want to understand the computer, this question, you know, I would say you have to understanding at different levels if you really want to build one. 31:51.440 --> 32:09.440 But for the brain, I think these levels of understanding, so the algorithms, which kind of computation, you know, the equivalent PowerPoint and the circuits, you know, the transistors, I think they are much more intertwined with each other. 32:09.440 --> 32:15.440 There is not, you know, a neatly level of the software separate from the hardware. 32:15.440 --> 32:29.440 And so that's why I think in the case of the brain, the problem is more difficult and more than for computers requires the interaction, the collaboration between different types of expertise. 32:29.440 --> 32:35.440 So the brain is a big hierarchical mess that you can't just disentangle levels. 32:35.440 --> 32:41.440 I think you can, but it's much more difficult and it's not completely obvious. 32:41.440 --> 32:47.440 And I said, I think he's one of the person I think is the greatest problem in science. 32:47.440 --> 32:52.440 So, you know, I think it's fair that it's difficult. 32:52.440 --> 32:53.440 That's a difficult one. 32:53.440 --> 32:58.440 That said, you do talk about compositionality and why it might be useful. 32:58.440 --> 33:07.440 And when you discuss why these neural networks in artificial or biological sense learn anything, you talk about compositionality. 33:07.440 --> 33:22.440 See, there's a sense that nature can be disentangled or well, all aspects of our cognition could be disentangled a little to some degree. 33:22.440 --> 33:31.440 So why do you think what, first of all, how do you see compositionality and why do you think it exists at all in nature? 33:31.440 --> 33:39.440 I spoke about, I use the term compositionality. 33:39.440 --> 33:54.440 When we looked at deep neural networks, multi layers and trying to understand when and why they are more powerful than more classical one layer networks, 33:54.440 --> 34:01.440 like linear classifier, kernel machines, so called. 34:01.440 --> 34:12.440 And what we found is that in terms of approximating or learning or representing a function, a mapping from an input to an output, 34:12.440 --> 34:20.440 like from an image to the label in the image, if this function has a particular structure, 34:20.440 --> 34:28.440 then deep networks are much more powerful than shallow networks to approximate the underlying function. 34:28.440 --> 34:33.440 And the particular structure is a structure of compositionality. 34:33.440 --> 34:45.440 If the function is made up of functions of function, so that you need to look on when you are interpreting an image, 34:45.440 --> 34:56.440 classifying an image, you don't need to look at all pixels at once, but you can compute something from small groups of pixels, 34:56.440 --> 35:04.440 and then you can compute something on the output of this local computation and so on. 35:04.440 --> 35:10.440 It is similar to what you do when you read a sentence, you don't need to read the first and the last letter, 35:10.440 --> 35:17.440 but you can read syllables, combine them in words, combine the words in sentences. 35:17.440 --> 35:20.440 So this is this kind of structure. 35:20.440 --> 35:27.440 So that's as part of a discussion of why deep neural networks may be more effective than the shallow methods. 35:27.440 --> 35:35.440 And is your sense for most things we can use neural networks for, 35:35.440 --> 35:43.440 those problems are going to be compositional in nature, like language, like vision. 35:43.440 --> 35:47.440 How far can we get in this kind of way? 35:47.440 --> 35:51.440 So here is almost philosophy. 35:51.440 --> 35:53.440 Well, let's go there. 35:53.440 --> 35:55.440 Yeah, let's go there. 35:55.440 --> 36:00.440 So friend of mine, Max Tagmark, who is a physicist at MIT. 36:00.440 --> 36:02.440 I've talked to him on this thing. 36:02.440 --> 36:04.440 Yeah, and he disagrees with you, right? 36:04.440 --> 36:09.440 We agree on most, but the conclusion is a bit different. 36:09.440 --> 36:14.440 His conclusion is that for images, for instance, 36:14.440 --> 36:23.440 the compositional structure of this function that we have to learn or to solve these problems 36:23.440 --> 36:35.440 comes from physics, comes from the fact that you have local interactions in physics between atoms and other atoms, 36:35.440 --> 36:42.440 between particle of matter and other particles, between planets and other planets, 36:42.440 --> 36:44.440 between stars and others. 36:44.440 --> 36:48.440 It's all local. 36:48.440 --> 36:55.440 And that's true, but you could push this argument a bit further. 36:55.440 --> 36:57.440 Not this argument, actually. 36:57.440 --> 37:02.440 You could argue that, you know, maybe that's part of the true, 37:02.440 --> 37:06.440 but maybe what happens is kind of the opposite, 37:06.440 --> 37:11.440 is that our brain is wired up as a deep network. 37:11.440 --> 37:22.440 So it can learn, understand, solve problems that have this compositional structure. 37:22.440 --> 37:29.440 And it cannot solve problems that don't have this compositional structure. 37:29.440 --> 37:37.440 So the problems we are accustomed to, we think about, we test our algorithms on, 37:37.440 --> 37:42.440 are this compositional structure because our brain is made up. 37:42.440 --> 37:46.440 And that's, in a sense, an evolutionary perspective that we've... 37:46.440 --> 37:54.440 So the ones that weren't dealing with the compositional nature of reality died off? 37:54.440 --> 38:05.440 Yes, but also could be, maybe the reason why we have this local connectivity in the brain, 38:05.440 --> 38:10.440 like simple cells in cortex looking only at the small part of the image, 38:10.440 --> 38:16.440 each one of them, and then other cells looking at the small number of the simple cells and so on. 38:16.440 --> 38:24.440 The reason for this may be purely that it was difficult to grow long range connectivity. 38:24.440 --> 38:33.440 So suppose it's, you know, for biology, it's possible to grow short range connectivity, 38:33.440 --> 38:39.440 but not long range also because there is a limited number of long range. 38:39.440 --> 38:44.440 And so you have this limitation from the biology. 38:44.440 --> 38:49.440 And this means you build a deep convolutional network. 38:49.440 --> 38:53.440 This would be something like a deep convolutional network. 38:53.440 --> 38:57.440 And this is great for solving certain class of problems. 38:57.440 --> 39:02.440 These are the ones we find easy and important for our life. 39:02.440 --> 39:06.440 And yes, they were enough for us to survive. 39:06.440 --> 39:13.440 And you can start a successful business on solving those problems with mobile eye. 39:13.440 --> 39:16.440 Driving is a compositional problem. 39:16.440 --> 39:25.440 So on the learning task, we don't know much about how the brain learns in terms of optimization. 39:25.440 --> 39:31.440 So the thing that's stochastic gradient descent is what artificial neural networks 39:31.440 --> 39:38.440 use for the most part to adjust the parameters in such a way that it's able to deal 39:38.440 --> 39:42.440 based on the labeled data, it's able to solve the problem. 39:42.440 --> 39:49.440 So what's your intuition about why it works at all? 39:49.440 --> 39:55.440 How hard of a problem it is to optimize a neural network, artificial neural network? 39:55.440 --> 39:57.440 Is there other alternatives? 39:57.440 --> 40:03.440 Just in general, your intuition is behind this very simplistic algorithm 40:03.440 --> 40:05.440 that seems to do pretty good, surprising. 40:05.440 --> 40:07.440 Yes, yes. 40:07.440 --> 40:16.440 So I find neuroscience, the architecture of cortex is really similar to the architecture of deep networks. 40:16.440 --> 40:26.440 So there is a nice correspondence there between the biology and this kind of local connectivity hierarchical 40:26.440 --> 40:28.440 architecture. 40:28.440 --> 40:35.440 The stochastic gradient descent, as you said, is a very simple technique. 40:35.440 --> 40:49.440 It seems pretty unlikely that biology could do that from what we know right now about cortex and neurons and synapses. 40:49.440 --> 40:58.440 So it's a big question open whether there are other optimization learning algorithms 40:58.440 --> 41:02.440 that can replace stochastic gradient descent. 41:02.440 --> 41:11.440 And my guess is yes, but nobody has found yet a real answer. 41:11.440 --> 41:17.440 I mean, people are trying, still trying, and there are some interesting ideas. 41:17.440 --> 41:27.440 The fact that stochastic gradient descent is so successful, this has become clearly not so mysterious. 41:27.440 --> 41:39.440 And the reason is that it's an interesting fact, you know, is a change in a sense in how people think about statistics. 41:39.440 --> 41:51.440 And this is the following is that typically when you had data and you had, say, a model with parameters, 41:51.440 --> 41:55.440 you are trying to fit the model to the data, you know, to fit the parameter. 41:55.440 --> 42:12.440 And typically the kind of kind of crowd wisdom type idea was you should have at least, you know, twice the number of data than the number of parameters. 42:12.440 --> 42:15.440 Maybe 10 times is better. 42:15.440 --> 42:24.440 Now, the way you train neural network these days is that they have 10 or 100 times more parameters than data. 42:24.440 --> 42:26.440 Exactly the opposite. 42:26.440 --> 42:34.440 And which, you know, it has been one of the puzzles about neural networks. 42:34.440 --> 42:40.440 How can you get something that really works when you have so much freedom? 42:40.440 --> 42:43.440 From that little data you can generalize somehow. 42:43.440 --> 42:44.440 Right, exactly. 42:44.440 --> 42:48.440 Do you think the stochastic nature of it is essential, the randomness? 42:48.440 --> 43:00.440 I think we have some initial understanding why this happens, but one nice side effect of having this over parameterization, more parameters than data, 43:00.440 --> 43:07.440 is that when you look for the minima of a loss function like stochastic degree of descent is doing, 43:07.440 --> 43:19.440 you find I made some calculations based on some old basic theorem of algebra called Bezu theorem. 43:19.440 --> 43:25.440 And that gives you an estimate of the number of solutions of a system of polynomial equation. 43:25.440 --> 43:38.440 Anyway, the bottom line is that there are probably more minima for a typical deep networks than atoms in the universe. 43:38.440 --> 43:43.440 Just to say there are a lot because of the over parameterization. 43:43.440 --> 43:44.440 Yes. 43:44.440 --> 43:48.440 More global minimum, zero minimum, good minimum. 43:48.440 --> 43:51.440 More global minimum. 43:51.440 --> 44:00.440 Yes, a lot of them, so you have a lot of solutions, so it's not so surprising that you can find them relatively easily. 44:00.440 --> 44:04.440 This is because of the over parameterization. 44:04.440 --> 44:09.440 The over parameterization sprinkles that entire space with solutions that are pretty good. 44:09.440 --> 44:11.440 It's not so surprising, right? 44:11.440 --> 44:17.440 It's like if you have a system of linear equation and you have more unknowns than equations, 44:17.440 --> 44:24.440 then we know you have an infinite number of solutions and the question is to pick one. 44:24.440 --> 44:27.440 That's another story, but you have an infinite number of solutions, 44:27.440 --> 44:32.440 so there are a lot of value of your unknowns that satisfy the equations. 44:32.440 --> 44:37.440 But it's possible that there's a lot of those solutions that aren't very good. 44:37.440 --> 44:38.440 What's surprising is that they're pretty good. 44:38.440 --> 44:39.440 So that's a separate question. 44:39.440 --> 44:43.440 Why can you pick one that generalizes one? 44:43.440 --> 44:46.440 That's a separate question with separate answers. 44:46.440 --> 44:53.440 One theorem that people like to talk about that inspires imagination of the power of neural networks 44:53.440 --> 45:00.440 is the universal approximation theorem that you can approximate any computable function 45:00.440 --> 45:04.440 with just a finite number of neurons and a single hidden layer. 45:04.440 --> 45:07.440 Do you find this theorem one surprising? 45:07.440 --> 45:12.440 Do you find it useful, interesting, inspiring? 45:12.440 --> 45:16.440 No, this one, I never found it very surprising. 45:16.440 --> 45:22.440 It was known since the 80s, since I entered the field, 45:22.440 --> 45:27.440 because it's basically the same as Viastras theorem, 45:27.440 --> 45:34.440 which says that I can approximate any continuous function with a polynomial of sufficiently, 45:34.440 --> 45:37.440 with a sufficient number of terms, monomials. 45:37.440 --> 45:41.440 It's basically the same, and the proofs are very similar. 45:41.440 --> 45:48.440 So your intuition was there was never any doubt that neural networks in theory could be very strong approximations. 45:48.440 --> 45:58.440 The interesting question is that if this theorem says you can approximate fine, 45:58.440 --> 46:06.440 but when you ask how many neurons, for instance, or in the case of how many monomials, 46:06.440 --> 46:11.440 I need to get a good approximation. 46:11.440 --> 46:20.440 Then it turns out that that depends on the dimensionality of your function, how many variables you have. 46:20.440 --> 46:25.440 But it depends on the dimensionality of your function in a bad way. 46:25.440 --> 46:35.440 For instance, suppose you want an error which is no worse than 10% in your approximation. 46:35.440 --> 46:40.440 If you want to approximate your function within 10%, 46:40.440 --> 46:48.440 then it turns out that the number of units you need are in the order of 10 to the dimensionality, d. 46:48.440 --> 46:50.440 How many variables? 46:50.440 --> 46:57.440 So if you have two variables, d is 2 and you have 100 units and OK. 46:57.440 --> 47:02.440 But if you have, say, 200 by 200 pixel images, 47:02.440 --> 47:06.440 now this is 40,000, whatever. 47:06.440 --> 47:09.440 We again go to the size of the universe pretty quickly. 47:09.440 --> 47:13.440 Exactly, 10 to the 40,000 or something. 47:13.440 --> 47:21.440 And so this is called the curse of dimensionality, not quite appropriately. 47:21.440 --> 47:27.440 And the hope is with the extra layers you can remove the curse. 47:27.440 --> 47:34.440 What we proved is that if you have deep layers or hierarchical architecture 47:34.440 --> 47:39.440 with the local connectivity of the type of convolutional deep learning, 47:39.440 --> 47:46.440 and if you're dealing with a function that has this kind of hierarchical architecture, 47:46.440 --> 47:50.440 then you avoid completely the curse. 47:50.440 --> 47:53.440 You've spoken a lot about supervised deep learning. 47:53.440 --> 47:58.440 What are your thoughts, hopes, views on the challenges of unsupervised learning 47:58.440 --> 48:04.440 with GANs, with generative adversarial networks? 48:04.440 --> 48:08.440 Do you see those as distinct, the power of GANs, 48:08.440 --> 48:12.440 do you see those as distinct from supervised methods in neural networks, 48:12.440 --> 48:16.440 or are they really all in the same representation ballpark? 48:16.440 --> 48:24.440 GANs is one way to get estimation of probability densities, 48:24.440 --> 48:29.440 which is a somewhat new way that people have not done before. 48:29.440 --> 48:38.440 I don't know whether this will really play an important role in intelligence, 48:38.440 --> 48:47.440 or it's interesting, I'm less enthusiastic about it than many people in the field. 48:47.440 --> 48:53.440 I have the feeling that many people in the field are really impressed by the ability 48:53.440 --> 49:00.440 of producing realistic looking images in this generative way. 49:00.440 --> 49:02.440 Which describes the popularity of the methods, 49:02.440 --> 49:10.440 but you're saying that while that's exciting and cool to look at, it may not be the tool that's useful for it. 49:10.440 --> 49:12.440 So you describe it kind of beautifully. 49:12.440 --> 49:17.440 Current supervised methods go N to infinity in terms of the number of labeled points, 49:17.440 --> 49:20.440 and we really have to figure out how to go to N to 1. 49:20.440 --> 49:24.440 And you're thinking GANs might help, but they might not be the right... 49:24.440 --> 49:28.440 I don't think for that problem, which I really think is important. 49:28.440 --> 49:35.440 I think they certainly have applications, for instance, in computer graphics. 49:35.440 --> 49:43.440 I did work long ago, which was a little bit similar in terms of, 49:43.440 --> 49:49.440 saying I have a network and I present images, 49:49.440 --> 49:59.440 so the input is images and output is, for instance, the pose of the image, a face, how much is smiling, 49:59.440 --> 50:02.440 is rotated 45 degrees or not. 50:02.440 --> 50:08.440 What about having a network that I train with the same data set, 50:08.440 --> 50:10.440 but now I invert input and output. 50:10.440 --> 50:16.440 Now the input is the pose or the expression, a number, certain numbers, 50:16.440 --> 50:19.440 and the output is the image and I train it. 50:19.440 --> 50:27.440 And we did pretty good interesting results in terms of producing very realistic looking images. 50:27.440 --> 50:35.440 It was less sophisticated mechanism, but the output was pretty less than GANs, 50:35.440 --> 50:38.440 but the output was pretty much of the same quality. 50:38.440 --> 50:43.440 So I think for computer graphics type application, 50:43.440 --> 50:48.440 definitely GANs can be quite useful and not only for that, 50:48.440 --> 51:01.440 but for helping, for instance, on this problem unsupervised example of reducing the number of labelled examples, 51:01.440 --> 51:10.440 I think people, it's like they think they can get out more than they put in. 51:10.440 --> 51:13.440 There's no free lunch, as you said. 51:13.440 --> 51:16.440 What's your intuition? 51:16.440 --> 51:24.440 How can we slow the growth of N to infinity in supervised learning? 51:24.440 --> 51:29.440 So, for example, Mobileye has very successfully, 51:29.440 --> 51:34.440 I mean essentially annotated large amounts of data to be able to drive a car. 51:34.440 --> 51:40.440 Now, one thought is, so we're trying to teach machines, the school of AI, 51:40.440 --> 51:45.440 and we're trying to, so how can we become better teachers, maybe? 51:45.440 --> 51:47.440 That's one way. 51:47.440 --> 51:58.440 I like that because, again, one caricature of the history of computer science, 51:58.440 --> 52:09.440 it begins with programmers, expensive, continuous labellers, cheap, 52:09.440 --> 52:16.440 and the future would be schools, like we have for kids. 52:16.440 --> 52:26.440 Currently, the labelling methods, we're not selective about which examples we teach networks with. 52:26.440 --> 52:33.440 I think the focus of making networks that learn much faster is often on the architecture side, 52:33.440 --> 52:37.440 but how can we pick better examples with which to learn? 52:37.440 --> 52:39.440 Do you have intuitions about that? 52:39.440 --> 52:50.440 Well, that's part of the problem, but the other one is, if we look at biology, 52:50.440 --> 52:58.440 the reasonable assumption, I think, is in the same spirit as I said, 52:58.440 --> 53:03.440 evolution is opportunistic and has weak priors. 53:03.440 --> 53:10.440 The way I think the intelligence of a child, a baby may develop, 53:10.440 --> 53:17.440 is by bootstrapping weak priors from evolution. 53:17.440 --> 53:26.440 For instance, you can assume that you have most organisms, 53:26.440 --> 53:37.440 including human babies, built in some basic machinery to detect motion and relative motion. 53:37.440 --> 53:46.440 In fact, we know all insects, from fruit flies to other animals, they have this. 53:46.440 --> 53:55.440 Even in the retinas, in the very peripheral part, it's very conserved across species, 53:55.440 --> 53:58.440 something that evolution discovered early. 53:58.440 --> 54:05.440 It may be the reason why babies tend to look in the first few days to moving objects, 54:05.440 --> 54:07.440 and not to not moving objects. 54:07.440 --> 54:11.440 Now, moving objects means, okay, they're attracted by motion, 54:11.440 --> 54:19.440 but motion also means that motion gives automatic segmentation from the background. 54:19.440 --> 54:26.440 So because of motion boundaries, either the object is moving, 54:26.440 --> 54:32.440 or the eye of the baby is tracking the moving object, and the background is moving. 54:32.440 --> 54:37.440 Yeah, so just purely on the visual characteristics of the scene, that seems to be the most useful. 54:37.440 --> 54:43.440 Right, so it's like looking at an object without background. 54:43.440 --> 54:49.440 It's ideal for learning the object, otherwise it's really difficult, because you have so much stuff. 54:49.440 --> 54:54.440 So suppose you do this at the beginning, first weeks, 54:54.440 --> 55:01.440 then after that you can recognize the object, now they are imprinted, the number one, 55:01.440 --> 55:05.440 even in the background, even without motion. 55:05.440 --> 55:10.440 So that's the, by the way, I just want to ask on the object recognition problem, 55:10.440 --> 55:16.440 so there is this being responsive to movement and doing edge detection, essentially. 55:16.440 --> 55:20.440 What's the gap between being effectively, 55:20.440 --> 55:27.440 effectively visually recognizing stuff, detecting where it is, and understanding the scene? 55:27.440 --> 55:32.440 Is this a huge gap in many layers, or is it close? 55:32.440 --> 55:35.440 No, I think that's a huge gap. 55:35.440 --> 55:44.440 I think present algorithm with all the success that we have, and the fact that are a lot of very useful, 55:44.440 --> 55:51.440 I think we are in a golden age for applications of low level vision, 55:51.440 --> 55:56.440 and low level speech recognition, and so on, you know, Alexa, and so on. 55:56.440 --> 56:01.440 There are many more things of similar level to be done, including medical diagnosis and so on, 56:01.440 --> 56:11.440 but we are far from what we call understanding of a scene, of language, of actions, of people. 56:11.440 --> 56:17.440 That is, despite the claims, that's, I think, very far. 56:17.440 --> 56:19.440 We're a little bit off. 56:19.440 --> 56:24.440 So in popular culture, and among many researchers, some of which I've spoken with, 56:24.440 --> 56:34.440 the Sewell Russell and Elon Musk, in and out of the AI field, there's a concern about the existential threat of AI. 56:34.440 --> 56:44.440 And how do you think about this concern, and is it valuable to think about large scale, 56:44.440 --> 56:51.440 long term, unintended consequences of intelligent systems we try to build? 56:51.440 --> 56:58.440 I always think it's better to worry first, you know, early rather than late. 56:58.440 --> 56:59.440 So worry is good. 56:59.440 --> 57:02.440 Yeah, I'm not against worrying at all. 57:02.440 --> 57:15.440 Personally, I think that, you know, it will take a long time before there is real reason to be worried. 57:15.440 --> 57:23.440 But as I said, I think it's good to put in place and think about possible safety against, 57:23.440 --> 57:35.440 what I find a bit misleading are things like that have been said by people I know, like Elon Musk and what is Bostrom in particular, 57:35.440 --> 57:39.440 and what is his first name, Nick Bostrom, right? 57:39.440 --> 57:46.440 And, you know, and a couple of other people that, for instance, AI is more dangerous than nuclear weapons. 57:46.440 --> 57:50.440 I think that's really wrong. 57:50.440 --> 57:59.440 That can be misleading, because in terms of priority, we should still be more worried about nuclear weapons 57:59.440 --> 58:05.440 and what people are doing about it and so on than AI. 58:05.440 --> 58:15.440 And you've spoken about them as obvious and yourself saying that you think you'll be about 100 years out 58:15.440 --> 58:20.440 before we have a general intelligence system that's on par with the human being. 58:20.440 --> 58:22.440 Do you have any updates for those predictions? 58:22.440 --> 58:23.440 Well, I think he said... 58:23.440 --> 58:25.440 He said 20, I think. 58:25.440 --> 58:26.440 He said 20, right. 58:26.440 --> 58:27.440 This was a couple of years ago. 58:27.440 --> 58:31.440 I have not asked him again, so I should have. 58:31.440 --> 58:38.440 Your own prediction, what's your prediction about when you'll be truly surprised 58:38.440 --> 58:42.440 and what's the confidence interval on that? 58:42.440 --> 58:46.440 You know, it's so difficult to predict the future and even the present. 58:46.440 --> 58:48.440 It's pretty hard to predict. 58:48.440 --> 58:50.440 Right, but I would be... 58:50.440 --> 58:52.440 As I said, this is completely... 58:52.440 --> 58:56.440 I would be more like Rod Brooks. 58:56.440 --> 58:59.440 I think he's about 200 years old. 58:59.440 --> 59:01.440 200 years. 59:01.440 --> 59:06.440 When we have this kind of AGI system, artificial intelligence system, 59:06.440 --> 59:12.440 you're sitting in a room with her, him, it, 59:12.440 --> 59:17.440 do you think it will be the underlying design of such a system 59:17.440 --> 59:19.440 and something we'll be able to understand? 59:19.440 --> 59:20.440 It will be simple? 59:20.440 --> 59:25.440 Do you think it will be explainable? 59:25.440 --> 59:27.440 Understandable by us? 59:27.440 --> 59:31.440 Your intuition, again, we're in the realm of philosophy a little bit. 59:31.440 --> 59:35.440 Well, probably no. 59:35.440 --> 59:42.440 But again, it depends what you really mean for understanding. 59:42.440 --> 59:53.440 I think we don't understand how deep networks work. 59:53.440 --> 59:56.440 I think we're beginning to have a theory now. 59:56.440 --> 59:59.440 But in the case of deep networks, 59:59.440 --> 1:00:06.440 or even in the case of the simpler kernel machines or linear classifier, 1:00:06.440 --> 1:00:12.440 we really don't understand the individual units or so. 1:00:12.440 --> 1:00:20.440 But we understand what the computation and the limitations and the properties of it are. 1:00:20.440 --> 1:00:24.440 It's similar to many things. 1:00:24.440 --> 1:00:29.440 Does it mean to understand how a fusion bomb works? 1:00:29.440 --> 1:00:35.440 How many of us, you know, many of us understand the basic principle 1:00:35.440 --> 1:00:40.440 and some of us may understand deeper details? 1:00:40.440 --> 1:00:44.440 In that sense, understanding is, as a community, as a civilization, 1:00:44.440 --> 1:00:46.440 can we build another copy of it? 1:00:46.440 --> 1:00:47.440 Okay. 1:00:47.440 --> 1:00:50.440 And in that sense, do you think there'll be, 1:00:50.440 --> 1:00:56.440 there'll need to be some evolutionary component where it runs away from our understanding? 1:00:56.440 --> 1:00:59.440 Or do you think it could be engineered from the ground up? 1:00:59.440 --> 1:01:02.440 The same way you go from the transistor to PowerPoint? 1:01:02.440 --> 1:01:03.440 Right. 1:01:03.440 --> 1:01:09.440 So many years ago, this was actually 40, 41 years ago, 1:01:09.440 --> 1:01:13.440 I wrote a paper with David Maher, 1:01:13.440 --> 1:01:19.440 who was one of the founding fathers of computer vision, computational vision. 1:01:19.440 --> 1:01:23.440 I wrote a paper about levels of understanding, 1:01:23.440 --> 1:01:28.440 which is related to the question we discussed earlier about understanding PowerPoint, 1:01:28.440 --> 1:01:31.440 understanding transistors and so on. 1:01:31.440 --> 1:01:38.440 And, you know, in that kind of framework, we had a level of the hardware 1:01:38.440 --> 1:01:41.440 and the top level of the algorithms. 1:01:41.440 --> 1:01:44.440 We did not have learning. 1:01:44.440 --> 1:01:54.440 Recently, I updated adding levels and one level I added to those three was learning. 1:01:54.440 --> 1:01:59.440 So, and you can imagine, you could have a good understanding 1:01:59.440 --> 1:02:04.440 of how you construct learning machine, like we do. 1:02:04.440 --> 1:02:13.440 But being unable to describe in detail what the learning machines will discover, right? 1:02:13.440 --> 1:02:19.440 Now, that would be still a powerful understanding if I can build a learning machine, 1:02:19.440 --> 1:02:25.440 even if I don't understand in detail every time it learns something. 1:02:25.440 --> 1:02:31.440 Just like our children, if they start listening to a certain type of music, 1:02:31.440 --> 1:02:33.440 I don't know, Miley Cyrus or something, 1:02:33.440 --> 1:02:37.440 you don't understand why they came to that particular preference, 1:02:37.440 --> 1:02:39.440 but you understand the learning process. 1:02:39.440 --> 1:02:41.440 That's very interesting. 1:02:41.440 --> 1:02:50.440 So, on learning for systems to be part of our world, 1:02:50.440 --> 1:02:56.440 it has a certain, one of the challenging things that you've spoken about is learning ethics, 1:02:56.440 --> 1:02:59.440 learning morals. 1:02:59.440 --> 1:03:06.440 And how hard do you think is the problem of, first of all, humans understanding our ethics? 1:03:06.440 --> 1:03:10.440 What is the origin on the neural and low level of ethics? 1:03:10.440 --> 1:03:12.440 What is it at the higher level? 1:03:12.440 --> 1:03:17.440 Is it something that's learnable from machines in your intuition? 1:03:17.440 --> 1:03:23.440 I think, yeah, ethics is learnable, very likely. 1:03:23.440 --> 1:03:36.440 I think it's one of these problems where I think understanding the neuroscience of ethics, 1:03:36.440 --> 1:03:42.440 people discuss there is an ethics of neuroscience. 1:03:42.440 --> 1:03:46.440 How a neuroscientist should or should not behave, 1:03:46.440 --> 1:03:53.440 can think of a neurosurgeon and the ethics that he or she has to be. 1:03:53.440 --> 1:03:57.440 But I'm more interested in the neuroscience of ethics. 1:03:57.440 --> 1:04:01.440 You're blowing my mind right now, the neuroscience of ethics, it's very meta. 1:04:01.440 --> 1:04:09.440 And I think that would be important to understand also for being able to design machines 1:04:09.440 --> 1:04:14.440 that are ethical machines in our sense of ethics. 1:04:14.440 --> 1:04:20.440 And you think there is something in neuroscience, there's patterns, 1:04:20.440 --> 1:04:25.440 tools in neuroscience that could help us shed some light on ethics 1:04:25.440 --> 1:04:29.440 or is it more on the psychologist's sociology at a much higher level? 1:04:29.440 --> 1:04:33.440 No, there is psychology, but there is also, in the meantime, 1:04:33.440 --> 1:04:41.440 there is evidence, fMRI, of specific areas of the brain 1:04:41.440 --> 1:04:44.440 that are involved in certain ethical judgment. 1:04:44.440 --> 1:04:49.440 And not only this, you can stimulate those areas with magnetic fields 1:04:49.440 --> 1:04:54.440 and change the ethical decisions. 1:04:54.440 --> 1:05:00.440 So that's work by a colleague of mine, Rebecca Sacks, 1:05:00.440 --> 1:05:04.440 and there are other researchers doing similar work. 1:05:04.440 --> 1:05:11.440 And I think this is the beginning, but ideally at some point 1:05:11.440 --> 1:05:17.440 we'll have an understanding of how this works and why it evolved, right? 1:05:17.440 --> 1:05:21.440 The big why question, yeah, it must have some purpose. 1:05:21.440 --> 1:05:29.440 Yeah, obviously it has some social purposes, probably. 1:05:29.440 --> 1:05:34.440 If neuroscience holds the key to at least eliminate some aspect of ethics, 1:05:34.440 --> 1:05:36.440 that means it could be a learnable problem. 1:05:36.440 --> 1:05:38.440 Yeah, exactly. 1:05:38.440 --> 1:05:41.440 And as we're getting into harder and harder questions, 1:05:41.440 --> 1:05:44.440 let's go to the hard problem of consciousness. 1:05:44.440 --> 1:05:51.440 Is this an important problem for us to think about and solve on the engineering 1:05:51.440 --> 1:05:55.440 of intelligence side of your work, of our dream? 1:05:55.440 --> 1:05:57.440 You know, it's unclear. 1:05:57.440 --> 1:06:04.440 So, again, this is a deep problem, partly because it's very difficult 1:06:04.440 --> 1:06:16.440 to define consciousness and there is a debate among neuroscientists 1:06:16.440 --> 1:06:22.440 about whether consciousness and philosophers, of course, 1:06:22.440 --> 1:06:30.440 whether consciousness is something that requires flesh and blood, so to speak, 1:06:30.440 --> 1:06:40.440 or could be, you know, that we could have silicon devices that are conscious, 1:06:40.440 --> 1:06:45.440 or up to a statement like everything has some degree of consciousness 1:06:45.440 --> 1:06:48.440 and some more than others. 1:06:48.440 --> 1:06:53.440 This is like Giulio Tonioni and Fee. 1:06:53.440 --> 1:06:56.440 We just recently talked to Christof Ko. 1:06:56.440 --> 1:07:00.440 Christof was my first graduate student. 1:07:00.440 --> 1:07:06.440 Do you think it's important to illuminate aspects of consciousness 1:07:06.440 --> 1:07:10.440 in order to engineer intelligence systems? 1:07:10.440 --> 1:07:14.440 Do you think an intelligence system would ultimately have consciousness? 1:07:14.440 --> 1:07:18.440 Are they intro linked? 1:07:18.440 --> 1:07:23.440 You know, most of the people working in artificial intelligence, I think, 1:07:23.440 --> 1:07:29.440 they answer, we don't strictly need consciousness to have an intelligence system. 1:07:29.440 --> 1:07:35.440 That's sort of the easier question, because it's a very engineering answer to the question. 1:07:35.440 --> 1:07:38.440 It has a touring test, we don't need consciousness. 1:07:38.440 --> 1:07:47.440 But if you were to go, do you think it's possible that we need to have that kind of self awareness? 1:07:47.440 --> 1:07:49.440 We may, yes. 1:07:49.440 --> 1:08:00.440 So, for instance, I personally think that when test a machine or a person in a touring test, 1:08:00.440 --> 1:08:10.440 in an extended touring testing, I think consciousness is part of what we require in that test, 1:08:10.440 --> 1:08:14.440 you know, implicitly to say that this is intelligent. 1:08:14.440 --> 1:08:17.440 Christof disagrees. 1:08:17.440 --> 1:08:19.440 Yes, he does. 1:08:19.440 --> 1:08:24.440 Despite many other romantic notions he holds, he disagrees with that one. 1:08:24.440 --> 1:08:26.440 Yes, that's right. 1:08:26.440 --> 1:08:29.440 So, you know, who would see? 1:08:29.440 --> 1:08:37.440 Do you think, as a quick question, Ernest Becker's fear of death, 1:08:37.440 --> 1:08:48.440 do you think mortality and those kinds of things are important for consciousness and for intelligence, 1:08:48.440 --> 1:08:53.440 the finiteness of life, finiteness of existence, 1:08:53.440 --> 1:09:00.440 or is that just a side effect of evolutionary side effect that's useful for natural selection? 1:09:00.440 --> 1:09:05.440 Do you think this kind of thing that this interview is going to run out of time soon, 1:09:05.440 --> 1:09:08.440 our life will run out of time soon? 1:09:08.440 --> 1:09:12.440 Do you think that's needed to make this conversation good and life good? 1:09:12.440 --> 1:09:14.440 You know, I never thought about it. 1:09:14.440 --> 1:09:16.440 It's a very interesting question. 1:09:16.440 --> 1:09:25.440 I think Steve Jobs in his commencement speech at Stanford argued that, you know, 1:09:25.440 --> 1:09:30.440 having a finite life was important for stimulating achievements. 1:09:30.440 --> 1:09:32.440 It was a different. 1:09:32.440 --> 1:09:34.440 You live every day like it's your last, right? 1:09:34.440 --> 1:09:35.440 Yeah. 1:09:35.440 --> 1:09:45.440 So, rationally, I don't think strictly you need mortality for consciousness, but... 1:09:45.440 --> 1:09:46.440 Who knows? 1:09:46.440 --> 1:09:49.440 They seem to go together in our biological system, right? 1:09:49.440 --> 1:09:51.440 Yeah. 1:09:51.440 --> 1:09:57.440 You've mentioned before and the students are associated with... 1:09:57.440 --> 1:10:01.440 AlphaGo immobilized the big recent success stories in AI. 1:10:01.440 --> 1:10:05.440 I think it's captivated the entire world of what AI can do. 1:10:05.440 --> 1:10:10.440 So, what do you think will be the next breakthrough? 1:10:10.440 --> 1:10:13.440 What's your intuition about the next breakthrough? 1:10:13.440 --> 1:10:16.440 Of course, I don't know where the next breakthrough is. 1:10:16.440 --> 1:10:22.440 I think that there is a good chance, as I said before, that the next breakthrough 1:10:22.440 --> 1:10:27.440 would also be inspired by, you know, neuroscience. 1:10:27.440 --> 1:10:31.440 But which one? 1:10:31.440 --> 1:10:32.440 I don't know. 1:10:32.440 --> 1:10:33.440 And there's... 1:10:33.440 --> 1:10:35.440 So, MIT has this quest for intelligence. 1:10:35.440 --> 1:10:36.440 Yeah. 1:10:36.440 --> 1:10:41.440 And there's a few moonshots which, in that spirit, which ones are you excited about? 1:10:41.440 --> 1:10:42.440 What... 1:10:42.440 --> 1:10:44.440 Which projects kind of... 1:10:44.440 --> 1:10:48.440 Well, of course, I'm excited about one of the moonshots with... 1:10:48.440 --> 1:10:52.440 Which is our center for brains, minds, and machines. 1:10:52.440 --> 1:10:57.440 The one which is fully funded by NSF. 1:10:57.440 --> 1:10:59.440 And it's a... 1:10:59.440 --> 1:11:02.440 It is about visual intelligence. 1:11:02.440 --> 1:11:05.440 And that one is particularly about understanding. 1:11:05.440 --> 1:11:07.440 Visual intelligence. 1:11:07.440 --> 1:11:16.440 Visual cortex and visual intelligence in the sense of how we look around ourselves 1:11:16.440 --> 1:11:25.440 and understand the world around ourselves, you know, meaning what is going on, 1:11:25.440 --> 1:11:31.440 how we could go from here to there without hitting obstacles. 1:11:31.440 --> 1:11:36.440 You know, whether there are other agents, people in the environment. 1:11:36.440 --> 1:11:41.440 These are all things that we perceive very quickly. 1:11:41.440 --> 1:11:47.440 And it's something actually quite close to being conscious, not quite. 1:11:47.440 --> 1:11:53.440 But there is this interesting experiment that was run at Google X, 1:11:53.440 --> 1:11:58.440 which is, in a sense, is just a virtual reality experiment, 1:11:58.440 --> 1:12:09.440 but in which they had subject sitting, say, in a chair with goggles, like Oculus and so on. 1:12:09.440 --> 1:12:11.440 Earphones. 1:12:11.440 --> 1:12:20.440 And they were seeing through the eyes of a robot nearby to cameras, microphones for receiving. 1:12:20.440 --> 1:12:23.440 So their sensory system was there. 1:12:23.440 --> 1:12:30.440 And the impression of all the subjects, very strong, they could not shake it off, 1:12:30.440 --> 1:12:35.440 was that they were where the robot was. 1:12:35.440 --> 1:12:42.440 They could look at themselves from the robot and still feel they were where the robot is. 1:12:42.440 --> 1:12:45.440 They were looking at their body. 1:12:45.440 --> 1:12:48.440 Their self had moved. 1:12:48.440 --> 1:12:54.440 So some aspect of seeing understanding has to have ability to place yourself, 1:12:54.440 --> 1:12:59.440 have a self awareness about your position in the world and what the world is. 1:12:59.440 --> 1:13:04.440 So we may have to solve the heart problem of consciousness to solve it. 1:13:04.440 --> 1:13:05.440 On their way, yes. 1:13:05.440 --> 1:13:07.440 It's quite a moonshot. 1:13:07.440 --> 1:13:14.440 So you've been an advisor to some incredible minds, including Demis Osabis, Christof Koch, 1:13:14.440 --> 1:13:21.440 Amna Shashwar, like you said, all went on to become seminal figures in their respective fields. 1:13:21.440 --> 1:13:28.440 From your own success as a researcher and from perspective as a mentor of these researchers, 1:13:28.440 --> 1:13:33.440 having guided them in the way of advice, 1:13:33.440 --> 1:13:39.440 what does it take to be successful in science and engineering careers? 1:13:39.440 --> 1:13:47.440 Whether you're talking to somebody in their teens, 20s and 30s, what does that path look like? 1:13:47.440 --> 1:13:52.440 It's curiosity and having fun. 1:13:52.440 --> 1:14:01.440 And I think it's important also having fun with other curious minds. 1:14:01.440 --> 1:14:06.440 It's the people you surround with to have fun and curiosity. 1:14:06.440 --> 1:14:09.440 You mentioned Steve Jobs. 1:14:09.440 --> 1:14:14.440 Is there also an underlying ambition that's unique that you saw, 1:14:14.440 --> 1:14:18.440 or is it really does boil down to insatiable curiosity and fun? 1:14:18.440 --> 1:14:20.440 Well, of course. 1:14:20.440 --> 1:14:29.440 It's being curious in an active and ambitious way, yes, definitely. 1:14:29.440 --> 1:14:38.440 But I think sometime in science, there are friends of mine who are like this. 1:14:38.440 --> 1:14:44.440 You know, there are some of the scientists who like to work by themselves 1:14:44.440 --> 1:14:54.440 and kind of communicate only when they complete their work or discover something. 1:14:54.440 --> 1:15:02.440 I think I always found the actual process of discovering something 1:15:02.440 --> 1:15:09.440 is more fun if it's together with other intelligent and curious and fun people. 1:15:09.440 --> 1:15:13.440 So if you see the fun in that process, the side effect of that process 1:15:13.440 --> 1:15:16.440 would be that you'll actually end up discovering something. 1:15:16.440 --> 1:15:25.440 So as you've led many incredible efforts here, what's the secret to being a good advisor, 1:15:25.440 --> 1:15:28.440 mentor, leader in a research setting? 1:15:28.440 --> 1:15:35.440 Is it a similar spirit or what advice could you give to people, young faculty and so on? 1:15:35.440 --> 1:15:42.440 It's partly repeating what I said about an environment that should be friendly and fun 1:15:42.440 --> 1:15:52.440 and ambitious and, you know, I think I learned a lot from some of my advisors and friends 1:15:52.440 --> 1:16:02.440 and some were physicists and there was, for instance, this behavior that was encouraged 1:16:02.440 --> 1:16:08.440 of when somebody comes with a new idea in the group, unless it's really stupid 1:16:08.440 --> 1:16:11.440 but you are always enthusiastic. 1:16:11.440 --> 1:16:14.440 And then you're enthusiastic for a few minutes, for a few hours. 1:16:14.440 --> 1:16:22.440 Then you start, you know, asking critically a few questions, testing this. 1:16:22.440 --> 1:16:28.440 But, you know, this is a process that is, I think it's very good. 1:16:28.440 --> 1:16:30.440 You have to be enthusiastic. 1:16:30.440 --> 1:16:33.440 Sometimes people are very critical from the beginning. 1:16:33.440 --> 1:16:35.440 That's not... 1:16:35.440 --> 1:16:37.440 Yes, you have to give it a chance. 1:16:37.440 --> 1:16:38.440 Yes. 1:16:38.440 --> 1:16:39.440 That's seed to grow. 1:16:39.440 --> 1:16:44.440 That said, with some of your ideas, which are quite revolutionary, so there's a witness, 1:16:44.440 --> 1:16:49.440 especially in the human vision side and neuroscience side, there could be some pretty heated arguments. 1:16:49.440 --> 1:16:51.440 Do you enjoy these? 1:16:51.440 --> 1:16:55.440 Is that a part of science and academic pursuits that you enjoy? 1:16:55.440 --> 1:16:56.440 Yeah. 1:16:56.440 --> 1:17:00.440 Is that something that happens in your group as well? 1:17:00.440 --> 1:17:02.440 Yeah, absolutely. 1:17:02.440 --> 1:17:14.440 I also spent some time in Germany again, there is this tradition in which people are more forthright, less kind than here. 1:17:14.440 --> 1:17:23.440 So, you know, in the US, when you write a bad letter, you still say, this guy is nice, you know. 1:17:23.440 --> 1:17:25.440 Yes, yes. 1:17:25.440 --> 1:17:26.440 So... 1:17:26.440 --> 1:17:28.440 Yeah, here in America it's degrees of nice. 1:17:28.440 --> 1:17:29.440 Yes. 1:17:29.440 --> 1:17:31.440 It's all just degrees of nice, yeah. 1:17:31.440 --> 1:17:44.440 Right, so as long as this does not become personal and it's really like, you know, a football game with its rules, that's great. 1:17:44.440 --> 1:17:46.440 It's fun. 1:17:46.440 --> 1:17:58.440 So, if you somehow find yourself in a position to ask one question of an oracle, like a genie, maybe a god, and you're guaranteed to get a clear answer, 1:17:58.440 --> 1:18:00.440 what kind of question would you ask? 1:18:00.440 --> 1:18:03.440 What would be the question you would ask? 1:18:03.440 --> 1:18:09.440 In the spirit of our discussion, it could be, how could I become ten times more intelligent? 1:18:09.440 --> 1:18:15.440 And so, but see, you only get a clear short answer. 1:18:15.440 --> 1:18:18.440 So, do you think there's a clear short answer to that? 1:18:18.440 --> 1:18:19.440 No. 1:18:19.440 --> 1:18:22.440 And that's the answer you'll get. 1:18:22.440 --> 1:18:23.440 Okay. 1:18:23.440 --> 1:18:26.440 So, you've mentioned Flowers of Algernon. 1:18:26.440 --> 1:18:27.440 Oh, yeah. 1:18:27.440 --> 1:18:32.440 There's a story that inspired you in your childhood. 1:18:32.440 --> 1:18:48.440 As this story of a mouse, a human achieving genius level intelligence, and then understanding what was happening while slowly becoming not intelligent again in this tragedy of gaining intelligence and losing intelligence. 1:18:48.440 --> 1:18:59.440 Do you think in that spirit, in that story, do you think intelligence is a gift or a curse from the perspective of happiness and meaning of life? 1:18:59.440 --> 1:19:10.440 You try to create an intelligent system that understands the universe, but on an individual level, the meaning of life, do you think intelligence is a gift? 1:19:10.440 --> 1:19:16.440 It's a good question. 1:19:16.440 --> 1:19:22.440 I don't know. 1:19:22.440 --> 1:19:34.440 As one of the, as one people who consider the smartest people in the world, in some, in some dimension at the very least, what do you think? 1:19:34.440 --> 1:19:35.440 I don't know. 1:19:35.440 --> 1:19:39.440 It may be invariant to intelligence, let's agree of happiness. 1:19:39.440 --> 1:19:43.440 It would be nice if it were. 1:19:43.440 --> 1:19:44.440 That's the hope. 1:19:44.440 --> 1:19:45.440 Yeah. 1:19:45.440 --> 1:19:49.440 You could be smart and happy and clueless and happy. 1:19:49.440 --> 1:19:51.440 Yeah. 1:19:51.440 --> 1:19:56.440 As always on the discussion of the meaning of life is probably a good place to end. 1:19:56.440 --> 1:19:58.440 Tomasso, thank you so much for talking today. 1:19:58.440 --> 1:19:59.440 Thank you. 1:19:59.440 --> 1:20:19.440 This was great.