WEBVTT 00:00.000 --> 00:03.080 The following is a conversation with Yanlacun. 00:03.080 --> 00:06.320 He's considered to be one of the fathers of deep learning, 00:06.320 --> 00:09.040 which, if you've been hiding under a rock, 00:09.040 --> 00:12.240 is the recent revolution in AI that has captivated the world 00:12.240 --> 00:16.160 with the possibility of what machines can learn from data. 00:16.160 --> 00:18.520 He's a professor at New York University, 00:18.520 --> 00:21.720 a vice president and chief AI scientist at Facebook, 00:21.720 --> 00:24.320 and co recipient of the Turing Award 00:24.320 --> 00:26.240 for his work on deep learning. 00:26.240 --> 00:28.880 He's probably best known as the founding father 00:28.880 --> 00:30.760 of convolutional neural networks, 00:30.760 --> 00:32.520 in particular, their application 00:32.520 --> 00:34.440 to optical character recognition 00:34.440 --> 00:37.280 and the famed MNIST dataset. 00:37.280 --> 00:40.160 He is also an outspoken personality, 00:40.160 --> 00:43.840 unafraid to speak his mind in a distinctive French accent 00:43.840 --> 00:45.760 and explore provocative ideas, 00:45.760 --> 00:48.400 both in the rigorous medium of academic research 00:48.400 --> 00:52.840 and the somewhat less rigorous medium of Twitter and Facebook. 00:52.840 --> 00:55.640 This is the Artificial Intelligence Podcast. 00:55.640 --> 00:58.000 If you enjoy it, subscribe on YouTube, 00:58.000 --> 00:59.520 give it five stars on iTunes, 00:59.520 --> 01:01.000 support it on Patreon, 01:01.000 --> 01:03.880 or simply connect with me on Twitter at Lex Freedman, 01:03.880 --> 01:06.880 spelled F R I D M A N. 01:06.880 --> 01:10.680 And now, here's my conversation with Leon Lacun. 01:11.760 --> 01:13.840 You said that 2001 Space Odyssey 01:13.840 --> 01:15.360 is one of your favorite movies. 01:16.280 --> 01:20.400 Hal 9000 decides to get rid of the astronauts 01:20.400 --> 01:23.080 for people who haven't seen the movie, Spoiler Alert, 01:23.080 --> 01:27.160 because he, it, she believes 01:27.160 --> 01:31.640 that the astronauts, they will interfere with the mission. 01:31.640 --> 01:34.720 Do you see Hal as flawed in some fundamental way 01:34.720 --> 01:38.480 or even evil, or did he do the right thing? 01:38.480 --> 01:39.360 Neither. 01:39.360 --> 01:43.280 There's no notion of evil in that, in that context, 01:43.280 --> 01:44.760 other than the fact that people die, 01:44.760 --> 01:48.760 but it was an example of what people call 01:48.760 --> 01:50.160 value misalignment, right? 01:50.160 --> 01:52.160 You give an objective to a machine, 01:52.160 --> 01:55.720 and the machine tries to achieve this objective. 01:55.720 --> 01:58.160 And if you don't put any constraints on this objective, 01:58.160 --> 02:00.960 like don't kill people and don't do things like this, 02:02.280 --> 02:06.280 the machine, given the power, will do stupid things 02:06.280 --> 02:08.040 just to achieve this, this objective, 02:08.040 --> 02:10.240 or damaging things to achieve this objective. 02:10.240 --> 02:12.480 It's a little bit like, I mean, we are used to this 02:12.480 --> 02:14.340 in the context of human society. 02:15.760 --> 02:20.760 We, we put in place laws to prevent people 02:21.000 --> 02:22.160 from doing bad things, 02:22.160 --> 02:24.840 because spontaneously they would do those bad things, right? 02:24.840 --> 02:28.400 So we have to shape their cost function, 02:28.400 --> 02:30.160 their objective function, if you want, through laws 02:30.160 --> 02:33.360 to kind of correct an education, obviously, 02:33.360 --> 02:35.200 to sort of correct for those. 02:36.160 --> 02:41.160 So maybe just pushing a little further on that point. 02:41.960 --> 02:44.360 Hal, you know, there's a mission. 02:44.360 --> 02:47.640 There's a fuzziness around the ambiguity 02:47.640 --> 02:49.800 around what the actual mission is. 02:49.800 --> 02:54.800 But, you know, do you think that there will be a time 02:55.120 --> 02:56.760 from a utilitarian perspective, 02:56.760 --> 02:59.680 when AI system, where it is not misalignment, 02:59.680 --> 03:02.840 where it is alignment for the greater good of society, 03:02.840 --> 03:05.920 that an AI system will make decisions that are difficult? 03:05.920 --> 03:06.840 Well, that's the trick. 03:06.840 --> 03:10.840 I mean, eventually we'll have to figure out how to do this. 03:10.840 --> 03:12.640 And again, we're not starting from scratch 03:12.640 --> 03:16.480 because we've been doing this with humans for millennia. 03:16.480 --> 03:19.160 So designing objective functions for people 03:19.160 --> 03:20.880 is something that we know how to do. 03:20.880 --> 03:24.600 And we don't do it by, you know, programming things, 03:24.600 --> 03:29.040 although the legal code is called code. 03:29.040 --> 03:30.760 So that tells you something. 03:30.760 --> 03:33.040 And it's actually the design of an objective function. 03:33.040 --> 03:34.600 That's really what legal code is, right? 03:34.600 --> 03:36.280 It tells you, here is what you can do, 03:36.280 --> 03:37.440 here is what you can't do. 03:37.440 --> 03:39.040 If you do it, you pay that much, 03:39.040 --> 03:40.720 that's an objective function. 03:41.680 --> 03:44.600 So there is this idea somehow that it's a new thing 03:44.600 --> 03:46.600 for people to try to design objective functions 03:46.600 --> 03:47.960 that are aligned with the common good. 03:47.960 --> 03:49.880 But no, we've been writing laws for millennia 03:49.880 --> 03:52.080 and that's exactly what it is. 03:52.080 --> 03:54.520 So that's where, you know, 03:54.520 --> 03:59.520 the science of lawmaking and computer science will... 04:00.560 --> 04:01.400 Come together. 04:01.400 --> 04:02.840 Will come together. 04:02.840 --> 04:06.760 So there's nothing special about how our AI systems 04:06.760 --> 04:09.480 is just the continuation of tools used 04:09.480 --> 04:11.720 to make some of these difficult ethical judgments 04:11.720 --> 04:13.000 that laws make. 04:13.000 --> 04:15.080 Yeah, and we have systems like this already 04:15.080 --> 04:20.000 that make many decisions for ourselves in society 04:20.000 --> 04:22.640 that need to be designed in a way that they... 04:22.640 --> 04:24.200 Like, you know, rules about things 04:24.200 --> 04:27.520 that sometimes have bad side effects. 04:27.520 --> 04:29.600 And we have to be flexible enough about those rules 04:29.600 --> 04:31.600 so that they can be broken when it's obvious 04:31.600 --> 04:33.000 that they shouldn't be applied. 04:34.040 --> 04:35.680 So you don't see this on the camera here, 04:35.680 --> 04:36.960 but all the decoration in this room 04:36.960 --> 04:39.760 is all pictures from 2001, it's based out of C. 04:39.760 --> 04:41.400 That's it. 04:41.400 --> 04:43.080 Wow, is that by accident? 04:43.080 --> 04:43.920 Or is there a lot? 04:43.920 --> 04:45.480 The accident is by design. 04:47.480 --> 04:48.480 Oh, wow. 04:48.480 --> 04:52.560 So if you were to build HAL 10,000, 04:52.560 --> 04:57.080 so an improvement of HAL 9,000, what would you improve? 04:57.080 --> 04:59.160 Well, first of all, I wouldn't ask you 04:59.160 --> 05:01.960 to hold secrets and tell lies 05:01.960 --> 05:03.840 because that's really what breaks it in the end. 05:03.840 --> 05:07.160 That's the fact that it's asking itself questions 05:07.160 --> 05:08.880 about the purpose of the mission. 05:08.880 --> 05:10.880 And it's, you know, pieces things together 05:10.880 --> 05:11.720 that it's heard, you know, 05:11.720 --> 05:13.960 all the secrecy of the preparation of the mission 05:13.960 --> 05:17.680 and the fact that it was discovery on the lunar surface 05:17.680 --> 05:19.120 that really was kept secret. 05:19.120 --> 05:22.320 And one part of HAL's memory knows this 05:22.320 --> 05:24.680 and the other part is, does not know it 05:24.680 --> 05:26.680 and is supposed to not tell anyone 05:26.680 --> 05:28.560 and that creates internal conflict. 05:28.560 --> 05:32.200 So you think there's never should be a set of things 05:32.200 --> 05:35.480 that an AI system should not be allowed, 05:36.560 --> 05:39.880 like a set of facts that should not be shared 05:39.880 --> 05:42.520 with the human operators? 05:42.520 --> 05:44.160 Well, I think, no, I think that, 05:44.160 --> 05:47.480 I think it should be a bit like in the design 05:47.480 --> 05:51.960 of autonomous AI systems. 05:51.960 --> 05:54.200 There should be the equivalent of, you know, 05:54.200 --> 05:59.040 the oath that hypocrites oaths 05:59.040 --> 06:02.560 that doctors sign up to, right? 06:02.560 --> 06:04.040 So there's certain things, certain rules 06:04.040 --> 06:05.960 that you have to abide by. 06:05.960 --> 06:09.000 And we can sort of hardwire this into our machines 06:09.000 --> 06:11.000 to kind of make sure they don't go. 06:11.000 --> 06:15.280 So I'm not, you know, an advocate of the $3 of robotics, 06:15.280 --> 06:17.120 you know, the azimov kind of thing 06:17.120 --> 06:18.560 because I don't think it's practical, 06:18.560 --> 06:23.240 but, you know, some level of limits. 06:23.240 --> 06:27.000 But to be clear, this is not, 06:27.000 --> 06:32.000 these are not questions that are kind of reworth asking today 06:32.040 --> 06:34.360 because we just don't have the technology to do this. 06:34.360 --> 06:36.440 We don't have autonomous intelligent machines. 06:36.440 --> 06:37.560 We have intelligent machines. 06:37.560 --> 06:41.000 Some are intelligent machines that are very specialized, 06:41.000 --> 06:43.360 but they don't really sort of satisfy an objective. 06:43.360 --> 06:46.520 They're just, you know, kind of trained to do one thing. 06:46.520 --> 06:50.000 So until we have some idea for design 06:50.000 --> 06:53.360 of a full fledged autonomous intelligent system, 06:53.360 --> 06:55.680 asking the question of how design is subjective, 06:55.680 --> 06:58.600 I think is a little too abstract. 06:58.600 --> 06:59.680 It's a little too abstract. 06:59.680 --> 07:01.600 There's useful elements to it 07:01.600 --> 07:04.240 in that it helps us understand 07:04.240 --> 07:07.960 our own ethical codes, humans. 07:07.960 --> 07:10.240 So even just as a thought experiment, 07:10.240 --> 07:14.280 if you imagine that an AGI system is here today, 07:14.280 --> 07:15.920 how would we program it 07:15.920 --> 07:18.360 is a kind of nice thought experiment of constructing, 07:18.360 --> 07:23.360 how should we have a system of laws for us humans? 07:24.360 --> 07:26.800 It's just a nice practical tool. 07:26.800 --> 07:29.760 And I think there's echoes of that idea too 07:29.760 --> 07:32.160 in the AI systems we have today. 07:32.160 --> 07:33.960 They don't have to be that intelligent. 07:33.960 --> 07:34.800 Yeah. 07:34.800 --> 07:35.640 Like autonomous vehicles. 07:35.640 --> 07:37.760 These things start creeping in 07:37.760 --> 07:39.200 that they're worth thinking about, 07:39.200 --> 07:41.880 but certainly they shouldn't be framed as how. 07:43.720 --> 07:46.720 Looking back, what is the most, 07:46.720 --> 07:49.440 I'm sorry if it's a silly question, 07:49.440 --> 07:51.440 but what is the most beautiful 07:51.440 --> 07:53.800 or surprising idea in deep learning 07:53.800 --> 07:56.320 or AI in general that you've ever come across? 07:56.320 --> 07:58.560 So personally, when you said back, 08:00.040 --> 08:01.960 and just had this kind of, 08:01.960 --> 08:03.920 oh, that's pretty cool moment. 08:03.920 --> 08:04.760 That's nice. 08:04.760 --> 08:05.600 That's surprising. 08:05.600 --> 08:06.560 I don't know if it's an idea 08:06.560 --> 08:11.040 rather than a sort of empirical fact. 08:12.200 --> 08:16.480 The fact that you can build gigantic neural nets, 08:16.480 --> 08:21.480 train them on relatively small amounts of data relatively 08:23.440 --> 08:24.840 with stochastic gradient descent, 08:24.840 --> 08:26.960 and that it actually works, 08:26.960 --> 08:29.280 breaks everything you read in every textbook, right? 08:29.280 --> 08:31.520 Every pre deep learning textbook 08:31.520 --> 08:33.920 I told you, you need to have fewer parameters 08:33.920 --> 08:35.560 and you have data samples. 08:37.080 --> 08:38.760 If you have nonconvex objective function, 08:38.760 --> 08:40.680 you have no guarantee of convergence. 08:40.680 --> 08:42.080 All those things that you read in textbook, 08:42.080 --> 08:43.480 and they tell you, stay away from this, 08:43.480 --> 08:45.160 and they're all wrong. 08:45.160 --> 08:48.080 Huge number of parameters, nonconvex, 08:48.080 --> 08:50.320 and somehow which is very relative 08:50.320 --> 08:53.480 to the number of parameters data, 08:53.480 --> 08:55.080 it's able to learn anything. 08:55.080 --> 08:57.520 Does that still surprise you today? 08:57.520 --> 09:02.000 Well, it was kind of obvious to me before I knew anything 09:02.000 --> 09:04.120 that this is a good idea. 09:04.120 --> 09:06.040 And then it became surprising that it worked 09:06.040 --> 09:08.240 because I started reading those textbooks. 09:09.240 --> 09:12.320 Okay, so do you talk through the intuition 09:12.320 --> 09:14.360 of why it was obvious to you if you remember? 09:14.360 --> 09:16.120 Well, okay, so the intuition was, 09:16.120 --> 09:19.960 it's sort of like those people in the late 19th century 09:19.960 --> 09:24.960 who proved that heavier than air flight was impossible, right? 09:25.480 --> 09:26.800 And of course you have birds, right? 09:26.800 --> 09:28.280 They do fly. 09:28.280 --> 09:30.320 And so on the face of it, 09:30.320 --> 09:33.200 it's obviously wrong as an empirical question, right? 09:33.200 --> 09:35.960 And so we have the same kind of thing that, 09:35.960 --> 09:38.560 you know, we know that the brain works. 09:38.560 --> 09:39.920 We don't know how, but we know it works. 09:39.920 --> 09:42.440 And we know it's a large network of neurons 09:42.440 --> 09:44.280 and interaction and that learning takes place 09:44.280 --> 09:45.360 by changing the connection. 09:45.360 --> 09:48.000 So kind of getting this level of inspiration 09:48.000 --> 09:49.320 without covering the details, 09:49.320 --> 09:52.520 but sort of trying to derive basic principles. 09:52.520 --> 09:56.800 You know, that kind of gives you a clue 09:56.800 --> 09:58.360 as to which direction to go. 09:58.360 --> 09:59.680 There's also the idea somehow 09:59.680 --> 10:02.080 that I've been convinced of since I was an undergrad 10:02.080 --> 10:05.480 that even before that intelligence 10:05.480 --> 10:06.880 is inseparable from learning. 10:06.880 --> 10:10.040 So the idea somehow that you can create 10:10.040 --> 10:14.080 an intelligent machine by basically programming, 10:14.080 --> 10:17.440 for me was a non starter, you know, from the start. 10:17.440 --> 10:20.280 Every intelligent entity that we know about 10:20.280 --> 10:24.000 arrives at this intelligence through learning. 10:25.000 --> 10:26.280 So learning, you know, machine learning 10:26.280 --> 10:28.240 was a completely obvious path. 10:30.000 --> 10:30.960 Also because I'm lazy. 10:30.960 --> 10:32.440 So, you know, kind of. 10:32.440 --> 10:35.200 These automate basically everything 10:35.200 --> 10:37.920 and learning is the automation of intelligence. 10:37.920 --> 10:39.240 Right. 10:39.240 --> 10:43.000 So do you think, so what is learning then? 10:43.000 --> 10:44.600 What falls under learning? 10:44.600 --> 10:48.320 Because do you think of reasoning as learning? 10:48.320 --> 10:51.320 Well, reasoning is certainly a consequence 10:51.320 --> 10:53.320 of learning as well, 10:53.320 --> 10:56.320 just like other functions of the brain. 10:56.320 --> 10:58.320 The big question about reasoning is, 10:58.320 --> 11:00.320 how do you make reasoning compatible 11:00.320 --> 11:02.320 with gradient based learning? 11:02.320 --> 11:04.320 Do you think neural networks can be made to reason? 11:04.320 --> 11:06.320 Yes, there is no question about that. 11:06.320 --> 11:08.320 Again, we have a good example, right? 11:10.320 --> 11:11.320 The question is how? 11:11.320 --> 11:13.320 So the question is how much prior structure 11:13.320 --> 11:15.320 do you have to put in the neural net 11:15.320 --> 11:17.320 so that something like human reasoning 11:17.320 --> 11:21.320 will emerge from it, you know, from learning? 11:21.320 --> 11:24.320 Another question is all of our kind of model 11:24.320 --> 11:27.320 of what reasoning is that are based on logic 11:27.320 --> 11:31.320 are discrete and are therefore incompatible 11:31.320 --> 11:33.320 with gradient based learning. 11:33.320 --> 11:35.320 And I'm a very strong believer in this idea 11:35.320 --> 11:36.320 of gradient based learning. 11:36.320 --> 11:39.320 I don't believe that other types of learning 11:39.320 --> 11:41.320 that don't use kind of gradient information 11:41.320 --> 11:42.320 if you want. 11:42.320 --> 11:43.320 So you don't like discrete mathematics. 11:43.320 --> 11:45.320 You don't like anything discrete? 11:45.320 --> 11:47.320 Well, that's, it's not that I don't like it. 11:47.320 --> 11:49.320 It's just that it's incompatible with learning 11:49.320 --> 11:51.320 and I'm a big fan of learning, right? 11:51.320 --> 11:56.320 So in fact, that's perhaps one reason why deep learning 11:56.320 --> 11:58.320 has been kind of looked at with suspicion 11:58.320 --> 11:59.320 by a lot of computer scientists 11:59.320 --> 12:00.320 because the math is very different. 12:00.320 --> 12:02.320 The math that you use for deep learning, 12:02.320 --> 12:05.320 you know, it kind of has more to do with, you know, 12:05.320 --> 12:08.320 cybernetics, the kind of math you do 12:08.320 --> 12:09.320 in electrical engineering 12:09.320 --> 12:12.320 than the kind of math you do in computer science. 12:12.320 --> 12:16.320 And, you know, nothing in machine learning is exact, right? 12:16.320 --> 12:19.320 Computer science is all about sort of, you know, 12:19.320 --> 12:21.320 obsessive compulsive attention to details 12:21.320 --> 12:24.320 of like, you know, every index has to be right 12:24.320 --> 12:26.320 and you can prove that an algorithm is correct, right? 12:26.320 --> 12:31.320 Machine learning is the science of sloppiness, really. 12:31.320 --> 12:33.320 That's beautiful. 12:33.320 --> 12:38.320 So, okay, maybe let's feel around in the dark 12:38.320 --> 12:41.320 of what is a neural network that reasons 12:41.320 --> 12:46.320 or a system that works with continuous functions 12:47.320 --> 12:52.320 that's able to do, build knowledge. 12:52.320 --> 12:54.320 However we think about reasoning, 12:54.320 --> 12:57.320 build on previous knowledge, build on extra knowledge, 12:57.320 --> 13:00.320 create new knowledge, generalize outside 13:00.320 --> 13:04.320 of any training set ever built, what does that look like? 13:04.320 --> 13:08.320 If, yeah, maybe do you have inklings of thoughts 13:08.320 --> 13:10.320 of what that might look like? 13:10.320 --> 13:12.320 Yeah, I mean, yes and no. 13:12.320 --> 13:14.320 If I had precise ideas about this, 13:14.320 --> 13:16.320 I think, you know, we'll be building it right now. 13:16.320 --> 13:18.320 But, and there are people working on this 13:18.320 --> 13:22.320 whose main research interest is actually exactly that, right? 13:22.320 --> 13:25.320 So, what you need to have is a working memory. 13:25.320 --> 13:29.320 So, you need to have some device, if you want, 13:29.320 --> 13:34.320 some subsystem that can store a relatively large number 13:34.320 --> 13:38.320 of factual, episodic information for, you know, 13:38.320 --> 13:40.320 reasonable amount of time. 13:40.320 --> 13:43.320 So, you know, in the brain, for example, 13:43.320 --> 13:45.320 there are kind of three main types of memory. 13:45.320 --> 13:52.320 One is the sort of memory of the state of your cortex. 13:52.320 --> 13:55.320 And that sort of disappears within 20 seconds. 13:55.320 --> 13:57.320 You can't remember things for more than about 20 seconds 13:57.320 --> 14:01.320 or a minute if you don't have any other form of memory. 14:01.320 --> 14:04.320 The second type of memory, which is longer term, 14:04.320 --> 14:06.320 is short term, is the hippocampus. 14:06.320 --> 14:08.320 So, you can, you know, you came into this building, 14:08.320 --> 14:13.320 you remember where the exit is, where the elevators are. 14:13.320 --> 14:15.320 You have some map of that building 14:15.320 --> 14:17.320 that's stored in your hippocampus. 14:17.320 --> 14:20.320 You might remember something about what I said, 14:20.320 --> 14:21.320 you know, a few minutes ago. 14:21.320 --> 14:22.320 I forgot it already. 14:22.320 --> 14:23.320 Of course, it's been erased. 14:23.320 --> 14:27.320 But, you know, that would be in your hippocampus. 14:27.320 --> 14:30.320 And then the longer term memory is in the synapse. 14:30.320 --> 14:32.320 The synapses, right? 14:32.320 --> 14:34.320 So, what you need if you want a system 14:34.320 --> 14:36.320 that's capable of reasoning is that you want 14:36.320 --> 14:39.320 the hippocampus like thing, right? 14:39.320 --> 14:41.320 And that's what people have tried to do 14:41.320 --> 14:43.320 with memory networks and, you know, 14:43.320 --> 14:45.320 neural engineering machines and stuff like that, right? 14:45.320 --> 14:49.320 And now with transformers, which have sort of a memory 14:49.320 --> 14:51.320 in their kind of self attention system. 14:51.320 --> 14:53.320 You can think of it this way. 14:53.320 --> 14:56.320 So, that's one element you need. 14:56.320 --> 14:59.320 Another thing you need is some sort of network 14:59.320 --> 15:04.320 that can access this memory, 15:04.320 --> 15:07.320 get an information back and then kind of crunch on it 15:07.320 --> 15:10.320 and then do this iteratively multiple times 15:10.320 --> 15:15.320 because a chain of reasoning is a process 15:15.320 --> 15:19.320 by which you can update your knowledge 15:19.320 --> 15:20.320 about the state of the world, 15:20.320 --> 15:22.320 about, you know, what's going to happen, et cetera. 15:22.320 --> 15:26.320 And that has to be this sort of recurrent operation, basically. 15:26.320 --> 15:30.320 And you think that kind of, if we think about a transformer, 15:30.320 --> 15:33.320 so that seems to be too small to contain the knowledge 15:33.320 --> 15:37.320 that's to represent the knowledge that's contained 15:37.320 --> 15:38.320 in Wikipedia, for example. 15:38.320 --> 15:41.320 Well, a transformer doesn't have this idea of recurrence. 15:41.320 --> 15:42.320 It's got a fixed number of layers 15:42.320 --> 15:44.320 and that's the number of steps that, you know, 15:44.320 --> 15:46.320 limits basically as a representation. 15:46.320 --> 15:50.320 But recurrence would build on the knowledge somehow. 15:50.320 --> 15:54.320 I mean, it would evolve the knowledge 15:54.320 --> 15:57.320 and expand the amount of information, 15:57.320 --> 16:00.320 perhaps, or useful information within that knowledge. 16:00.320 --> 16:04.320 But is this something that just can emerge with size? 16:04.320 --> 16:06.320 Because it seems like everything we have now is too small. 16:06.320 --> 16:09.320 No, it's not clear. 16:09.320 --> 16:12.320 I mean, how you access and write into an associated memory 16:12.320 --> 16:13.320 in an efficient way. 16:13.320 --> 16:15.320 I mean, sort of the original memory network 16:15.320 --> 16:17.320 maybe had something like the right architecture, 16:17.320 --> 16:20.320 but if you try to scale up a memory network 16:20.320 --> 16:22.320 so that the memory contains all of Wikipedia, 16:22.320 --> 16:24.320 it doesn't quite work. 16:24.320 --> 16:27.320 So there's a need for new ideas there. 16:27.320 --> 16:29.320 But it's not the only form of reasoning. 16:29.320 --> 16:31.320 So there's another form of reasoning, 16:31.320 --> 16:36.320 which is very classical also in some types of AI, 16:36.320 --> 16:40.320 and it's based on, let's call it energy minimization. 16:40.320 --> 16:44.320 So you have some sort of objective, 16:44.320 --> 16:50.320 some energy function that represents the quality 16:50.320 --> 16:52.320 or the negative quality. 16:52.320 --> 16:54.320 Energy goes up when things get bad 16:54.320 --> 16:56.320 and they get low when things get good. 16:56.320 --> 17:00.320 So let's say you want to figure out what gestures 17:00.320 --> 17:07.320 do I need to do to grab an object or walk out the door. 17:07.320 --> 17:09.320 If you have a good model of your own body, 17:09.320 --> 17:11.320 a good model of the environment, 17:11.320 --> 17:13.320 using this kind of energy minimization, 17:13.320 --> 17:16.320 you can do planning. 17:16.320 --> 17:21.320 And it's in optimal control, it's called model predictive control. 17:21.320 --> 17:23.320 You have a model of what's going to happen in the world 17:23.320 --> 17:25.320 as a consequence of your actions. 17:25.320 --> 17:28.320 And that allows you to buy energy minimization, 17:28.320 --> 17:29.320 figure out a sequence of action 17:29.320 --> 17:31.320 that optimizes a particular objective function, 17:31.320 --> 17:34.320 which measures the number of times you're going to hit something 17:34.320 --> 17:39.320 and the energy you're going to spend doing the gesture and etc. 17:39.320 --> 17:42.320 So that's a form of reasoning. 17:42.320 --> 17:43.320 Planning is a form of reasoning. 17:43.320 --> 17:47.320 And perhaps what led to the ability of humans to reason 17:47.320 --> 17:53.320 is the fact that species that appear before us 17:53.320 --> 17:56.320 had to do some sort of planning to be able to hunt and survive 17:56.320 --> 17:59.320 and survive the winter in particular. 17:59.320 --> 18:03.320 And so it's the same capacity that you need to have. 18:03.320 --> 18:09.320 So in your intuition, if we look at expert systems, 18:09.320 --> 18:13.320 and encoding knowledge as logic systems, 18:13.320 --> 18:16.320 as graphs in this kind of way, 18:16.320 --> 18:20.320 is not a useful way to think about knowledge? 18:20.320 --> 18:24.320 Graphs are a little brittle or logic representation. 18:24.320 --> 18:28.320 So basically, variables that have values 18:28.320 --> 18:31.320 and then constrained between them that are represented by rules 18:31.320 --> 18:33.320 is a little too rigid and too brittle. 18:33.320 --> 18:38.320 So some of the early efforts in that respect 18:38.320 --> 18:41.320 were to put probabilities on them. 18:41.320 --> 18:44.320 So a rule, if you have this and that symptom, 18:44.320 --> 18:47.320 you have this disease with that probability 18:47.320 --> 18:50.320 and you should prescribe that antibiotic with that probability. 18:50.320 --> 18:54.320 That's the mysine system from the 70s. 18:54.320 --> 18:59.320 And that branch of AI led to business networks 18:59.320 --> 19:02.320 and graphical models and causal inference 19:02.320 --> 19:05.320 and variational method. 19:05.320 --> 19:10.320 So there is certainly a lot of interesting work going on 19:10.320 --> 19:11.320 in this area. 19:11.320 --> 19:13.320 The main issue with this is knowledge acquisition. 19:13.320 --> 19:19.320 How do you reduce a bunch of data to a graph of this type? 19:19.320 --> 19:23.320 It relies on the expert on the human being to encode, 19:23.320 --> 19:24.320 to add knowledge. 19:24.320 --> 19:27.320 And that's essentially impractical. 19:27.320 --> 19:29.320 So that's a big question. 19:29.320 --> 19:32.320 The second question is, do you want to represent knowledge 19:32.320 --> 19:36.320 as symbols and do you want to manipulate them with logic? 19:36.320 --> 19:38.320 And again, that's incompatible with learning. 19:38.320 --> 19:42.320 So one suggestion with Jeff Hinton 19:42.320 --> 19:44.320 has been advocating for many decades 19:44.320 --> 19:48.320 is replace symbols by vectors. 19:48.320 --> 19:50.320 Think of it as pattern of activities 19:50.320 --> 19:54.320 in a bunch of neurons or units or whatever you want to call them. 19:54.320 --> 19:58.320 And replace logic by continuous functions. 19:58.320 --> 20:01.320 And that becomes now compatible. 20:01.320 --> 20:04.320 There's a very good set of ideas 20:04.320 --> 20:07.320 written in a paper about 10 years ago 20:07.320 --> 20:12.320 by Leon Botou who is here at Facebook. 20:12.320 --> 20:14.320 The title of the paper is 20:14.320 --> 20:15.320 From Machine Learning to Machine Reasoning. 20:15.320 --> 20:19.320 And his idea is that a learning system 20:19.320 --> 20:22.320 should be able to manipulate objects that are in a space 20:22.320 --> 20:24.320 and then put the result back in the same space. 20:24.320 --> 20:27.320 So it's this idea of working memory basically. 20:27.320 --> 20:30.320 And it's very enlightening. 20:30.320 --> 20:33.320 And in a sense, that might learn something 20:33.320 --> 20:37.320 like the simple expert systems. 20:37.320 --> 20:41.320 I mean, you can learn basic logic operations there. 20:41.320 --> 20:43.320 Yeah, quite possibly. 20:43.320 --> 20:46.320 There's a big debate on how much prior structure 20:46.320 --> 20:48.320 you have to put in for this kind of stuff to emerge. 20:48.320 --> 20:51.320 That's the debate I have with Gary Marcus and people like that. 20:51.320 --> 20:54.320 Yeah, so and the other person, 20:54.320 --> 20:57.320 so I just talked to Judea Pearl 20:57.320 --> 21:00.320 and he mentioned causal inference world. 21:00.320 --> 21:04.320 So his worry is that the current neural networks 21:04.320 --> 21:09.320 are not able to learn what causes 21:09.320 --> 21:12.320 what causal inference between things. 21:12.320 --> 21:15.320 So I think he's right and wrong about this. 21:15.320 --> 21:21.320 If he's talking about the sort of classic type of neural nets, 21:21.320 --> 21:23.320 people sort of didn't worry too much about this. 21:23.320 --> 21:26.320 But there's a lot of people now working on causal inference. 21:26.320 --> 21:28.320 There's a paper that just came out last week 21:28.320 --> 21:29.320 by Leon Boutou, among others, 21:29.320 --> 21:32.320 the Vila Pespas and a bunch of other people. 21:32.320 --> 21:36.320 Exactly on that problem of how do you kind of, 21:36.320 --> 21:39.320 you know, get a neural net to sort of pay attention 21:39.320 --> 21:41.320 to real causal relationships, 21:41.320 --> 21:46.320 which may also solve issues of bias in data 21:46.320 --> 21:48.320 and things like this. 21:48.320 --> 21:51.320 I'd like to read that paper because that ultimately 21:51.320 --> 21:56.320 challenges also seems to fall back on the human expert 21:56.320 --> 22:01.320 to ultimately decide causality between things. 22:01.320 --> 22:04.320 People are not very good at establishing causality, first of all. 22:04.320 --> 22:06.320 So first of all, you talk to physicists 22:06.320 --> 22:08.320 and physicists actually don't believe in causality 22:08.320 --> 22:12.320 because look at all the basic laws of macro physics 22:12.320 --> 22:15.320 are time reversible, so there's no causality. 22:15.320 --> 22:17.320 The era of time is not real. 22:17.320 --> 22:20.320 It's as soon as you start looking at macroscopic systems 22:20.320 --> 22:22.320 where there is unpredictable randomness 22:22.320 --> 22:25.320 where there is clearly an hour of time, 22:25.320 --> 22:28.320 but it's a big mystery in physics, actually, how that emerges. 22:28.320 --> 22:34.320 Is it emergent or is it part of the fundamental fabric of reality? 22:34.320 --> 22:36.320 Or is it a bias of intelligent systems 22:36.320 --> 22:39.320 that, you know, because of the second law of thermodynamics, 22:39.320 --> 22:41.320 we perceive a particular hour of time, 22:41.320 --> 22:44.320 but in fact, it's kind of arbitrary, right? 22:44.320 --> 22:47.320 So yeah, physicists, mathematicians, they don't care about, 22:47.320 --> 22:51.320 I mean, the math doesn't care about the flow of time. 22:51.320 --> 22:53.320 Well, certainly macro physics doesn't. 22:53.320 --> 22:58.320 People themselves are not very good at establishing causal relationships. 22:58.320 --> 23:02.320 If you ask, I think it was in one of Seymour Papert's book 23:02.320 --> 23:06.320 on, like, children learning. 23:06.320 --> 23:08.320 You know, he studied with Jean Piaget. 23:08.320 --> 23:12.320 He's the guy who coauthored the book Perception with Marvin Minsky 23:12.320 --> 23:14.320 that kind of killed the first wave of neural nets. 23:14.320 --> 23:17.320 But he was actually a learning person. 23:17.320 --> 23:22.320 He, in the sense of studying learning in humans and machines. 23:22.320 --> 23:24.320 That's why he got interested in Perceptron. 23:24.320 --> 23:33.320 And he wrote that if you ask a little kid about what is the cause of the wind, 23:33.320 --> 23:36.320 a lot of kids will say, they will think for a while and they will say, 23:36.320 --> 23:38.320 oh, it's the branches in the trees. 23:38.320 --> 23:40.320 They move and that creates wind, right? 23:40.320 --> 23:42.320 So they get the causal relationship backwards. 23:42.320 --> 23:45.320 And it's because they're understanding of the world and intuitive physics. 23:45.320 --> 23:46.320 It's not that great, right? 23:46.320 --> 23:49.320 I mean, these are like, you know, four or five year old kids. 23:49.320 --> 23:53.320 You know, it gets better and then you understand that this, it can be, right? 23:53.320 --> 24:00.320 But there are many things which we can, because of our common sense understanding of things, 24:00.320 --> 24:02.320 what people call common sense. 24:02.320 --> 24:03.320 Yeah. 24:03.320 --> 24:05.320 And we're understanding of physics. 24:05.320 --> 24:09.320 We can, there's a lot of stuff that we can figure out causality, even with diseases. 24:09.320 --> 24:13.320 We can figure out what's not causing what often. 24:13.320 --> 24:19.320 There's a lot of mystery, of course, but the idea is that you should be able to encode that into systems. 24:19.320 --> 24:22.320 Because it seems unlikely they'd be able to figure that out themselves. 24:22.320 --> 24:26.320 Well, whenever we can do intervention, but you know, all of humanity has been completely deluded 24:26.320 --> 24:32.320 for millennia, probably since existence, about a very, very wrong causal relationship 24:32.320 --> 24:38.320 where whatever you can explain, you're attributed to, you know, some deity, some divinity, right? 24:38.320 --> 24:40.320 And that's a cup out. 24:40.320 --> 24:42.320 That's a way of saying like, I don't know the cause. 24:42.320 --> 24:44.320 So, you know, God did it, right? 24:44.320 --> 24:54.320 So you mentioned Marvin Minsky and the irony of, you know, maybe causing the first day I winter. 24:54.320 --> 24:56.320 You were there in the 90s. 24:56.320 --> 24:58.320 You were there in the 80s, of course. 24:58.320 --> 25:02.320 In the 90s, what do you think people lost faith in deep learning in the 90s 25:02.320 --> 25:06.320 and found it again a decade later, over a decade later? 25:06.320 --> 25:07.320 Yeah. 25:07.320 --> 25:09.320 Deep learning, yeah, it was just called neural nets. 25:09.320 --> 25:11.320 You know, that works. 25:11.320 --> 25:13.320 Yeah, they lost interest. 25:13.320 --> 25:18.320 I mean, I think I would put that around 1995, at least the machine learning community. 25:18.320 --> 25:28.320 There was always a neural net community, but it became kind of disconnected from sort of mainstream machine learning if you want. 25:28.320 --> 25:32.320 There were, it was basically electrical engineering that kept at it. 25:32.320 --> 25:33.320 Right. 25:33.320 --> 25:35.320 And computer science. 25:35.320 --> 25:36.320 Just gave up. 25:36.320 --> 25:37.320 Neural nets. 25:37.320 --> 25:39.320 I don't, I don't know. 25:39.320 --> 25:47.320 You know, I was too close to it to really sort of analyze it with sort of a unbiased eye if you want. 25:47.320 --> 25:50.320 But I would, I would, I would make a few guesses. 25:50.320 --> 26:03.320 So the first one is at the time neural nets were, it was very hard to make them work in a sense that you would, you know, implement backprop in your favorite language. 26:03.320 --> 26:06.320 And that favorite language was not Python. 26:06.320 --> 26:07.320 It was not MATLAB. 26:07.320 --> 26:10.320 It was not any of those things because they didn't exist. 26:10.320 --> 26:11.320 Right. 26:11.320 --> 26:14.320 You had to write it in Fortran or C or something like this. 26:14.320 --> 26:15.320 Right. 26:15.320 --> 26:18.320 So you would experiment with it. 26:18.320 --> 26:26.320 You would probably make some very basic mistakes, like, you know, badly initialize your weights, make the network too small because you're already in the textbook, you know, you don't want too many parameters. 26:26.320 --> 26:27.320 Right. 26:27.320 --> 26:31.320 And of course, you know, and you would train on XOR because you didn't have any other data set to trade on. 26:31.320 --> 26:33.320 And of course, you know, it works half the time. 26:33.320 --> 26:35.320 So you would say, I give up. 26:35.320 --> 26:39.320 Also, you would train it with batch gradient, which, you know, isn't that sufficient. 26:39.320 --> 26:46.320 So there was a lot of bad good tricks that you had to know to make those things work or you had to reinvent. 26:46.320 --> 26:50.320 And a lot of people just didn't and they just couldn't make it work. 26:50.320 --> 26:52.320 So that's one thing. 26:52.320 --> 27:08.320 The investment in software platform to be able to kind of, you know, display things, figure out why things don't work, kind of get a good intuition for how to get them to work, have enough flexibility so you can create, you know, network architectures like convolutional nets and stuff like that. 27:08.320 --> 27:09.320 It was hard. 27:09.320 --> 27:10.320 I mean, you had to write everything from scratch. 27:10.320 --> 27:13.320 And again, you didn't have any Python or MATLAB or anything. 27:13.320 --> 27:14.320 Right. 27:14.320 --> 27:25.320 I read that, sorry to interrupt, but I read that you wrote in Lisp the, your first versions of Lynette with the convolutional networks, which by the way, one of my favorite languages. 27:25.320 --> 27:27.320 That's how I knew you were legit. 27:27.320 --> 27:29.320 Touring award, whatever. 27:29.320 --> 27:31.320 You programmed in Lisp. 27:31.320 --> 27:32.320 It's still my favorite language. 27:32.320 --> 27:35.320 But it's not that we programmed in Lisp. 27:35.320 --> 27:37.320 It's that we had to write a Lisp interpreter. 27:37.320 --> 27:38.320 Okay. 27:38.320 --> 27:40.320 Because it's not like we use one that existed. 27:40.320 --> 27:48.320 So we wrote a Lisp interpreter that we hooked up to, you know, a back end library that we wrote also for sort of neural net computation. 27:48.320 --> 28:01.320 And then after a few years around 1991, we invented this idea of basically having modules that know how to forward propagate and back propagate gradients and then interconnecting those modules in a graph. 28:01.320 --> 28:11.320 Leon but who had made proposals on this about this in the late 80s, and we're able to implement this using a list system. 28:11.320 --> 28:14.320 Eventually, we wanted to use that system to make build production code for character recognition at Bell Labs. 28:14.320 --> 28:22.320 So we actually wrote a compiler for that Lisp interpreter so that Petris Seymard, who is now Microsoft, kind of did the bulk of it with Leon and me. 28:22.320 --> 28:33.320 And so we could write our system in Lisp and then compile to C and then we'll have a self contain complete system that could kind of do the entire thing. 28:33.320 --> 28:36.320 Neither PyTorch nor Transparency can do this today. 28:36.320 --> 28:37.320 Yeah. 28:37.320 --> 28:38.320 Okay. 28:38.320 --> 28:39.320 It's coming. 28:39.320 --> 28:40.320 Yeah. 28:40.320 --> 28:44.320 I mean, there's something like that in PyTorch called, you know, Torch script. 28:44.320 --> 28:50.320 And so, you know, we had to write a Lisp interpreter, we had to write a Lisp compiler, we had to invest a huge amount of effort to do this. 28:50.320 --> 28:56.320 And not everybody, if you don't completely believe in the concept, you're not going to invest the time to do this. 28:56.320 --> 28:57.320 Right. 28:57.320 --> 29:03.320 Now, at the time also, you know, or today, this would turn into Torch or PyTorch or Transparency or whatever. 29:03.320 --> 29:07.320 We'd put it in open source, everybody would use it and, you know, realize it's good. 29:07.320 --> 29:17.320 Back before 1995, working at AT&T, there's no way the lawyers would let you release anything in open source of this nature. 29:17.320 --> 29:20.320 And so we could not distribute our code, really. 29:20.320 --> 29:29.320 And on that point, and sorry to go on a million tangents, but on that point, I also read that there was some almost pat, like a patent on convolutional networks. 29:29.320 --> 29:31.320 Yes, there was. 29:31.320 --> 29:35.320 So that, first of all, I mean, just. 29:35.320 --> 29:37.320 There were two, actually. 29:37.320 --> 29:39.320 That ran out. 29:39.320 --> 29:41.320 Thankfully, in 2007. 29:41.320 --> 29:44.320 In 2007. 29:44.320 --> 29:48.320 What, can we, can we just talk about that first? 29:48.320 --> 29:50.320 I know you're a Facebook, but you're also an NYU. 29:50.320 --> 29:58.320 And what does it mean to patent ideas like these software ideas, essentially? 29:58.320 --> 30:01.320 Or what are mathematical ideas? 30:01.320 --> 30:03.320 Or what are they? 30:03.320 --> 30:04.320 Okay. 30:04.320 --> 30:05.320 So they're not mathematical ideas. 30:05.320 --> 30:07.320 So there are, you know, algorithms. 30:07.320 --> 30:15.320 And there was a period where the US patent office would allow the patent of software as long as it was embodied. 30:15.320 --> 30:18.320 The Europeans are very different. 30:18.320 --> 30:21.320 They don't, they don't quite accept that they have a different concept. 30:21.320 --> 30:28.320 But, you know, I don't, I no longer, I mean, I never actually strongly believed in this, but I don't believe in this kind of patent. 30:28.320 --> 30:33.320 Facebook basically doesn't believe in this kind of patent. 30:33.320 --> 30:39.320 Google files patents because they've been burned with Apple. 30:39.320 --> 30:41.320 And so now they do this for defensive purpose. 30:41.320 --> 30:44.320 But usually they say, we're not going to see you if you're in French. 30:44.320 --> 30:47.320 Facebook has a, has a similar policy. 30:47.320 --> 30:50.320 They say, you know, we have a patent on certain things for defensive purpose. 30:50.320 --> 30:54.320 We're not going to see you if you're in French unless you through us. 30:54.320 --> 30:59.320 So the, the industry does not believe in, in patents. 30:59.320 --> 31:03.320 They're there because of, you know, the legal landscape and, and, and various things. 31:03.320 --> 31:07.320 But, but I don't really believe in patents for this kind of stuff. 31:07.320 --> 31:09.320 Okay. So that's, that's a great thing. 31:09.320 --> 31:11.320 So I, I tell you a worst story. 31:11.320 --> 31:12.320 Yeah. 31:12.320 --> 31:19.320 So what happens was the first, the first patent about convolutional net was about kind of the early version of convolutional net that didn't have separate pooling layers. 31:19.320 --> 31:24.320 It had, you know, convolutional layers with tried more than one, if you want, right? 31:24.320 --> 31:31.320 And then there was a second one on convolutional nets with separate pooling layers, trained with backprop. 31:31.320 --> 31:35.320 And there were files filed in 89 and 1990 or something like this. 31:35.320 --> 31:39.320 At the time, the life, life of a patent was 17 years. 31:39.320 --> 31:48.320 So here's what happened over the next few years is that we started developing character recognition technology around convolutional nets. 31:48.320 --> 31:55.320 And in 1994, a check reading system was deployed in ATM machines. 31:55.320 --> 32:00.320 In 1995, it was for large check reading machines in back offices, et cetera. 32:00.320 --> 32:08.320 And those systems were developed by an engineering group that we were collaborating with AT&T and they were commercialized by NCR, 32:08.320 --> 32:11.320 which at the time was a subsidiary of AT&T. 32:11.320 --> 32:18.320 Now AT&T split up in 1996, early 1996. 32:18.320 --> 32:22.320 And the lawyers just looked at all the patents and they distributed the patents among the various companies. 32:22.320 --> 32:28.320 They gave the convolutional net patent to NCR because they were actually selling products that used it. 32:28.320 --> 32:31.320 But nobody at NCR had any idea what a convolutional net was. 32:31.320 --> 32:32.320 Yeah. 32:32.320 --> 32:33.320 Okay. 32:33.320 --> 32:40.320 So between 1996 and 2007, there's a whole period until 2002 where I didn't actually work on 32:40.320 --> 32:42.320 machine learning or convolutional net. 32:42.320 --> 32:45.320 I resumed working on this around 2002. 32:45.320 --> 32:51.320 And between 2002 and 2007, I was working on them crossing my finger that nobody at NCR would notice and nobody noticed. 32:51.320 --> 32:52.320 Yeah. 32:52.320 --> 33:02.320 And I hope that this kind of somewhat, as you said, lawyers aside, relative openness of the community now will continue. 33:02.320 --> 33:05.320 It accelerates the entire progress of the industry. 33:05.320 --> 33:17.320 And the problems that Facebook and Google and others are facing today is not whether Facebook or Google or Microsoft or IBM or whoever is ahead of the other. 33:17.320 --> 33:20.320 It's that we don't have the technology to build these things we want to build. 33:20.320 --> 33:24.320 We want to build intelligent virtual assistants that have common sense. 33:24.320 --> 33:26.320 We don't have monopoly on good ideas for this. 33:26.320 --> 33:27.320 We don't believe we do. 33:27.320 --> 33:30.320 Maybe others do believe they do, but we don't. 33:30.320 --> 33:31.320 Okay. 33:31.320 --> 33:37.320 If a startup tells you they have a secret to human level intelligence and common sense, don't believe them. 33:37.320 --> 33:38.320 They don't. 33:38.320 --> 33:50.320 And it's going to take the entire work of the world research community for a while to get to the point where you can go off and in each of those companies can start to build things on this. 33:50.320 --> 33:51.320 We're not there yet. 33:51.320 --> 33:52.320 Absolutely. 33:52.320 --> 34:03.320 And this calls to the gap between the space of ideas and the rigorous testing of those ideas of practical application that you often speak to. 34:03.320 --> 34:17.320 You've written advice saying, don't get fooled by people who claim to have a solution to artificial general intelligence who claim to have an AI system that works just like the human brain or who claim to have figured out how the brain works. 34:17.320 --> 34:23.320 That's them, what the error rate they get on MNIST or ImageNet. 34:23.320 --> 34:25.320 This is a little dated, by the way. 34:25.320 --> 34:26.320 $2,000. 34:26.320 --> 34:27.320 I mean, five years. 34:27.320 --> 34:28.320 Who's counting? 34:28.320 --> 34:29.320 Okay. 34:29.320 --> 34:33.320 But I think your opinion is the MNIST and ImageNet. 34:33.320 --> 34:35.320 Yes, maybe dated. 34:35.320 --> 34:36.320 There may be new benchmarks, right? 34:36.320 --> 34:47.320 But I think that philosophy is one you still in somewhat hold that benchmarks and the practical testing, the practical application is where you really get to test the ideas. 34:47.320 --> 34:49.320 Well, it may not be completely practical. 34:49.320 --> 35:00.320 Like, for example, you know, it could be a toy data set, but it has to be some sort of task that the community as a whole is accepted as some sort of standard, you know, kind of benchmark if you want. 35:00.320 --> 35:01.320 It doesn't need to be real. 35:01.320 --> 35:17.320 So for example, many years ago here at FAIR, people, you know, Cheson West and Antoine Bourne and a few others proposed the baby tasks, which were kind of a toy problem to test the ability of machines to reason actually to access working memory and things like this. 35:17.320 --> 35:20.320 And it was very useful, even though it wasn't a real task. 35:20.320 --> 35:23.320 MNIST is kind of halfway a real task. 35:23.320 --> 35:26.320 So, you know, toy problems can be very useful. 35:26.320 --> 35:39.320 I guess that I was really struck by the fact that a lot of people, particularly a lot of people with money to invest would be fooled by people telling them, oh, we have, you know, the algorithm of the cortex and you should give us 50 million. 35:39.320 --> 35:40.320 Yes, absolutely. 35:40.320 --> 35:48.320 So there's a lot of people who who try to take advantage of the hype for business reasons and so on. 35:48.320 --> 36:00.320 But let me sort of talk to this idea that new ideas, the ideas that push the field forward may not yet have a benchmark or it may be very difficult to establish a benchmark. 36:00.320 --> 36:01.320 I agree. 36:01.320 --> 36:02.320 That's part of the process. 36:02.320 --> 36:04.320 Establishing benchmarks is part of the process. 36:04.320 --> 36:18.320 So what are your thoughts about, so we have these benchmarks on around stuff we can do with images from classification to captioning to just every kind of information you can pull off from images and the surface level. 36:18.320 --> 36:20.320 There's audio data set. 36:20.320 --> 36:22.320 There's some video. 36:22.320 --> 36:25.320 What can we start natural language? 36:25.320 --> 36:41.320 What kind of stuff, what kind of benchmarks do you see that start creeping on to more something like intelligence, like reasoning, like maybe you don't like the term but AGI echoes of that kind of formulation. 36:41.320 --> 36:48.320 A lot of people are working on interactive environments in which you can you can train and test intelligence systems. 36:48.320 --> 37:02.320 So there, for example, you know, it's the classical paradigm of supervised running is that you have a data set, you partition it into a training set, validation set, test set, and there's a clear protocol, right. 37:02.320 --> 37:13.320 But what if the that assumes that the samples are statistically independent, you can exchange them, the order in which you see them doesn't shouldn't matter, you know, things like that. 37:13.320 --> 37:23.320 But what if the answer you give determines the next sample you see, which is the case, for example, in robotics, right, you robot does something and then it gets exposed to a new room. 37:23.320 --> 37:26.320 And depending on where it goes, the room would be different. 37:26.320 --> 37:30.320 So that's the that creates the exploration problem. 37:30.320 --> 37:44.320 What if the samples, so that creates also a dependency between samples, right, you, you, if you move, if you can only move in space, the next sample you're going to see is going to be probably in the same building, most likely. 37:44.320 --> 37:56.320 So, so, so the all the assumptions about the validity of this training set, test set hypothesis break, whenever machine can take an action that has an influence in the in the world, and it's what is going to see. 37:56.320 --> 38:08.320 So people are setting up artificial environments where where that takes place, right, the robot runs around a 3D model of a house and can interact with objects and things like this. 38:08.320 --> 38:20.320 So you do robotics by simulation, you have those, you know, opening a gym type thing or Mujoko kind of simulated robots and you have games, you know, things like that. 38:20.320 --> 38:25.320 So that that's where the field is going really this kind of environment. 38:25.320 --> 38:35.320 Now, back to the question of AGI, like, I don't like the term AGI, because it implies that human intelligence is general. 38:35.320 --> 38:40.320 And human intelligence is nothing like general, it's very, very specialized. 38:40.320 --> 38:45.320 We think is general, we'd like to think of ourselves as having general intelligence, we don't, we're very specialized. 38:45.320 --> 38:47.320 We're only slightly more general. 38:47.320 --> 38:48.320 Why does it feel general? 38:48.320 --> 38:51.320 So you kind of the term general. 38:51.320 --> 39:07.320 I think what's impressive about humans is ability to learn, as we were talking about learning, to learn in just so many different domains is perhaps not arbitrarily general, but just you can learn in many domains and integrate that knowledge somehow. 39:07.320 --> 39:08.320 Okay. 39:08.320 --> 39:09.320 The knowledge persists. 39:09.320 --> 39:11.320 So let me take a very specific example. 39:11.320 --> 39:12.320 Yes. 39:12.320 --> 39:13.320 It's not an example. 39:13.320 --> 39:16.320 It's more like a quasi mathematical demonstration. 39:16.320 --> 39:22.320 So you have about one million fibers coming out of one of your eyes, okay, two million total, but let's, let's talk about just one of them. 39:22.320 --> 39:26.320 It's one million nerve fibers, your optical nerve. 39:26.320 --> 39:30.320 Let's imagine that they are binary, so they can be active or inactive, right? 39:30.320 --> 39:36.320 So the input to your visual cortex is one million bits. 39:36.320 --> 39:47.320 Now they're connected to your brain in a particular way and your brain has connections that are kind of a little bit like a convolution that they kind of local, you know, in space and things like this. 39:47.320 --> 39:50.320 Now imagine I play a trick on you. 39:50.320 --> 39:52.320 It's a pretty nasty trick, I admit. 39:52.320 --> 40:00.320 I cut your optical nerve and I put a device that makes a random perturbation of a permutation of all the nerve fibers. 40:00.320 --> 40:08.320 So now what comes to your brain is a fixed but random permutation of all the pixels. 40:08.320 --> 40:19.320 There's no way in hell that your visual cortex, even if I do this to you in infancy, will actually learn vision to the same level of quality that you can. 40:19.320 --> 40:20.320 Got it. 40:20.320 --> 40:22.320 And you're saying there's no way you've learned that? 40:22.320 --> 40:28.320 No, because now two pixels that are nearby in the world will end up in very different places in your visual cortex. 40:28.320 --> 40:33.320 And your neurons there have no connections with each other because they only connect it locally. 40:33.320 --> 40:38.320 So this whole, our entire, the hardware is built in many ways to support. 40:38.320 --> 40:39.320 The locality of the real world? 40:39.320 --> 40:40.320 Yeah. 40:40.320 --> 40:41.320 Yes. 40:41.320 --> 40:42.320 That's specialization. 40:42.320 --> 40:44.320 Yeah, but it's still pretty damn impressive. 40:44.320 --> 40:46.320 So it's not perfect generalization. 40:46.320 --> 40:47.320 It's not even close. 40:47.320 --> 40:48.320 No, no. 40:48.320 --> 40:50.320 It's not that it's not even close. 40:50.320 --> 40:51.320 It's not at all. 40:51.320 --> 40:52.320 Yeah, it's not. 40:52.320 --> 40:54.320 So how many Boolean functions? 40:54.320 --> 41:03.320 Let's imagine you want to train your visual system to recognize particular patterns of those one million bits. 41:03.320 --> 41:05.320 So that's a Boolean function. 41:05.320 --> 41:07.320 Either the pattern is here or not here. 41:07.320 --> 41:13.320 It's a two way classification with one million binary inputs. 41:13.320 --> 41:16.320 How many such Boolean functions are there? 41:16.320 --> 41:21.320 You have two to the one million combinations of inputs. 41:21.320 --> 41:24.320 For each of those, you have an output bit. 41:24.320 --> 41:29.320 And so you have two to the two to the one million Boolean functions of this type. 41:29.320 --> 41:30.320 Okay. 41:30.320 --> 41:33.320 Which is an unimaginably large number. 41:33.320 --> 41:37.320 How many of those functions can actually be computed by your visual cortex? 41:37.320 --> 41:41.320 And the answer is a tiny, tiny, tiny, tiny, tiny, tiny sliver. 41:41.320 --> 41:43.320 Like an enormously tiny sliver. 41:43.320 --> 41:44.320 Yeah. 41:44.320 --> 41:45.320 Yeah. 41:45.320 --> 41:48.320 So we are ridiculously specialized. 41:48.320 --> 41:51.320 Okay. 41:51.320 --> 41:54.320 That's an argument against the word general. 41:54.320 --> 42:09.320 I agree with your intuition, but I'm not sure it seems the brain is impressively capable of adjusting to things. 42:09.320 --> 42:16.320 It's because we can't imagine tasks that are outside of our comprehension. 42:16.320 --> 42:20.320 So we think we are general because we're general of all the things that we can apprehend. 42:20.320 --> 42:21.320 So yeah. 42:21.320 --> 42:24.320 But there is a huge world out there of things that we have no idea. 42:24.320 --> 42:26.320 We call that heat, by the way. 42:26.320 --> 42:27.320 Heat. 42:27.320 --> 42:28.320 Heat. 42:28.320 --> 42:33.320 So at least physicists call that heat or they call it entropy, which is kind of... 42:33.320 --> 42:39.320 You have a thing full of gas, right? 42:39.320 --> 42:40.320 Close system for gas. 42:40.320 --> 42:41.320 Right? 42:41.320 --> 42:42.320 Close or no close. 42:42.320 --> 42:51.320 It has, you know, pressure, it has temperature, it has, you know, and you can write equations, 42:51.320 --> 42:55.320 PV equal and RT, you know, things like that, right? 42:55.320 --> 43:00.320 When you reduce the volume, the temperature goes up, the pressure goes up, you know, things like that, right? 43:00.320 --> 43:02.320 For perfect gas, at least. 43:02.320 --> 43:05.320 Those are the things you can know about that system. 43:05.320 --> 43:10.320 And it's a tiny, tiny number of bits compared to the complete information of the state of the entire system. 43:10.320 --> 43:17.320 Because the state of the entire system will give you the position and momentum of every molecule of the gas. 43:17.320 --> 43:23.320 And what you don't know about it is the entropy and you interpret it as heat. 43:23.320 --> 43:27.320 The energy contained in that thing is what we call heat. 43:27.320 --> 43:34.320 Now, it's very possible that, in fact, there is some very strong structure in how those molecules are moving. 43:34.320 --> 43:38.320 It's just that they are in a way that we are just not wired to perceive. 43:38.320 --> 43:39.320 Yeah, we're ignorant of it. 43:39.320 --> 43:44.320 And there's, in your infinite amount of things, we're not wired to perceive. 43:44.320 --> 43:45.320 Yeah. 43:45.320 --> 43:47.320 And you're right, that's a nice way to put it. 43:47.320 --> 43:54.320 We're general to all the things we can imagine, which is a very tiny subset of all the things that are possible. 43:54.320 --> 43:58.320 So it's like comagraph complexity or the comagraph's chat in some kind of complexity. 43:58.320 --> 43:59.320 Yeah. 43:59.320 --> 44:07.320 You know, every bit string or every integer is random, except for all the ones that you can actually write down. 44:07.320 --> 44:15.320 Yeah, okay, so beautiful, but, you know, so we can just call it artificial intelligence. 44:15.320 --> 44:17.320 We don't need to have a general. 44:17.320 --> 44:18.320 Or human level. 44:18.320 --> 44:20.320 Human level intelligence is good. 44:20.320 --> 44:33.320 You know, you'll start, anytime you touch human, it gets interesting because, you know, it's because we attach ourselves to human 44:33.320 --> 44:36.320 and it's difficult to define what human intelligence is. 44:36.320 --> 44:42.320 Nevertheless, my definition is maybe a damn impressive intelligence. 44:42.320 --> 44:46.320 Okay, damn impressive demonstration of intelligence, whatever. 44:46.320 --> 44:53.320 And so on that topic, most successes in deep learning have been in supervised learning. 44:53.320 --> 44:57.320 What is your view on unsupervised learning? 44:57.320 --> 45:07.320 Is there a hope to reduce involvement of human input and still have successful systems that have practically use? 45:07.320 --> 45:09.320 Yeah, I mean, there's definitely a hope. 45:09.320 --> 45:11.320 It's more than a hope, actually. 45:11.320 --> 45:13.320 It's, you know, mounting evidence for it. 45:13.320 --> 45:15.320 And that's basically all I do. 45:15.320 --> 45:20.320 Like the only thing I'm interested in at the moment is I call itself supervised learning, not unsupervised. 45:20.320 --> 45:25.320 Because unsupervised learning is a loaded term. 45:25.320 --> 45:31.320 People who know something about machine learning, you know, tell you, so you're doing clustering or PCA, which is not the case. 45:31.320 --> 45:37.320 And the white public, you know, when you say unsupervised learning, oh my God, you know, machines are going to learn by themselves and without supervision. 45:37.320 --> 45:39.320 You know, they see this as... 45:39.320 --> 45:41.320 Where's the parents? 45:41.320 --> 45:49.320 Yeah, so I call myself supervised learning because, in fact, the underlying algorithms that are used are the same algorithms as the supervised learning algorithms. 45:49.320 --> 45:59.320 Except that what we try them to do is not predict a particular set of variables, like the category of an image. 45:59.320 --> 46:05.320 And not to predict a set of variables that have been provided by human labelers. 46:05.320 --> 46:11.320 But what you're trying the machine to do is basically reconstruct a piece of its input that it's being... 46:11.320 --> 46:15.320 It's being masked out, essentially. You can think of it this way, right? 46:15.320 --> 46:20.320 So show a piece of video to a machine and ask it to predict what's going to happen next. 46:20.320 --> 46:28.320 And of course, after a while, you can show what happens and the machine will kind of train itself to do better at that task. 46:28.320 --> 46:35.320 You can do, like all the latest, most successful models in natural language processing, use self supervised learning. 46:35.320 --> 46:38.320 You know, sort of bird style systems, for example, right? 46:38.320 --> 46:43.320 You show it a window of a dozen words on a text corpus. 46:43.320 --> 46:51.320 You take out 15% of the words and then you train the machine to predict the words that are missing. 46:51.320 --> 46:56.320 That's self supervised learning. It's not predicting the future, it's just predicting things in the middle. 46:56.320 --> 46:59.320 But you could have it predict the future. That's what language models do. 46:59.320 --> 47:05.320 So in an unsupervised way, you construct a model of language. Do you think... 47:05.320 --> 47:09.320 Or video or the physical world or whatever, right? 47:09.320 --> 47:12.320 How far do you think that can take us? 47:12.320 --> 47:17.320 Do you think very far it understands anything? 47:17.320 --> 47:23.320 To some level, it has, you know, a shadow understanding of text. 47:23.320 --> 47:26.320 But it needs to, I mean, to have kind of true human level intelligence. 47:26.320 --> 47:29.320 I think you need to ground language in reality. 47:29.320 --> 47:32.320 So some people are attempting to do this, right? 47:32.320 --> 47:37.320 Having systems that kind of have some visual representation of what is being talked about. 47:37.320 --> 47:40.320 Which is one reason you need those interactive environments, actually. 47:40.320 --> 47:44.320 But this is like a huge technical problem that is not solved. 47:44.320 --> 47:49.320 And that explains why self supervised learning works in the context of natural language. 47:49.320 --> 47:55.320 That does not work in the context, or at least not well, in the context of image recognition and video. 47:55.320 --> 47:57.320 Although it's making progress quickly. 47:57.320 --> 48:04.320 And the reason, that reason is the fact that it's much easier to represent uncertainty in the prediction. 48:04.320 --> 48:09.320 In the context of natural language than it is in the context of things like video and images. 48:09.320 --> 48:17.320 So for example, if I ask you to predict what words I'm missing, you know, 15% of the words that I've taken out. 48:17.320 --> 48:19.320 The possibilities are small. 48:19.320 --> 48:22.320 It's small, right? There is 100,000 words in the lexicon. 48:22.320 --> 48:27.320 And what the machine spits out is a big probability vector, right? 48:27.320 --> 48:30.320 It's a bunch of numbers between the one one that's on to one. 48:30.320 --> 48:33.320 And we know how to do this with computers. 48:33.320 --> 48:37.320 So there, representing uncertainty in the prediction is relatively easy. 48:37.320 --> 48:42.320 And that's, in my opinion, why those techniques work for NLP. 48:42.320 --> 48:48.320 For images, if you ask, if you block a piece of an image and you ask the system reconstruct that piece of the image, 48:48.320 --> 48:54.320 there are many possible answers that are all perfectly legit, right? 48:54.320 --> 48:58.320 And how do you represent that, this set of possible answers? 48:58.320 --> 49:00.320 You can't train a system to make one prediction. 49:00.320 --> 49:04.320 You can train an old net to say, here it is, that's the image. 49:04.320 --> 49:07.320 Because there's a whole set of things that are compatible with it. 49:07.320 --> 49:12.320 So how do you get the machine to represent not a single output, but a whole set of outputs? 49:12.320 --> 49:20.320 And, you know, similarly with video prediction, there's a lot of things that can happen in the future of video. 49:20.320 --> 49:22.320 You're looking at me right now. I'm not moving my head very much. 49:22.320 --> 49:26.320 But, you know, I might, you know, turn my head to the left or to the right. 49:26.320 --> 49:30.320 If you don't have a system that can predict this, 49:30.320 --> 49:34.320 and you train it with least square to kind of minimize the error with a prediction on what I'm doing, 49:34.320 --> 49:39.320 what you get is a blurry image of myself in all possible future positions that I might be in. 49:39.320 --> 49:41.320 Which is not a good prediction. 49:41.320 --> 49:45.320 But so there might be other ways to do the self supervision, right? 49:45.320 --> 49:47.320 For visual scenes. 49:47.320 --> 49:49.320 Like what? 49:49.320 --> 49:55.320 I mean, if I knew I wouldn't tell you, I'd publish it first. I don't know. 49:55.320 --> 49:57.320 No, there might be. 49:57.320 --> 50:05.320 So, I mean, these are kind of, there might be artificial ways of like self play in games to where you can simulate part of the environment. 50:05.320 --> 50:10.320 Oh, that doesn't solve the problem. It's just a way of generating data. 50:10.320 --> 50:16.320 But because you have more of a control, that may mean you can control, yeah, it's a way to generate data. 50:16.320 --> 50:21.320 That's right. And because you can do huge amounts of data generation, that doesn't, you're right. 50:21.320 --> 50:26.320 Well, it's a creeps up on the problem from the side of data. 50:26.320 --> 50:28.320 I don't think that's the right way to creep up on the problem. 50:28.320 --> 50:31.320 It doesn't solve this problem of handling uncertainty in the world, right? 50:31.320 --> 50:42.320 So, if you have a machine learn a predictive model of the world in a game that is deterministic or quasi deterministic, it's easy, right? 50:42.320 --> 50:49.320 Just, you know, give a few frames of the game to a connet, put a bunch of layers, and then have the game generates the next few frames. 50:49.320 --> 50:54.320 And if the game is deterministic, it works fine. 50:54.320 --> 51:02.320 And that includes, you know, feeding the system with the action that your little character is going to take. 51:02.320 --> 51:09.320 The problem comes from the fact that the real world and most games are not entirely predictable. 51:09.320 --> 51:13.320 And so there you get those blurry predictions, and you can't do planning with blurry predictions. 51:13.320 --> 51:23.320 Right, so if you have a perfect model of the world, you can, in your head, run this model with a hypothesis for a sequence of actions, 51:23.320 --> 51:27.320 and you're going to predict the outcome of that sequence of actions. 51:27.320 --> 51:32.320 But if your model is imperfect, how can you plan? 51:32.320 --> 51:34.320 Yeah, it quickly explodes. 51:34.320 --> 51:39.320 What are your thoughts on the extension of this, which topic I'm super excited about. 51:39.320 --> 51:44.320 It's connected to something you were talking about in terms of robotics, is active learning. 51:44.320 --> 51:50.320 So, as opposed to sort of completely unsupervised or self supervised learning, 51:50.320 --> 51:58.320 you ask the system for human help for selecting parts you want annotated next. 51:58.320 --> 52:02.320 So if you think about a robot exploring a space, or a baby exploring a space, 52:02.320 --> 52:08.320 or a system exploring a data set, every once in a while asking for human input. 52:08.320 --> 52:12.320 Do you see value in that kind of work? 52:12.320 --> 52:14.320 I don't see transformative value. 52:14.320 --> 52:20.320 It's going to make things that we can already do more efficient, or they will learn slightly more efficiently, 52:20.320 --> 52:25.320 but it's not going to make machines sort of significantly more intelligent, I think. 52:25.320 --> 52:34.320 And by the way, there is no opposition, there is no conflict between self supervised learning, reinforcement learning, 52:34.320 --> 52:38.320 and supervised learning, or imitation learning, or active learning. 52:38.320 --> 52:43.320 I see self supervised learning as a preliminary to all of the above. 52:43.320 --> 52:44.320 Yes. 52:44.320 --> 52:54.320 So, the example I use very often is, how is it that, so if you use classical reinforcement learning, 52:54.320 --> 52:57.320 deep reinforcement learning, if you want. 52:57.320 --> 53:05.320 The best methods today, so called model free reinforcement learning, to learn to play Atari games, 53:05.320 --> 53:11.320 take about 80 hours of training to reach the level that any human can reach in about 15 minutes. 53:11.320 --> 53:17.320 They get better than humans, but it takes them a long time. 53:17.320 --> 53:27.320 Alpha star, okay, the, you know, all your vinyls and his teams, the system to play, to play Starcraft, 53:27.320 --> 53:34.320 plays, you know, a single map, a single type of player, 53:34.320 --> 53:45.320 and can reach better than human level with about the equivalent of 200 years of training playing against itself. 53:45.320 --> 53:50.320 It's 200 years, right? It's not something that no human can, could ever do. 53:50.320 --> 53:52.320 I mean, I'm not sure what lesson to take away from that. 53:52.320 --> 54:01.320 Okay, now, take those algorithms, the best RL algorithms we have today, to train a car to drive itself. 54:01.320 --> 54:05.320 It would probably have to drive millions of hours, it will have to kill thousands of pedestrians, 54:05.320 --> 54:09.320 it will have to run into thousands of trees, it will have to run off cliffs, 54:09.320 --> 54:15.320 and it had to run off cliffs multiple times before it figures out that it's a bad idea, first of all, 54:15.320 --> 54:18.320 and second of all, before it figures out how not to do it. 54:18.320 --> 54:24.320 And so, I mean, this type of learning obviously does not reflect the kind of learning that animals and humans do. 54:24.320 --> 54:27.320 There is something missing that's really, really important there. 54:27.320 --> 54:31.320 And my hypothesis, which I've been advocating for like five years now, 54:31.320 --> 54:39.320 is that we have predictive models of the world that include the ability to predict under uncertainty, 54:39.320 --> 54:45.320 and what allows us to not run off a cliff when we learn to drive. 54:45.320 --> 54:51.320 Most of us can learn to drive in about 20 or 30 hours of training without ever crashing, causing any accident. 54:51.320 --> 54:56.320 If we drive next to a cliff, we know that if we turn the wheel to the right, 54:56.320 --> 55:00.320 the car is going to run off the cliff and nothing good is going to come out of this, 55:00.320 --> 55:03.320 because we have a pretty good model of intuitive physics that tells us the car is going to fall. 55:03.320 --> 55:05.320 We know about gravity. 55:05.320 --> 55:12.320 Babies run this around the age of eight or nine months that objects don't float, they fall. 55:12.320 --> 55:16.320 And we have a pretty good idea of the effect of turning the wheel on the car, 55:16.320 --> 55:18.320 and we know we need to stay on the road. 55:18.320 --> 55:23.320 So there's a lot of things that we bring to the table, which is basically our predictive model of the world, 55:23.320 --> 55:31.320 and that model allows us to not do stupid things and to basically stay within the context of things we need to do. 55:31.320 --> 55:35.320 We still face unpredictable situations, and that's how we learn, 55:35.320 --> 55:39.320 but that allows us to learn really, really, really quickly. 55:39.320 --> 55:42.320 So that's called model based reinforcement learning. 55:42.320 --> 55:48.320 There's some imitation and supervised learning because we have a driving instructor that tells us occasionally what to do, 55:48.320 --> 55:52.320 but most of the learning is learning the model. 55:52.320 --> 55:55.320 Learning physics that we've done since we were babies. 55:55.320 --> 55:57.320 That's where almost all the learning... 55:57.320 --> 56:00.320 And the physics is somewhat transferable from... 56:00.320 --> 56:02.320 It's transferable from scene to scene. 56:02.320 --> 56:05.320 Stupid things are the same everywhere. 56:05.320 --> 56:08.320 Yeah. I mean, if you have an experience of the world, 56:08.320 --> 56:16.320 you don't need to be from a particularly intelligent species to know that if you spill water from a container, 56:16.320 --> 56:19.320 the rest is going to get wet. 56:19.320 --> 56:21.320 You might get wet. 56:21.320 --> 56:24.320 So cats know this, right? 56:24.320 --> 56:25.320 Yeah. 56:25.320 --> 56:30.320 So the main problem we need to solve is how do we learn models of the world? 56:30.320 --> 56:31.320 And that's what I'm interested in. 56:31.320 --> 56:34.320 That's what self supervised learning is all about. 56:34.320 --> 56:39.320 If you were to try to construct a benchmark for... 56:39.320 --> 56:41.320 Let's look at MNIST. 56:41.320 --> 56:43.320 I love that dataset. 56:43.320 --> 56:53.320 Do you think it's useful, interesting, slash possible to perform well on MNIST with just one example of each digit? 56:53.320 --> 56:58.320 And how would we solve that problem? 56:58.320 --> 56:59.320 The answer is probably yes. 56:59.320 --> 57:03.320 The question is what other type of learning are you allowed to do? 57:03.320 --> 57:08.320 So if what you're allowed to do is train on some gigantic dataset of labeled digit that's called transfer learning. 57:08.320 --> 57:10.320 And we know that works. 57:10.320 --> 57:13.320 We do this at Facebook like in production, right? 57:13.320 --> 57:20.320 We train large convolution nest to predict hashtags that people type on Instagram and we train on billions of images, literally billions. 57:20.320 --> 57:24.320 And then we chop off the last layer and fine tune on whatever task we want. 57:24.320 --> 57:25.320 That works really well. 57:25.320 --> 57:28.320 You can beat the ImageNet record with this. 57:28.320 --> 57:31.320 We actually open sourced the whole thing like a few weeks ago. 57:31.320 --> 57:33.320 Yeah, that's still pretty cool. 57:33.320 --> 57:40.320 But yeah, so what would be impressive and what's useful and impressive, what kind of transfer learning would be useful and impressive? 57:40.320 --> 57:42.320 Is it Wikipedia, that kind of thing? 57:42.320 --> 57:43.320 No, no. 57:43.320 --> 57:46.320 I don't think transfer learning is really where we should focus. 57:46.320 --> 57:59.320 We should try to have a kind of scenario for a benchmark where you have unlabeled data and it's a very large number of unlabeled data. 57:59.320 --> 58:10.320 It could be video clips, it could be where you do frame prediction, it could be images where you could choose to mask a piece of it. 58:10.320 --> 58:15.320 It could be whatever, but they're unlabeled and you're not allowed to label them. 58:15.320 --> 58:26.320 So you do some training on this and then you train on a particular supervised task, ImageNet or NIST. 58:26.320 --> 58:35.320 And you measure how your test error or validation error decreases as you increase the number of labeled training samples. 58:35.320 --> 58:47.320 And what you'd like to see is that your error decreases much faster than if you train from scratch, from random weights. 58:47.320 --> 58:56.320 So that to reach the same level of performance than a completely supervised, purely supervised system would reach, you would need way fewer samples. 58:56.320 --> 59:02.320 So that's the crucial question because it will answer the question to people interested in medical image analysis. 59:02.320 --> 59:17.320 Okay, if I want to get a particular level of error rate for this task, I know I need a million samples, can I do self supervised pre training to reduce this to about 100 or something? 59:17.320 --> 59:20.320 And you think the answer there is self supervised pre training? 59:20.320 --> 59:24.320 Yeah, some form of it. 59:24.320 --> 59:27.320 Telling you active learning, but you disagree? 59:27.320 --> 59:33.320 No, it's not useless, it's just not going to lead to a quantum leap, it's just going to make things that we already do. 59:33.320 --> 59:36.320 So you're way smarter than me, I just disagree with you. 59:36.320 --> 59:40.320 But I don't have anything to back that, it's just intuition. 59:40.320 --> 59:46.320 So I worked a lot of large scale data sets and there's something that might be magic in active learning. 59:46.320 --> 59:49.320 But okay, at least I said it publicly. 59:49.320 --> 59:52.320 At least I'm being an idiot publicly. 59:52.320 --> 1:00:05.320 Okay, it's not being an idiot, it's working with the data you have. I mean, certainly people are doing things like, okay, I have 3000 hours of imitation learning for cell driving car, but most of those are incredibly boring. 1:00:05.320 --> 1:00:12.320 What I like is select 10% of them that are kind of the most informative and with just that, I would probably reach the same. 1:00:12.320 --> 1:00:16.320 So it's a weak form of active learning if you want. 1:00:16.320 --> 1:00:20.320 Yes, but there might be a much stronger version. 1:00:20.320 --> 1:00:23.320 That's right. And that's an open question if it exists. 1:00:23.320 --> 1:00:26.320 The question is how much stronger can you get? 1:00:26.320 --> 1:00:35.320 Elon Musk is confident, talked to him recently, he's confident that large scale data and deep learning can solve the autonomous driving problem. 1:00:35.320 --> 1:00:40.320 What are your thoughts on the limits possibilities of deep learning in this space? 1:00:40.320 --> 1:00:42.320 It's obviously part of the solution. 1:00:42.320 --> 1:00:50.320 I mean, I don't think we'll ever have a cell driving system or it is not in the foreseeable future that does not use deep learning. 1:00:50.320 --> 1:00:52.320 Now, how much of it? 1:00:52.320 --> 1:01:03.320 So in the history of sort of engineering, particularly sort of AI like systems, there's generally a first phase where everything is built by hand. 1:01:03.320 --> 1:01:08.320 Then there is a second phase, and that was the case for autonomous driving, you know, 20, 30 years ago. 1:01:08.320 --> 1:01:18.320 There's a phase where there's a little bit of learning is used, but there's a lot of engineering that's involved in kind of, you know, taking care of corner cases and putting limits, etc. 1:01:18.320 --> 1:01:20.320 Because the learning system is not perfect. 1:01:20.320 --> 1:01:26.320 And then as technology progresses, we end up relying more and more on learning. 1:01:26.320 --> 1:01:31.320 That's the history of character recognition, so history of speech recognition, now computer vision, natural language processing. 1:01:31.320 --> 1:01:43.320 And I think the same is going to happen with autonomous driving that currently the methods that are closest to providing some level of autonomy, 1:01:43.320 --> 1:01:50.320 some, you know, decent level of autonomy where you don't expect a driver to kind of do anything, is where you constrain the world. 1:01:50.320 --> 1:02:00.320 So you only run within, you know, 100 square kilometers or square miles in Phoenix, but the weather is nice and the roads are wide, which is what Waymo is doing. 1:02:00.320 --> 1:02:13.320 You completely over engineer the car with tons of lidars and sophisticated sensors that are too expensive for consumer cars, but they're fine if you just run a fleet. 1:02:13.320 --> 1:02:20.320 And you engineer the thing, the hell out of the everything else, you map the entire world, so you have complete 3D model of everything. 1:02:20.320 --> 1:02:30.320 So the only thing that the perception system has to take care of is moving objects and construction and sort of, you know, things that weren't in your map. 1:02:30.320 --> 1:02:33.320 And you can engineer a good, you know, slam system. 1:02:33.320 --> 1:02:43.320 So that's kind of the current approach that's closest to some level of autonomy, but I think eventually the long term solution is going to rely more and more on learning 1:02:43.320 --> 1:02:50.320 and possibly using a combination of self supervised learning and model based reinforcement or something like that. 1:02:50.320 --> 1:02:57.320 But ultimately learning will be not just at the core, but really the fundamental part of the system. 1:02:57.320 --> 1:03:00.320 Yeah, it already is, but it will become more and more. 1:03:00.320 --> 1:03:04.320 What do you think it takes to build a system with human level intelligence? 1:03:04.320 --> 1:03:12.320 You talked about the AI system in the movie, her being way out of reach, our current reach, this might be outdated as well, but 1:03:12.320 --> 1:03:13.320 this is your way out of reach. 1:03:13.320 --> 1:03:15.320 It's the way out of reach. 1:03:15.320 --> 1:03:18.320 What would it take to build her? 1:03:18.320 --> 1:03:19.320 Do you think? 1:03:19.320 --> 1:03:24.320 So I can tell you the first two obstacles that we have to clear, but I don't know how many obstacles there are after this. 1:03:24.320 --> 1:03:32.320 So the image I usually use is that there is a bunch of mountains that we have to climb and we can see the first one, but we don't know if there are 50 mountains behind it or not. 1:03:32.320 --> 1:03:43.320 And this might be a good sort of metaphor for why AI researchers in the past have been overly optimistic about the result of AI. 1:03:43.320 --> 1:03:52.320 For example, Noah and Simon wrote the general problem solver and they call it the general problem solver. 1:03:52.320 --> 1:03:59.320 And of course, the first thing you realize is that all the problems you want to solve are exponential and so you can't actually use it for anything useful. 1:03:59.320 --> 1:04:02.320 Yeah, so yeah, all you see is the first peak. 1:04:02.320 --> 1:04:05.320 So what are the first couple of peaks for her? 1:04:05.320 --> 1:04:09.320 So the first peak, which is precisely what I'm working on, is cell supervision. 1:04:09.320 --> 1:04:17.320 How do we get machines to run models of the world by observation, kind of like babies and like young animals? 1:04:17.320 --> 1:04:23.320 So we've been working with, you know, cognitive scientists. 1:04:23.320 --> 1:04:32.320 So this Emmanuel Dupu, who is at Faire in Paris, is a half time, is also a researcher in French University. 1:04:32.320 --> 1:04:42.320 And he has this chart that shows how many months of life baby humans can learn different concepts. 1:04:42.320 --> 1:04:46.320 And you can measure this in various ways. 1:04:46.320 --> 1:04:56.320 So things like distinguishing animate objects from inanimate objects, you can tell the difference at age two, three months. 1:04:56.320 --> 1:05:03.320 Whether an object is going to stay stable is going to fall, you know, about four months you can tell. 1:05:03.320 --> 1:05:05.320 You know, there are various things like this. 1:05:05.320 --> 1:05:13.320 And then things like gravity, the fact that objects are not supposed to float in the air but are supposed to fall, you run this around the age of eight or nine months. 1:05:13.320 --> 1:05:19.320 So you look at a lot of eight month old babies, you give them a bunch of toys on their high chair. 1:05:19.320 --> 1:05:22.320 First thing they do is throw them on the ground and they look at them. 1:05:22.320 --> 1:05:27.320 It's because, you know, they're learning about, actively learning about gravity. 1:05:27.320 --> 1:05:33.320 So they're not trying to know you, but they need to do the experiment, right? 1:05:33.320 --> 1:05:39.320 So, you know, how do we get machines to learn like babies mostly by observation with a little bit of interaction 1:05:39.320 --> 1:05:46.320 and learning those models of the world because I think that's really a crucial piece of an intelligent autonomous system. 1:05:46.320 --> 1:05:51.320 So if you think about the architecture of an intelligent autonomous system, it needs to have a predictive model of the world. 1:05:51.320 --> 1:05:57.320 So something that says, here is a world at time t, here is a state of the world at time t plus one if I take this action. 1:05:57.320 --> 1:05:59.320 And it's not a single answer. 1:05:59.320 --> 1:06:01.320 It can be a distribution. 1:06:01.320 --> 1:06:05.320 Yeah, well, we don't know how to represent distributions in highly measured space. 1:06:05.320 --> 1:06:07.320 So it's got to be something weaker than that. 1:06:07.320 --> 1:06:10.320 With some representation of uncertainty. 1:06:10.320 --> 1:06:15.320 If you have that, then you can do what optimal control theory is called model predictive control, 1:06:15.320 --> 1:06:21.320 which means that you can run your model with a hypothesis for a sequence of action and then see the result. 1:06:21.320 --> 1:06:25.320 Now what you need, the other thing you need is some sort of objective that you want to optimize. 1:06:25.320 --> 1:06:28.320 Am I reaching the goal of grabbing the subject? 1:06:28.320 --> 1:06:30.320 Am I minimizing energy? 1:06:30.320 --> 1:06:31.320 Am I whatever, right? 1:06:31.320 --> 1:06:34.320 So there is some sort of objective that you have to minimize. 1:06:34.320 --> 1:06:40.320 And so in your head, if you have this model, you can figure out the sequence of action that will optimize your objective. 1:06:40.320 --> 1:06:46.320 That objective is something that ultimately is rooted in your basal ganglia, at least in the human brain. 1:06:46.320 --> 1:06:47.320 That's what it's. 1:06:47.320 --> 1:06:52.320 Basal ganglia computes your level of contentment or miscontentment. 1:06:52.320 --> 1:06:53.320 I don't know if that's a word. 1:06:53.320 --> 1:06:55.320 Unhappiness, okay. 1:06:55.320 --> 1:06:57.320 Discontentment. 1:06:57.320 --> 1:06:58.320 Discontentment. 1:06:58.320 --> 1:07:10.320 And so your entire behavior is driven towards kind of minimizing that objective, which is maximizing your contentment computed by your basal ganglia. 1:07:10.320 --> 1:07:16.320 And what you have is an objective function, which is basically a predictor of what your basal ganglia is going to tell you. 1:07:16.320 --> 1:07:23.320 So you're not going to put your hand on fire because you know it's going to burn and you're going to get hurt. 1:07:23.320 --> 1:07:29.320 And you're predicting this because of your model of the world and your sort of predictor of this objective, right? 1:07:29.320 --> 1:07:43.320 So if you have those three components, you have four components, you have the hardwired contentment objective computer, if you want, calculator. 1:07:43.320 --> 1:07:44.320 And then you have the three components. 1:07:44.320 --> 1:07:48.320 One is the objective predictor, which basically predicts your level of contentment. 1:07:48.320 --> 1:08:01.320 One is the model of the world, and there's a third module I didn't mention, which is the module that will figure out the best course of action to optimize an objective given your model. 1:08:01.320 --> 1:08:02.320 Okay? 1:08:02.320 --> 1:08:03.320 Yeah. 1:08:03.320 --> 1:08:08.320 Collision policy, policy network or something like that, right? 1:08:08.320 --> 1:08:15.320 Now, you need those three components to act autonomously intelligently, and you can be stupid in three different ways. 1:08:15.320 --> 1:08:18.320 You can be stupid because your model of the world is wrong. 1:08:18.320 --> 1:08:24.320 You can be stupid because your objective is not aligned with what you actually want to achieve. 1:08:24.320 --> 1:08:26.320 Okay? 1:08:26.320 --> 1:08:29.320 In humans, that would be a psychopath. 1:08:29.320 --> 1:08:40.320 And then the third thing, the third way you can be stupid is that you have the right model, you have the right objective, but you're unable to figure out a course of action to optimize your objective given your model. 1:08:40.320 --> 1:08:41.320 Right. 1:08:41.320 --> 1:08:43.320 Okay? 1:08:43.320 --> 1:08:47.320 Some people who are in charge of big countries actually have all three that are wrong. 1:08:47.320 --> 1:08:50.320 All right. 1:08:50.320 --> 1:08:51.320 Which countries? 1:08:51.320 --> 1:08:52.320 I don't know. 1:08:52.320 --> 1:08:53.320 Okay. 1:08:53.320 --> 1:09:04.320 So if we think about this agent, if we think about the movie Her, you've criticized the art project that is Sophia the Robot. 1:09:04.320 --> 1:09:14.320 And what that project essentially does is uses our natural inclination to anthropomorphize things that look like human and give them more. 1:09:14.320 --> 1:09:20.320 Do you think that could be used by AI systems like in the movie Her? 1:09:20.320 --> 1:09:26.320 So do you think that body is needed to create a feeling of intelligence? 1:09:26.320 --> 1:09:32.320 Well, if Sophia was just an art piece, I would have no problem with it, but it's presented as something else. 1:09:32.320 --> 1:09:35.320 Let me add that comment real quick. 1:09:35.320 --> 1:09:42.320 If creators of Sophia could change something about their marketing or behavior in general, what would it be? 1:09:42.320 --> 1:09:45.320 I'm just about everything. 1:09:45.320 --> 1:09:50.320 I mean, don't you think, here's a tough question. 1:09:50.320 --> 1:09:52.320 Let me, so I agree with you. 1:09:52.320 --> 1:09:59.320 So Sophia is not, the general public feels that Sophia can do way more than she actually can. 1:09:59.320 --> 1:10:00.320 That's right. 1:10:00.320 --> 1:10:09.320 And the people who created Sophia are not honestly publicly communicating, trying to teach the public. 1:10:09.320 --> 1:10:10.320 Right. 1:10:10.320 --> 1:10:13.320 But here's a tough question. 1:10:13.320 --> 1:10:29.320 Don't you think the same thing is scientists in industry and research are taking advantage of the same misunderstanding in the public when they create AI companies or publish stuff? 1:10:29.320 --> 1:10:31.320 Some companies, yes. 1:10:31.320 --> 1:10:34.320 I mean, there is no sense of, there's no desire to delude. 1:10:34.320 --> 1:10:38.320 There's no desire to kind of overclaim what something is done. 1:10:38.320 --> 1:10:39.320 Right. 1:10:39.320 --> 1:10:42.320 You publish a paper on AI that has this result on ImageNet. 1:10:42.320 --> 1:10:43.320 It's pretty clear. 1:10:43.320 --> 1:10:45.320 I mean, it's not even interesting anymore. 1:10:45.320 --> 1:10:48.320 But I don't think there is that. 1:10:48.320 --> 1:10:57.320 I mean, the reviewers are generally not very forgiving of unsupported claims of this type. 1:10:57.320 --> 1:11:05.320 And, but there are certainly quite a few startups that have had a huge amount of hype around this that I find extremely damaging. 1:11:05.320 --> 1:11:07.320 And I've been calling it out when I've seen it. 1:11:07.320 --> 1:11:15.320 So, yeah, but to go back to your original question, like the necessity of embodiment, I think, I don't think embodiment is necessary. 1:11:15.320 --> 1:11:17.320 I think grounding is necessary. 1:11:17.320 --> 1:11:22.320 So I don't think we're going to get machines that really understand language without some level of grounding in the real world. 1:11:22.320 --> 1:11:29.320 And it's not clear to me that language is a high enough bandwidth medium to communicate how the real world works. 1:11:29.320 --> 1:11:30.320 I think for this... 1:11:30.320 --> 1:11:33.320 Can you talk about what grounding means to you? 1:11:33.320 --> 1:11:34.320 So grounding means that... 1:11:34.320 --> 1:11:41.320 So there is this classic problem of common sense reasoning, you know, the Winograd schema, right? 1:11:41.320 --> 1:11:49.320 And so I tell you the trophy doesn't fit in the suitcase because it's too big, or the trophy doesn't fit in the suitcase because it's too small. 1:11:49.320 --> 1:11:53.320 And the it in the first case refers to the trophy in the second case to the suitcase. 1:11:53.320 --> 1:11:58.320 And the reason you can figure this out is because you know what the trophy in the suitcase are, you know, one is supposed to fit in the other one, 1:11:58.320 --> 1:12:05.320 and you know the notion of size and a big object doesn't fit in a small object unless it's a target, you know, things like that, right? 1:12:05.320 --> 1:12:11.320 So you have this knowledge of how the world works, of geometry and things like that. 1:12:11.320 --> 1:12:18.320 I don't believe you can learn everything about the world by just being told in language how the world works. 1:12:18.320 --> 1:12:26.320 You need some low level perception of the world, you know, be it visual touch, you know, whatever, but some higher bandwidth perception of the world. 1:12:26.320 --> 1:12:31.320 So by reading all the world's text, you still may not have enough information. 1:12:31.320 --> 1:12:32.320 That's right. 1:12:32.320 --> 1:12:37.320 There's a lot of things that just will never appear in text and that you can't really infer. 1:12:37.320 --> 1:12:43.320 So I think common sense will emerge from, you know, certainly a lot of language interaction, 1:12:43.320 --> 1:12:51.320 but also with watching videos or perhaps even interacting in virtual environments and possibly, you know, robot interacting in the real world. 1:12:51.320 --> 1:12:55.320 But I don't actually believe necessarily that this last one is absolutely necessary. 1:12:55.320 --> 1:12:59.320 But I think there's a need for some grounding. 1:12:59.320 --> 1:13:04.320 But the final product doesn't necessarily need to be embodied, you're saying? 1:13:04.320 --> 1:13:05.320 No. 1:13:05.320 --> 1:13:07.320 It just needs to have an awareness grounding. 1:13:07.320 --> 1:13:08.320 Right. 1:13:08.320 --> 1:13:16.320 It needs to know how the world works to have, you know, to not be frustrated, frustrating to talk to. 1:13:16.320 --> 1:13:20.320 And you talked about emotions being important. 1:13:20.320 --> 1:13:22.320 That's a whole other topic. 1:13:22.320 --> 1:13:33.320 Well, so, you know, I talked about this, the base of ganglia as the, you know, the thing that calculates your level of misconstantment, contentment. 1:13:33.320 --> 1:13:38.320 This is the other module that sort of tries to do a prediction of whether you're going to be content or not. 1:13:38.320 --> 1:13:40.320 That's the source of some emotion. 1:13:40.320 --> 1:13:47.320 So fear, for example, is an anticipation of bad things that can happen to you, right? 1:13:47.320 --> 1:13:52.320 You have this inkling that there is some chance that something really bad is going to happen to you and that creates fear. 1:13:52.320 --> 1:13:56.320 When you know for sure that something bad is going to happen to you, you kind of give up, right? 1:13:56.320 --> 1:13:57.320 It's not going to be anymore. 1:13:57.320 --> 1:13:59.320 It's uncertainty that creates fear. 1:13:59.320 --> 1:14:04.320 So the punchline is we're not going to have autonomous intelligence without emotions. 1:14:04.320 --> 1:14:06.320 Okay. 1:14:06.320 --> 1:14:08.320 Whatever the heck emotions are. 1:14:08.320 --> 1:14:13.320 So you mentioned very practical things of fear, but there's a lot of other mess around it. 1:14:13.320 --> 1:14:16.320 But there are kind of the results of, you know, drives. 1:14:16.320 --> 1:14:17.320 Yeah. 1:14:17.320 --> 1:14:19.320 There's deeper biological stuff going on. 1:14:19.320 --> 1:14:21.320 And I've talked to a few folks on this. 1:14:21.320 --> 1:14:27.320 There's this fascinating stuff that ultimately connects to our brain. 1:14:27.320 --> 1:14:30.320 If we create an AGI system. 1:14:30.320 --> 1:14:31.320 Sorry. 1:14:31.320 --> 1:14:32.320 Human level intelligence. 1:14:32.320 --> 1:14:34.320 Human level intelligence system. 1:14:34.320 --> 1:14:37.320 And you get to ask her one question. 1:14:37.320 --> 1:14:40.320 What would that question be? 1:14:40.320 --> 1:14:45.320 You know, I think the first one we'll create will probably not be that smart. 1:14:45.320 --> 1:14:47.320 They'll be like a four year old. 1:14:47.320 --> 1:14:48.320 Okay. 1:14:48.320 --> 1:14:53.320 So you would have to ask her a question to know she's not that smart. 1:14:53.320 --> 1:14:54.320 Yeah. 1:14:54.320 --> 1:14:57.320 Well, what's a good question to ask, you know, to be impressed? 1:14:57.320 --> 1:15:00.320 With the cause of wind. 1:15:00.320 --> 1:15:06.320 And if she answers, oh, it's because the leaves of the tree are moving and that creates wind. 1:15:06.320 --> 1:15:08.320 She's onto something. 1:15:08.320 --> 1:15:12.320 And if she says, that's a stupid question, she's really onto something. 1:15:12.320 --> 1:15:13.320 No. 1:15:13.320 --> 1:15:17.320 And then you tell her, actually, you know, here is the real thing. 1:15:17.320 --> 1:15:20.320 And she says, oh, yeah, that makes sense. 1:15:20.320 --> 1:15:26.320 So questions that, that reveal the ability to do common sense reasoning about the physical world. 1:15:26.320 --> 1:15:27.320 Yeah. 1:15:27.320 --> 1:15:29.320 And you know, some of that will cause an inference. 1:15:29.320 --> 1:15:31.320 Causal inference. 1:15:31.320 --> 1:15:33.320 Well, it was a huge honor. 1:15:33.320 --> 1:15:35.320 Congratulations on your touring award. 1:15:35.320 --> 1:15:37.320 Thank you so much for talking today. 1:15:37.320 --> 1:15:38.320 Thank you. 1:15:38.320 --> 1:15:58.320 Thank you.