WEBVTT 00:00:00.080 --> 00:00:01.997 The following is a conversation with Yan 00:00:02.080 --> 00:00:04.477 Lakun. He's considered to be one of the 00:00:04.560 --> 00:00:09.357 fathers of deep learning, which if you've been hiding under a rock, is the 00:00:09.440 --> 00:00:12.637 recent revolution in AI that has captivated the world with the 00:00:12.720 --> 00:00:14.717 possibility of what machines can learn 00:00:14.800 --> 00:00:17.677 from data. He's a professor at New York 00:00:17.760 --> 00:00:19.757 University, a vice president and chief 00:00:19.840 --> 00:00:23.197 AI scientist at Facebook, and co-reient 00:00:23.280 --> 00:00:27.677 of the touring award for his work on deep learning. He's probably best known 00:00:27.760 --> 00:00:31.917 as the founding father of convolutional neural networks, in particular their 00:00:32.000 --> 00:00:33.437 application to optical character 00:00:33.520 --> 00:00:36.557 recognition and the famed emnest data 00:00:36.640 --> 00:00:40.077 set. He is also an outspoken personality 00:00:40.160 --> 00:00:44.397 unafraid to speak his mind in a distinctive French accent and explore 00:00:44.480 --> 00:00:46.637 provocative ideas both in the rigorous 00:00:46.720 --> 00:00:49.437 medium of academic research and the 00:00:49.520 --> 00:00:51.437 somewhat less rigorous medium of Twitter 00:00:51.520 --> 00:00:54.077 and Facebook. This is the artificial 00:00:54.160 --> 00:00:58.717 intelligence podcast. If you enjoy it, subscribe on YouTube, give it five stars 00:00:58.800 --> 00:01:01.037 on iTunes, support it on Patreon, or 00:01:01.120 --> 00:01:03.277 simply connect with me on Twitter at Lex 00:01:03.360 --> 00:01:07.517 Freedman, spelled F R I D M. And now, 00:01:07.600 --> 00:01:13.000 here's my conversation with Yan Lun. 00:01:21.220 --> 00:01:23.240 [music] 00:01:28.880 --> 00:01:31.277 You said that 2001 Space Odyssey is one 00:01:31.360 --> 00:01:33.357 of your favorite movies. 00:01:33.440 --> 00:01:36.877 Hell 9000 decides to get rid of the 00:01:36.960 --> 00:01:38.397 astronauts for people who haven't seen 00:01:38.480 --> 00:01:42.797 the movie. Spoiler alert, because he it 00:01:42.880 --> 00:01:46.637 she believes that the astronauts they 00:01:46.720 --> 00:01:49.037 will interfere with the mission. Do you 00:01:49.120 --> 00:01:54.637 see Hal as flawed in some fundamental way or even evil or did he do the right 00:01:54.720 --> 00:01:58.717 thing? >> Neither. There's no notion of evil in 00:01:58.800 --> 00:02:01.117 that in that context other than the fact 00:02:01.200 --> 00:02:06.317 that people die. But it was an example of what people call uh value 00:02:06.400 --> 00:02:10.077 misalignment, right? And you give an objective to a machine and the machine 00:02:10.160 --> 00:02:14.797 strives to achieve this objective and if you don't put any constraints on this 00:02:14.880 --> 00:02:16.717 objective like don't kill people and 00:02:16.800 --> 00:02:19.677 don't do things like this. 00:02:19.760 --> 00:02:24.717 The machine given the power will do stupid things just to achieve this this 00:02:24.800 --> 00:02:28.397 objective or damaging things to achieve this objective. It's a little bit like I 00:02:28.480 --> 00:02:30.237 mean we are used to this in the context 00:02:30.320 --> 00:02:32.957 of human society. 00:02:33.040 --> 00:02:36.877 We we put in place 00:02:36.960 --> 00:02:40.397 laws to prevent people from doing bad things because spontaneously they would 00:02:40.480 --> 00:02:44.877 do those bad things, right? So we have to shape their 00:02:44.960 --> 00:02:48.077 cost function, their objective function if you want through laws to kind of 00:02:48.160 --> 00:02:50.797 correct and education obviously to sort 00:02:50.880 --> 00:02:55.117 of correct for for those. So maybe just 00:02:55.200 --> 00:02:59.197 pushing a little further on on that point, 00:02:59.280 --> 00:03:01.517 Hal, you know, there's a mission. 00:03:01.600 --> 00:03:04.157 there's a this fuzziness around the the 00:03:04.240 --> 00:03:06.077 ambiguity around what the actual mission 00:03:06.160 --> 00:03:08.557 is but 00:03:08.640 --> 00:03:11.357 you know do you think that there will be 00:03:11.440 --> 00:03:15.917 a time from a utilitarian perspective where an AI system where it is not 00:03:16.000 --> 00:03:20.557 misalignment where it is alignment for the greater good of society that an AI 00:03:20.640 --> 00:03:23.037 system will make decisions that are difficult 00:03:23.120 --> 00:03:27.997 >> well that's the trick I mean eventually we'll have to figure out how to do this 00:03:28.080 --> 00:03:30.797 and again we're not starting from scratch because we've been doing this 00:03:30.880 --> 00:03:33.677 with humans for for millennia. 00:03:33.760 --> 00:03:35.837 >> So designing objective functions for 00:03:35.920 --> 00:03:37.437 people is something that we know how to 00:03:37.520 --> 00:03:40.957 do and we don't do it by um you know 00:03:41.040 --> 00:03:43.597 programming things although 00:03:43.680 --> 00:03:48.397 the legal code is called code. Um so that tells you something and it's 00:03:48.480 --> 00:03:51.277 actually the design of an objective function. That's really what legal code 00:03:51.360 --> 00:03:54.877 is right. It tells you here is what you can do here is what you can't do. If you 00:03:54.960 --> 00:03:58.877 do it you pay that much. That's that's an objective function. 00:03:58.960 --> 00:04:02.957 So there is this idea somehow that it's a new thing for people to try to design 00:04:03.040 --> 00:04:05.917 objective functions that are aligned with the common good. But no, we've been 00:04:06.000 --> 00:04:07.357 writing laws for millennia and that's 00:04:07.440 --> 00:04:09.997 exactly what it is. Um so this that's 00:04:10.080 --> 00:04:13.837 where you know uh the science of uh law 00:04:13.920 --> 00:04:18.477 making and uh and computer science will >> come together 00:04:18.560 --> 00:04:21.597 >> will come together. >> So it's nothing there's nothing special 00:04:21.680 --> 00:04:24.637 about hell or AI systems. It's just a 00:04:24.720 --> 00:04:28.797 continuation of tools used to make some of these difficult ethical judgments 00:04:28.880 --> 00:04:31.757 that laws make. >> Yeah. And we and we have systems like 00:04:31.840 --> 00:04:37.117 this already that you know uh make many decisions for for ourselves in society 00:04:37.200 --> 00:04:40.957 that you know need to be designed in a way that they like you know rules about 00:04:41.040 --> 00:04:45.837 things that sometime sometimes have bad side effects and we have to be flexible 00:04:45.920 --> 00:04:48.957 enough about those rules so that they can be broken when it's obvious that 00:04:49.040 --> 00:04:52.637 they shouldn't be applied. So, you don't see this on the camera 00:04:52.720 --> 00:04:56.317 here, but all the decoration in this room is all pictures from 2001 of Space 00:04:56.400 --> 00:05:00.557 Odyssey. [laughter] >> Wow. Is that by accident or is there a 00:05:00.640 --> 00:05:00.797 lot? 00:05:00.880 --> 00:05:04.557 >> It's not by accident. It's by design. 00:05:04.640 --> 00:05:07.757 >> Oh, wow. So, if you were um if you were 00:05:07.840 --> 00:05:12.637 to build HAL 10,000, so an improvement of HAL 9,000, 00:05:12.720 --> 00:05:14.717 what would you improve? Well, first of 00:05:14.800 --> 00:05:17.757 all, I wouldn't ask it to hold secrets 00:05:17.840 --> 00:05:21.757 and tell lies because that's really what breaks it in the end. That's the the 00:05:21.840 --> 00:05:26.157 fact that it's asking itself questions about the purpose of the mission and 00:05:26.240 --> 00:05:29.997 it's you know pieces things together that it's heard you know all the secrecy 00:05:30.080 --> 00:05:34.077 of the preparation of the mission and the fact that it was discovery on on the 00:05:34.160 --> 00:05:35.677 lunar surface that really was kept 00:05:35.760 --> 00:05:38.957 secret and and one part of how's memory 00:05:39.040 --> 00:05:43.117 knows this and the other part is does not know it and is supposed to not tell 00:05:43.200 --> 00:05:45.197 anyone and that creates internal 00:05:45.280 --> 00:05:46.797 conflict. So you think there's never 00:05:46.880 --> 00:05:50.317 should be a set of things that an AI 00:05:50.400 --> 00:05:53.677 system should not be allowed 00:05:53.760 --> 00:05:56.717 like a a set of facts that should not be 00:05:56.800 --> 00:05:59.357 shared with the human operators. 00:05:59.440 --> 00:06:01.437 Well, I think no. I think that I think 00:06:01.520 --> 00:06:06.797 it should be a bit like in the design of 00:06:06.880 --> 00:06:09.357 autonomous AI systems. Um, there should 00:06:09.440 --> 00:06:12.237 be the equivalent of, you know, the the 00:06:12.320 --> 00:06:16.237 the the oath that hypocratic oath Yeah. 00:06:16.320 --> 00:06:21.197 >> that uh doctors sign up to, right? So, there's certain things, certain rules 00:06:21.280 --> 00:06:25.197 that that that you have to abide by and we can sort of hardwire this into into 00:06:25.280 --> 00:06:29.197 our into our machines to kind of make sure they don't go. So I'm not you know 00:06:29.280 --> 00:06:33.837 an advocate of the the three law three laws of robotics you know the azimov 00:06:33.920 --> 00:06:35.277 kind of thing because I don't think it's 00:06:35.360 --> 00:06:38.477 practical but but you know some some 00:06:38.560 --> 00:06:41.757 level of uh of limits but but to be 00:06:41.840 --> 00:06:44.637 clear uh this is not these are not 00:06:44.720 --> 00:06:48.317 questions that are kind of really worth 00:06:48.400 --> 00:06:52.237 asking today because we just don't have the technology to do this. We don't we 00:06:52.320 --> 00:06:54.637 don't have autonomous intelligent machines. We have intelligent machines, 00:06:54.720 --> 00:06:57.997 semi- intelligent machines that are very specialized, 00:06:58.080 --> 00:07:01.277 but they don't they don't really sort of satisfy an objective. They're just, you 00:07:01.360 --> 00:07:03.757 know, kind of trained to do one thing. 00:07:03.840 --> 00:07:07.037 >> So, until we have some idea for design 00:07:07.120 --> 00:07:11.837 of a full-fledged autonomous intelligent system, asking the question of how how 00:07:11.920 --> 00:07:15.677 we design is objective, I think is a little a little too abstract. 00:07:15.760 --> 00:07:17.117 >> It's a little too abstract. There's 00:07:17.200 --> 00:07:19.757 useful elements to it in that it helps 00:07:19.840 --> 00:07:23.997 us understand our own ethical codes, 00:07:24.080 --> 00:07:26.397 humans. So even just as a thought 00:07:26.480 --> 00:07:29.437 experiment, if you imagine that an AGI 00:07:29.520 --> 00:07:32.397 system is here today, h how would we 00:07:32.480 --> 00:07:34.317 program it is a kind of nice thought 00:07:34.400 --> 00:07:37.677 experiment of constructing how should we 00:07:37.760 --> 00:07:40.717 have a law have a system of laws for us 00:07:40.800 --> 00:07:42.877 humans. It's it's just a nice practical 00:07:42.960 --> 00:07:45.677 tool. And I think there's echoes of that 00:07:45.760 --> 00:07:49.197 idea too in the AI systems we have today 00:07:49.280 --> 00:07:53.357 that don't have to be that intelligent like autonomous vehicles. These things 00:07:53.440 --> 00:07:55.757 start creeping in that they're worth 00:07:55.840 --> 00:07:57.037 thinking about but certainly they 00:07:57.120 --> 00:08:00.877 shouldn't be framed as as hell. Yeah. 00:08:00.960 --> 00:08:04.237 >> Looking back, what is the most I'm sorry 00:08:04.320 --> 00:08:07.677 if it's a silly question, but what is 00:08:07.760 --> 00:08:12.317 the most beautiful or surprising idea in deep learning or AI in general that 00:08:12.400 --> 00:08:14.077 you've ever come across sort of 00:08:14.160 --> 00:08:17.117 personally where you sat back 00:08:17.200 --> 00:08:21.997 and and just had this kind of a that's pretty cool moment. That's nice 00:08:22.080 --> 00:08:24.397 surprising. >> I don't know if it's an idea rather than 00:08:24.480 --> 00:08:27.277 a sort of empirical 00:08:27.360 --> 00:08:29.357 uh fact. 00:08:29.440 --> 00:08:31.997 The fact that 00:08:32.080 --> 00:08:36.797 you can build gigantic neural nets train them on, 00:08:36.880 --> 00:08:40.557 you know, relatively small amounts of data relatively 00:08:40.640 --> 00:08:44.557 with stoastic gradient and descent and that it actually works. Uh breaks 00:08:44.640 --> 00:08:48.637 everything you read in every textbook, right? Every pre deep learning textbook 00:08:48.720 --> 00:08:53.037 that told you you need to have fewer parameters than you have data samples. 00:08:53.120 --> 00:08:56.397 Um you know if you have a non-convex objective function, you have no 00:08:56.480 --> 00:08:59.197 guarantee of convergence. You know all those things that you read in textbook 00:08:59.280 --> 00:09:02.397 and they tell you stay away from this and they're all wrong. 00:09:02.480 --> 00:09:05.117 >> Huge number of parameters non-convex 00:09:05.200 --> 00:09:10.797 and somehow which just vary relative to the number of uh parameters data it's 00:09:10.880 --> 00:09:14.717 able to learn anything >> right >> does that still surprise you today? 00:09:14.800 --> 00:09:17.517 >> Well it it was kind of obvious to me 00:09:17.600 --> 00:09:21.997 before I knew anything that that this is a good idea and then it became 00:09:22.080 --> 00:09:26.397 surprising that it worked because I started reading those textbooks. 00:09:26.480 --> 00:09:29.997 >> Okay. Wait. So, okay. So, can you talk through the intuition of why it was 00:09:30.080 --> 00:09:31.437 obvious to you if you remember? 00:09:31.520 --> 00:09:34.797 >> Well, okay. So, the intuition was it's it's sort of like, you know, those 00:09:34.880 --> 00:09:37.437 people in the late 19th century who 00:09:37.520 --> 00:09:41.437 proved that heavier than than air flight 00:09:41.520 --> 00:09:45.597 was impossible, right? And of course, you have birds, right? They do fly. And 00:09:45.680 --> 00:09:50.237 so, on the face of it, it it's obviously wrong as an empirical question, right? 00:09:50.320 --> 00:09:54.397 And so, we have the same kind of thing that, you know, the we know that the 00:09:54.480 --> 00:09:58.557 brain works. We don't know how but we know it works and we know it's a large 00:09:58.640 --> 00:10:01.917 network of neurons in interaction and that learning takes place by changing 00:10:02.000 --> 00:10:05.837 the connections. So kind of getting this level of inspiration without covering 00:10:05.920 --> 00:10:11.837 the details but sort of trying to derive basic principles 00:10:11.920 --> 00:10:15.677 you know that kind of gives you a clue as to which direction to go. There's 00:10:15.760 --> 00:10:19.357 also the idea somehow that I' I've been convinced of since I was in undergrad 00:10:19.440 --> 00:10:22.877 that even before that intelligence is 00:10:22.960 --> 00:10:25.277 inseparable from learning. So you the 00:10:25.360 --> 00:10:27.277 idea somehow that you can create an 00:10:27.360 --> 00:10:30.077 intelligent machine by basically 00:10:30.160 --> 00:10:33.597 programming for me was a non-starter you 00:10:33.680 --> 00:10:36.237 know from the start every intelligent 00:10:36.320 --> 00:10:40.717 entity that we know about arrives at this intelligence through 00:10:40.800 --> 00:10:43.197 learning so learning you know machine 00:10:43.280 --> 00:10:46.397 learning was completely obvious path 00:10:46.480 --> 00:10:50.637 uh also because I'm lazy so you know [laughter] 00:10:50.720 --> 00:10:54.477 >> you automate basically everything and learning is the automation of >> exactly 00:10:54.560 --> 00:10:56.637 >> intelligence, >> right? 00:10:56.720 --> 00:10:59.677 >> So do you think so what is learning 00:10:59.760 --> 00:11:02.717 then? What falls under learning? Because 00:11:02.800 --> 00:11:06.397 do you think of reasoning is learning? 00:11:06.480 --> 00:11:08.957 >> Well, reasoning is certainly a a a 00:11:09.040 --> 00:11:11.597 consequence of of learning as well just 00:11:11.680 --> 00:11:14.397 like other functions of uh of the brain. 00:11:14.480 --> 00:11:19.197 The big question about reasoning is how do you make reasoning compatible with 00:11:19.280 --> 00:11:22.077 gradient-based learning? >> Do you think neural networks can be made 00:11:22.160 --> 00:11:25.837 to reason? Yes, there is no question about that. Again, we have a good 00:11:25.920 --> 00:11:29.997 example, right? The question is is how? So, the question 00:11:30.080 --> 00:11:34.237 is how much prior structure do you have to put in the neural net so that 00:11:34.320 --> 00:11:38.557 something like human reasoning will emerge from it you know from learning. 00:11:38.640 --> 00:11:41.117 Uh another question is all of our kind 00:11:41.200 --> 00:11:43.757 of model of what reasoning is that are 00:11:43.840 --> 00:11:49.437 based on logic are discrete and and and are therefore incompatible with gradient 00:11:49.520 --> 00:11:52.157 based learning. Mhm. >> And I'm a very strong believer in this 00:11:52.240 --> 00:11:57.117 idea gradient based learning. I don't believe that uh other types of learning 00:11:57.200 --> 00:11:59.677 that don't use kind of gradient information if you want. 00:11:59.760 --> 00:12:01.197 >> So you don't like discrete mathematics. 00:12:01.280 --> 00:12:02.797 You don't like anything discrete. 00:12:02.880 --> 00:12:05.917 >> Well, that's it's not that I don't like it. It's just that it's it's 00:12:06.000 --> 00:12:07.917 incompatible with learning and I'm I'm a 00:12:08.000 --> 00:12:13.277 big fan of learning, right? So in fact, that's perhaps one reason why 00:12:13.360 --> 00:12:15.997 deep learning has been kind of looked at with suspicion by a lot of computer 00:12:16.080 --> 00:12:19.517 scientists because the math is very different. the math that uh you use for 00:12:19.600 --> 00:12:21.357 deep learning, you know, kind of has 00:12:21.440 --> 00:12:24.237 more to do with, you know, cybernetics, 00:12:24.320 --> 00:12:28.477 uh the kind of math you do in electrical engineering than the kind of math you do 00:12:28.560 --> 00:12:32.557 in computer science. [snorts] And and you know, nothing in in machine learning 00:12:32.640 --> 00:12:34.477 is exact, right? Computer science is all 00:12:34.560 --> 00:12:39.597 about sort of, you know, obviously compulsive attention to details of like, 00:12:39.680 --> 00:12:43.117 you know, every index has to be right and you can prove that an algorithm is 00:12:43.200 --> 00:12:47.587 correct, right? Uh machine learning is the science of sloppiness. Really? 00:12:47.670 --> 00:12:49.690 [laughter] 00:12:49.920 --> 00:12:53.917 >> That's beautiful. So okay, maybe let's 00:12:54.000 --> 00:12:56.557 feel around in the dark of what is a 00:12:56.640 --> 00:13:00.397 neural network that reasons or a system 00:13:00.480 --> 00:13:02.077 that 00:13:02.160 --> 00:13:05.037 is works with continuous functions 00:13:05.120 --> 00:13:09.597 that's able to do uh build knowledge. 00:13:09.680 --> 00:13:11.917 However we think about reasoning, build 00:13:12.000 --> 00:13:14.397 on previous knowledge, build on extra 00:13:14.480 --> 00:13:16.717 knowledge, create new knowledge, 00:13:16.800 --> 00:13:19.437 generalize outside of any training set 00:13:19.520 --> 00:13:22.717 ever built. What does that look like? If 00:13:22.800 --> 00:13:25.517 uh yeah, maybe do you have inklings of 00:13:25.600 --> 00:13:27.517 thoughts of what that might look like? 00:13:27.600 --> 00:13:31.917 >> Well, yeah. I mean, yes and no. If I had precise ideas about this, I think, you 00:13:32.000 --> 00:13:36.157 know, we'd be building it right now. But and there are people working on this or 00:13:36.240 --> 00:13:38.477 whose main research interest is actually 00:13:38.560 --> 00:13:43.197 exactly that right. So what you need to have is a working memory. So you need to 00:13:43.280 --> 00:13:44.877 have 00:13:44.960 --> 00:13:47.357 some device if you want uh some 00:13:47.440 --> 00:13:51.197 subsystem that can store uh a relatively 00:13:51.280 --> 00:13:53.757 large number of factual episodic 00:13:53.840 --> 00:13:57.117 information for you know a reasonable 00:13:57.200 --> 00:14:00.077 amount of time. So you you know in the 00:14:00.160 --> 00:14:01.517 in the brain for example there are kind 00:14:01.600 --> 00:14:04.157 of three main types of memory. one is 00:14:04.240 --> 00:14:09.037 the the sort of memory of the the the state of your 00:14:09.120 --> 00:14:12.637 cortex and that sort of disappears within 20 00:14:12.720 --> 00:14:16.477 seconds. You can't remember things for more than about 20 seconds or a minute 00:14:16.560 --> 00:14:20.157 if if if you don't have any other form of memory. Uh [snorts] the second type 00:14:20.240 --> 00:14:23.757 of memory which is longer term but still shortterm is the hippoc campus. So you 00:14:23.840 --> 00:14:25.517 can you know you came into this building 00:14:25.600 --> 00:14:27.597 you remember where the where the the 00:14:27.680 --> 00:14:31.037 exit is where the uh elevators are. Um, 00:14:31.120 --> 00:14:34.717 you have some map of that building that's stored in your hippoc campus. 00:14:34.800 --> 00:14:38.477 >> You might remember something about what I said, you know, a few minutes ago. 00:14:38.560 --> 00:14:41.917 >> I forgot it all already. >> It's been erased, but you know, but that 00:14:42.000 --> 00:14:44.397 that would be in your in your hypoc campus. 00:14:44.480 --> 00:14:46.237 >> Uh, and then the the longer term memory 00:14:46.320 --> 00:14:49.517 is in the synapse, the synapses, right? 00:14:49.600 --> 00:14:52.877 >> Um, so what you need if you want a system that's capable of reasoning is 00:14:52.960 --> 00:14:57.277 that you want the hypocampus like thing, right? 00:14:57.360 --> 00:15:00.637 And [snorts] that's what people have tried to do with memory networks and you 00:15:00.720 --> 00:15:02.637 know neural machines and stuff like that 00:15:02.720 --> 00:15:07.677 right and and now with transformers which have sort of a a memory in their 00:15:07.760 --> 00:15:12.797 kind of self attention system you can you can think of it this way so 00:15:12.880 --> 00:15:16.957 so that's one element you need another thing you need is some sort of network 00:15:17.040 --> 00:15:19.037 that can 00:15:19.120 --> 00:15:22.237 access this memory 00:15:22.320 --> 00:15:26.317 get an information back and then kind of crunch on it and 00:15:26.400 --> 00:15:28.077 do this iteratively multiple times 00:15:28.160 --> 00:15:31.037 because a chain of reasoning 00:15:31.120 --> 00:15:34.717 is is a process by which you you you 00:15:34.800 --> 00:15:38.637 kind of you update your knowledge about the state of the world about you know 00:15:38.720 --> 00:15:42.957 what's going to happen etc and that that has to be this sort of recurrent 00:15:43.040 --> 00:15:46.637 operation basically >> and you think that kind of if if we 00:15:46.720 --> 00:15:48.877 think about a transformer so that seems 00:15:48.960 --> 00:15:54.077 to be too small to contain the knowledge that's that's uh to represent the 00:15:54.160 --> 00:15:56.317 knowledge that's contained in Wikipedia media for example 00:15:56.400 --> 00:15:59.997 >> but transformer doesn't have this idea of recurrence it's got a fixed number of 00:16:00.080 --> 00:16:03.597 layers and that's the number of steps that you know limits basically it 00:16:03.680 --> 00:16:07.197 representation >> but recurrence would build on the 00:16:07.280 --> 00:16:09.197 knowledge somehow I mean it would it 00:16:09.280 --> 00:16:13.277 would evolve the knowledge and expand 00:16:13.360 --> 00:16:17.597 the amount of information perhaps or useful information within that knowledge 00:16:17.680 --> 00:16:22.557 >> but is is this something that just can emerge with size because it seems like 00:16:22.640 --> 00:16:26.637 everything we have now is too small no it's not It's not it's not clear I mean 00:16:26.720 --> 00:16:30.957 how how you access and write into an associative memory in efficient way. I 00:16:31.040 --> 00:16:34.157 mean sort of the original memory network maybe had something like the right 00:16:34.240 --> 00:16:38.637 architecture but uh if you try to scale up a memory network so that the memory 00:16:38.720 --> 00:16:42.237 contains all Wikipedia it doesn't quite work >> right 00:16:42.320 --> 00:16:46.397 >> so so there's there's a need for new ideas there okay but it's not the only 00:16:46.480 --> 00:16:50.077 form of reasoning so there's another form of reasoning which is through which 00:16:50.160 --> 00:16:52.877 is very classical also in some types of 00:16:52.960 --> 00:16:56.877 AI uh and it's based on u let's call it 00:16:56.960 --> 00:17:00.717 energy minimization okay so you have uh 00:17:00.800 --> 00:17:02.637 some sort of objective, some energy 00:17:02.720 --> 00:17:06.957 function that represents the the the um 00:17:07.040 --> 00:17:09.917 the quality or the negative quality. 00:17:10.000 --> 00:17:14.557 Okay, energy goes up when things get bad and they get low when things get good. 00:17:14.640 --> 00:17:16.877 So, let's say you you want to figure 00:17:16.960 --> 00:17:18.797 out, you know, what gestures do I need 00:17:18.880 --> 00:17:23.677 to to do to grab an object or walk out 00:17:23.760 --> 00:17:27.117 the door. [snorts] If you have a good model of uh your own 00:17:27.200 --> 00:17:31.437 body, a good model of the environment using this kind of energy minimization, 00:17:31.520 --> 00:17:33.357 you can make a you can make you can do 00:17:33.440 --> 00:17:38.557 planning and it's uh in optimal control is called it's called model model 00:17:38.640 --> 00:17:41.437 predictive control. You have a model of what's going to happen in the world as a 00:17:41.520 --> 00:17:45.757 consequence of your actions and that allows you to by energy minimization 00:17:45.840 --> 00:17:48.477 figure out a sequence of action that optimizes a particular objective 00:17:48.560 --> 00:17:51.517 function which measures you know minimizes number of times you're going 00:17:51.600 --> 00:17:55.437 to hit something and the energy you're going to spend doing the gesture and 00:17:55.520 --> 00:17:59.597 etc. So so that's a form of reasoning. 00:17:59.680 --> 00:18:00.877 Planning is a form of reasoning. And 00:18:00.960 --> 00:18:03.837 perhaps what led to the ability of 00:18:03.920 --> 00:18:07.837 humans to reason is the fact that or you 00:18:07.920 --> 00:18:10.397 know species you know that appear before 00:18:10.480 --> 00:18:12.557 us had to do some sort of planning to be 00:18:12.640 --> 00:18:17.677 able to hunt and survive and survive the winter in particular and so you know 00:18:17.760 --> 00:18:19.277 it's the same capacity that that you 00:18:19.360 --> 00:18:24.637 need to have. So in your intuition is um 00:18:24.720 --> 00:18:27.197 if we look at expert systems and 00:18:27.280 --> 00:18:30.637 encoding knowledge as logic systems 00:18:30.720 --> 00:18:34.477 as graphs or in this kind of way is not 00:18:34.560 --> 00:18:37.437 a useful way to think about knowledge. 00:18:37.520 --> 00:18:40.237 >> Graphs are a little brittle or or logic 00:18:40.320 --> 00:18:45.437 uh representation. So basically you know variables that that have values and then 00:18:45.520 --> 00:18:48.877 constraint between them that are represented by rules uh is a little too 00:18:48.960 --> 00:18:52.717 rigid and too brittle right. So one of the you know some of the early efforts 00:18:52.800 --> 00:18:54.717 in that respect 00:18:54.800 --> 00:18:57.997 um were were to put probabilities on 00:18:58.080 --> 00:19:01.997 them. So a rule you know you know if you have this and that symptom you know you 00:19:02.080 --> 00:19:05.277 have this uh disease with that probability and you should prescribe 00:19:05.360 --> 00:19:07.437 that antibiotic with that probability 00:19:07.520 --> 00:19:09.677 right this the m system from the from 00:19:09.760 --> 00:19:13.037 the 70s um and that that's what that 00:19:13.120 --> 00:19:16.317 branch of AI led to uh you know basics 00:19:16.400 --> 00:19:19.037 and graphical models and causal 00:19:19.120 --> 00:19:21.437 inference and viral you know 00:19:21.520 --> 00:19:26.237 >> method so so there there is I mean certainly 00:19:26.320 --> 00:19:28.157 a lot of interesting work going on in 00:19:28.240 --> 00:19:29.757 this area. The main issue with this is 00:19:29.840 --> 00:19:36.077 is knowledge acquisition. How do you uh reduce a bunch of data to a graph of 00:19:36.160 --> 00:19:39.357 this type? >> Yeah. It relies on the expert to on a 00:19:39.440 --> 00:19:42.157 human being to encode to add knowledge 00:19:42.240 --> 00:19:44.317 >> and that's essentially impractical. 00:19:44.400 --> 00:19:47.757 >> Yeah, it's not scalable. >> That's that's a big question. The second 00:19:47.840 --> 00:19:52.637 question is do you want to represent knowledge as symbols and do you want to 00:19:52.720 --> 00:19:54.717 manipulate them with logic? And again 00:19:54.800 --> 00:19:58.557 that's incompatible with learning. So uh 00:19:58.640 --> 00:20:00.637 one suggestion which Jeffington has been 00:20:00.720 --> 00:20:03.117 advocating for many decades is replace 00:20:03.200 --> 00:20:07.117 symbols by uh vectors. Think of it as 00:20:07.200 --> 00:20:09.597 pattern of activities in a bunch of 00:20:09.680 --> 00:20:14.797 neurons or units or whatever you want to call them and replace logic by 00:20:14.880 --> 00:20:16.717 continuous functions. >> Mhm. 00:20:16.800 --> 00:20:19.037 >> Okay. And that becomes now compatible. 00:20:19.120 --> 00:20:23.117 There's a very good set of ideas by uh 00:20:23.200 --> 00:20:25.197 written in a paper about 10 years ago by 00:20:25.280 --> 00:20:29.837 Leon Bou on who is here at Facebook. Um 00:20:29.920 --> 00:20:32.957 um the title of the paper is from machine learning to machine reasoning 00:20:33.040 --> 00:20:37.197 and his idea is that uh a learning learning system should be able to 00:20:37.280 --> 00:20:41.197 manipulate objects that are in the same spa in a space and then put the result 00:20:41.280 --> 00:20:45.437 back in the same space. So it's this idea of working memory basically [snorts] 00:20:45.520 --> 00:20:47.757 and it's uh it's very enlightening 00:20:47.840 --> 00:20:50.477 >> and in a sense that might learn 00:20:50.560 --> 00:20:53.277 something like the simple expert 00:20:53.360 --> 00:20:57.837 systems. I mean it's you can learn basic logic 00:20:57.920 --> 00:21:00.797 operations there. >> Yeah, quite possibly. Yeah, there's a 00:21:00.880 --> 00:21:04.797 big debate on sort of how much prior structure you have to put in for this 00:21:04.880 --> 00:21:08.077 kind of stuff to emerge. That's the debate I have with Gary Marcus and 00:21:08.160 --> 00:21:12.717 people like that. >> Yeah. Yeah. So and the other person so I 00:21:12.800 --> 00:21:14.637 just talked to Judea Pearl 00:21:14.720 --> 00:21:16.397 >> from you mentioned causal inference 00:21:16.480 --> 00:21:20.397 world. So his worry is that the current 00:21:20.480 --> 00:21:25.837 neural networks are not able to learn 00:21:25.920 --> 00:21:29.917 what causes what causal inference between things. 00:21:30.000 --> 00:21:32.077 >> So I think I think he's right and wrong 00:21:32.160 --> 00:21:34.717 about this. uh if he's talking about the 00:21:34.800 --> 00:21:39.197 sort of classic uh type of neural nets, people sort of 00:21:39.280 --> 00:21:42.317 didn't worry too much about this, but there's a lot of people now working on 00:21:42.400 --> 00:21:45.757 calls on inference and there's a paper that just came out last week by Leonu 00:21:45.840 --> 00:21:47.517 among others, David Lopez and a bunch of 00:21:47.600 --> 00:21:51.197 other people exactly on that problem of 00:21:51.280 --> 00:21:54.237 um how do you kind of you know get a 00:21:54.320 --> 00:21:56.957 neural net to sort of pay attention to 00:21:57.040 --> 00:21:59.597 real causal relationships which may also 00:21:59.680 --> 00:22:04.317 solve issues of bias. in data and things like 00:22:04.400 --> 00:22:06.957 this. So >> I' I'd like to read that paper because 00:22:07.040 --> 00:22:09.037 uh that ultimately the challenge there 00:22:09.120 --> 00:22:12.557 is also seems to fall back on the human 00:22:12.640 --> 00:22:17.677 expert to ultimately decide causality 00:22:17.760 --> 00:22:20.397 between things. >> People are not very good at establishing 00:22:20.480 --> 00:22:24.477 causality first of all. So first of all you talk to physicists and physicists 00:22:24.560 --> 00:22:27.677 actually don't believe in causality because look at the 00:22:27.760 --> 00:22:32.637 >> all the basic laws of microfysics are time reversible. So there is no causality. 00:22:32.720 --> 00:22:34.317 >> The arrow of time is not real. 00:22:34.400 --> 00:22:35.917 >> It's it's as soon as you start looking 00:22:36.000 --> 00:22:38.077 at microscopic systems where there is 00:22:38.160 --> 00:22:43.037 unpredictable randomness where there is clearly an arrow of time but it's a big 00:22:43.120 --> 00:22:46.237 mystery in physics actually how that emerges. Um 00:22:46.320 --> 00:22:49.437 >> is it emergent or is it part of the 00:22:49.520 --> 00:22:54.397 fundamental fabric of reality? Yeah. Or is it a bias of intelligent systems that 00:22:54.480 --> 00:22:57.917 you know because of the second law of thermodynamics we perceive a particular 00:22:58.000 --> 00:23:02.237 hour of time but in fact it's kind of arbitrary right 00:23:02.320 --> 00:23:06.557 >> so yeah physicists mathematicians they don't care about I mean the math doesn't 00:23:06.640 --> 00:23:10.237 care about the flow of time >> well certainly certainly micro physics 00:23:10.320 --> 00:23:12.477 doesn't people themselves are not very 00:23:12.560 --> 00:23:15.597 good at establishing causal causal 00:23:15.680 --> 00:23:17.997 relationships if you ask this I think it 00:23:18.080 --> 00:23:21.757 was in one of simar paper's book on on 00:23:21.840 --> 00:23:26.397 um like children learning you know he studied with Jean Pia you know he's the 00:23:26.480 --> 00:23:29.757 guy who co-authored [snorts] the book perceptron with Marvin Minsky that kind 00:23:29.840 --> 00:23:31.357 of killed the first wave of neural nets 00:23:31.440 --> 00:23:34.077 but but he was actually a a learning 00:23:34.160 --> 00:23:39.117 person he uh in the sense of studying learning in humans and and machines 00:23:39.200 --> 00:23:40.077 that's why he got interested in 00:23:40.160 --> 00:23:43.917 perceptron and uh and he he he wrote 00:23:44.000 --> 00:23:47.277 that if you ask a little kid uh about 00:23:47.360 --> 00:23:51.037 what is the cause of of the wind a lot 00:23:51.120 --> 00:23:54.317 of kids will say they will think for a while and they'll say oh it's the the 00:23:54.400 --> 00:23:57.837 branches in the trees they move and that creates wind right so they get the 00:23:57.920 --> 00:24:01.517 causal relationship backwards and it's because their understanding of the world 00:24:01.600 --> 00:24:04.397 and intuitive physics is not that great right I mean these are like you know 00:24:04.480 --> 00:24:08.317 four or five year old kids >> you know it gets better and then you 00:24:08.400 --> 00:24:11.117 understand that this it can't be right 00:24:11.200 --> 00:24:14.637 but there are many things which we can 00:24:14.720 --> 00:24:19.357 because of our common sense understanding of things what people call 00:24:19.440 --> 00:24:23.837 common sense >> and with understanding of physics we can 00:24:23.920 --> 00:24:26.877 >> there's a lot of stuff that we can figure out causality even with diseases 00:24:26.960 --> 00:24:30.717 we can figure out what's not causing what 00:24:30.800 --> 00:24:33.197 >> often there's a lot of mystery of course 00:24:33.280 --> 00:24:37.757 but the idea is that you should be able to encode that into systems it seems 00:24:37.840 --> 00:24:39.917 unlikely they'd be able to figure that out themselves 00:24:40.000 --> 00:24:43.277 >> well whenever we can do intervention but you know all of humanity has been 00:24:43.360 --> 00:24:45.517 completely deluded for millennia 00:24:45.600 --> 00:24:50.797 probably since existence about a very very wrong causal relationship where 00:24:50.880 --> 00:24:55.837 whatever you can explain you attribute it to you know some deity some divinity right 00:24:55.920 --> 00:24:59.197 >> um and that's a copout that's a way of saying like I don't know the cause so 00:24:59.280 --> 00:25:01.517 you know God did it right so you 00:25:01.600 --> 00:25:05.037 mentioned Marvin Minsky and uh the irony 00:25:05.120 --> 00:25:07.277 of um 00:25:07.360 --> 00:25:10.477 uh you know maybe causing the first AI 00:25:10.560 --> 00:25:15.437 winter you were there in the '9s you were there in the 80s of course uh in 00:25:15.520 --> 00:25:17.437 the '9s why do you think people lost 00:25:17.520 --> 00:25:22.637 faith in deep learning in the '9s and found it again a decade later, over a 00:25:22.720 --> 00:25:24.637 decade later. >> Yeah, it wasn't called deep learning 00:25:24.720 --> 00:25:28.717 yet. It was just called neural nets, but um 00:25:28.800 --> 00:25:33.997 >> yeah, they lost interest. I mean, I think I would put that around 1995, at 00:25:34.080 --> 00:25:35.277 least the machine learning community. 00:25:35.360 --> 00:25:39.357 There was always a neural net community, but it but became 00:25:39.440 --> 00:25:41.837 kind of disconnected from sort of 00:25:41.920 --> 00:25:45.357 mainstream machine learning if you want. 00:25:45.440 --> 00:25:49.997 There were it was basically electrical engineering that kept at it, >> right? 00:25:50.080 --> 00:25:52.797 >> Uh and computer science >> just gave up 00:25:52.880 --> 00:25:56.077 >> gave up on uh on neural nets. I don't I 00:25:56.160 --> 00:25:57.517 don't know, you know, I was too close to 00:25:57.600 --> 00:26:00.077 it to really 00:26:00.160 --> 00:26:03.357 sort of analyze it with sort of a 00:26:03.440 --> 00:26:07.197 unbiased eye if you want. But I I would I would I would I would make a few 00:26:07.280 --> 00:26:10.317 guesses. So the first one is at the time 00:26:10.400 --> 00:26:12.397 neural nets 00:26:12.480 --> 00:26:14.957 were it was very hard to make them work 00:26:15.040 --> 00:26:18.077 in the sense that you would you know 00:26:18.160 --> 00:26:20.397 implement back prop in your favorite 00:26:20.480 --> 00:26:25.517 language um and that favorite language was not Python it was not matlab it was 00:26:25.600 --> 00:26:26.877 not any of those things because they 00:26:26.960 --> 00:26:29.197 didn't exist right you had to write it 00:26:29.280 --> 00:26:33.437 in forc or something like this right um 00:26:33.520 --> 00:26:37.357 so you would uh experiment with it you would probably make some very basic 00:26:37.440 --> 00:26:40.957 mistakes takes like you know badly initialize your weights make the network 00:26:41.040 --> 00:26:43.277 too small because you read in the textbook you know you don't want too 00:26:43.360 --> 00:26:46.637 many parameters right and of course you know and you would train on exor because 00:26:46.720 --> 00:26:49.997 you didn't have any other data set to train it on and of course you know it 00:26:50.080 --> 00:26:52.797 works half the time so you' say I give 00:26:52.880 --> 00:26:57.837 up also you train it with batch gradient which you know isn't efficient so 00:26:57.920 --> 00:27:01.437 there's a lot of there's bag of tricks that you had to know to make those 00:27:01.520 --> 00:27:03.997 things work or you you had to reinvent 00:27:04.080 --> 00:27:08.557 and a lot of people just didn't and they just couldn't make it work. Um, so 00:27:08.640 --> 00:27:10.717 that's one thing. The investment in 00:27:10.800 --> 00:27:15.837 software platform to be able to kind of, you know, display things, figure out why 00:27:15.920 --> 00:27:19.197 things don't work, kind of get a good intuition for how to get them to work, 00:27:19.280 --> 00:27:22.637 have enough flexibility so you can create, you know, network architectures 00:27:22.720 --> 00:27:26.717 like convolutional nets and stuff like that. It was hard. I mean, you had to 00:27:26.800 --> 00:27:29.597 write everything from scratch and again you didn't have any Python or MATLAB or 00:27:29.680 --> 00:27:32.957 anything, right? >> [snorts] >> I read that, sorry to interrupt, but I 00:27:33.040 --> 00:27:35.997 read that you wrote in in lisp the your 00:27:36.080 --> 00:27:40.477 first uh versions of uh Lynette with convolutional networks, which by the 00:27:40.560 --> 00:27:42.557 way, one of my favorite languages. 00:27:42.640 --> 00:27:46.877 >> That's how I knew you were legit. Uh touring award, whatever. This what you 00:27:46.960 --> 00:27:50.557 programmed in Lisp. That's >> it's still my favorite language. But 00:27:50.640 --> 00:27:53.597 >> it's not that we programmed in Lisp, it's that we had to write a list 00:27:53.680 --> 00:27:58.317 interpreter. >> Okay? Because it's not like we used one that existed. So 00:27:58.400 --> 00:28:02.797 >> we wrote a list interpreter that we hooked up to you know a backend library 00:28:02.880 --> 00:28:07.117 that we wrote also for sort of neural net computation. And then after a few 00:28:07.200 --> 00:28:12.317 years around 1991 we invented this idea of uh basically having modules that know 00:28:12.400 --> 00:28:15.197 how to forward propagate and back propagate gradients and then 00:28:15.280 --> 00:28:17.037 interconnecting those modules in a 00:28:17.120 --> 00:28:22.637 graph. [snorts] Leotu had made proposals on this about this in the late 80s and 00:28:22.720 --> 00:28:26.557 we're able to implement this using a list system. Eventually we wanted to use 00:28:26.640 --> 00:28:30.637 that system to make build production code for character 00:28:30.720 --> 00:28:33.597 recognition at Bell Lab. So we actually wrote a compiler for that list 00:28:33.680 --> 00:28:37.837 interpreter so that Patricy Mart who's now Microsoft kind of did did the bulk 00:28:37.920 --> 00:28:40.557 of it with Leon and and me and uh and so 00:28:40.640 --> 00:28:45.037 we could write our system in list and then compile to C and then we'll have a 00:28:45.120 --> 00:28:47.277 self-contained complete system that 00:28:47.360 --> 00:28:50.557 could kind of do the entire thing. 00:28:50.640 --> 00:28:55.277 Neither PyTorch nor transform can do this today yet. Okay, it's coming. 00:28:55.360 --> 00:28:58.797 >> Yeah, >> I mean there's something like that in 00:28:58.880 --> 00:29:01.757 PyTorch called you know to script 00:29:01.840 --> 00:29:03.757 >> and so you know we had to write all this 00:29:03.840 --> 00:29:07.437 interpreter. to this compiler. where to invest a huge amount of effort to do 00:29:07.520 --> 00:29:12.477 this and not everybody if if you don't completely believe in the concept you're 00:29:12.560 --> 00:29:14.397 not going to invest the time to do this right 00:29:14.480 --> 00:29:17.677 >> now at the time also you know or today 00:29:17.760 --> 00:29:21.517 >> this would turn into torture pietor torture tensorflow or whatever we put it 00:29:21.600 --> 00:29:22.957 in open source everybody would use it 00:29:23.040 --> 00:29:25.597 and you know realize it's good back 00:29:25.680 --> 00:29:28.957 before 1995 working at AT&T there's no 00:29:29.040 --> 00:29:31.357 way the lawyers would let you release 00:29:31.440 --> 00:29:36.477 anything in open source of this nature And so we could not distribute our code 00:29:36.560 --> 00:29:39.997 really. >> And on that point and sorry to go on a 00:29:40.080 --> 00:29:42.237 million tangents but on that point I 00:29:42.320 --> 00:29:47.837 also read that there was some almost pat like a patent on convolution networks 00:29:47.920 --> 00:29:51.037 labs. >> [sighs and gasps] >> So that 00:29:51.120 --> 00:29:55.117 >> first of all I mean just >> there's two actually [laughter] 00:29:55.200 --> 00:29:57.197 >> uh that ran out uh the that 00:29:57.280 --> 00:29:59.197 >> thankfully in 2007 00:29:59.280 --> 00:30:03.917 >> in 2007 I what 00:30:04.000 --> 00:30:07.357 can we can we just talk about that for I know you're a Facebook but you're also 00:30:07.440 --> 00:30:09.917 at NYU and and 00:30:10.000 --> 00:30:12.877 what does it mean to patent ideas like 00:30:12.960 --> 00:30:14.397 these 00:30:14.480 --> 00:30:20.077 software ideas essentially or what or mathematical ical ideas or what are 00:30:20.160 --> 00:30:22.877 they? >> Okay, so they're not mathematical ideas. 00:30:22.960 --> 00:30:24.797 There are, you know, algorithms and 00:30:24.880 --> 00:30:27.917 there was a period where the US patent 00:30:28.000 --> 00:30:30.637 office would allow the patent of 00:30:30.720 --> 00:30:35.517 software as long as it was embodied. Uh the Europeans are very different. They 00:30:35.600 --> 00:30:40.077 don't they don't quite accept that. They have a different concept. But you know, 00:30:40.160 --> 00:30:43.677 I don't I no longer I mean I never actually strongly believed in this, but 00:30:43.760 --> 00:30:45.117 I I don't believe in this kind of 00:30:45.200 --> 00:30:51.000 patent. Facebook basically doesn't believe in this kind of patent. 00:30:51.280 --> 00:30:55.037 Google fights patents because they've been 00:30:55.120 --> 00:30:59.277 burned with Apple. And so now they do this for defensive purpose. But usually 00:30:59.360 --> 00:31:02.157 they say we're not going to sue you if you infringe. 00:31:02.240 --> 00:31:04.397 >> Facebook has a has a similar policy. 00:31:04.480 --> 00:31:08.077 They say, you know, we patent on certain things for defensive purpose. We're not 00:31:08.160 --> 00:31:09.677 going to see you if you infringe unless 00:31:09.760 --> 00:31:15.117 you sue us. Um so the the industry does 00:31:15.200 --> 00:31:17.117 not believe in in patents. They're there 00:31:17.200 --> 00:31:22.557 because of you know the legal landscape and and and various things but but I 00:31:22.640 --> 00:31:24.877 don't really believe in patents for this kind of stuff. 00:31:24.960 --> 00:31:27.517 >> Okay. So that's that's a great thing. So I I 00:31:27.600 --> 00:31:29.197 >> I'll tell you a worst story actually. So 00:31:29.280 --> 00:31:33.517 what happens was the the first the first patent about convolutional net was about 00:31:33.600 --> 00:31:35.997 kind of the early version of convolutional net that didn't have 00:31:36.080 --> 00:31:40.717 separate pooling layers. It had convolutional layers which tried more 00:31:40.800 --> 00:31:44.877 than one if you want right >> and then there was a second one on 00:31:44.960 --> 00:31:46.877 convolutional nets with separate pooling 00:31:46.960 --> 00:31:50.077 layers uh train with back prop and they 00:31:50.160 --> 00:31:52.637 were filed filed in ' 89 and 1990 00:31:52.720 --> 00:31:57.757 something like this at the time the life life of a patent was 17 years so here's 00:31:57.840 --> 00:32:01.757 what what happened over the next few years is that we started developing um 00:32:01.840 --> 00:32:04.317 character recognition technology around 00:32:04.400 --> 00:32:09.277 conviction nets [snorts] and uh in 1994 00:32:09.360 --> 00:32:11.837 a check reading system was deployed in 00:32:11.920 --> 00:32:15.357 ATM machines. In 1995 it was for large 00:32:15.440 --> 00:32:20.637 check reading machines in back offices etc. And those systems were 00:32:20.720 --> 00:32:24.237 developed by an engineering group that we were collaborating with at AT&T and 00:32:24.320 --> 00:32:26.157 they were commercialized by NCR which at 00:32:26.240 --> 00:32:28.717 the time was a subsidiary of AT&T. Mhm. 00:32:28.800 --> 00:32:33.357 >> Now, AT&T split up in 1996, 00:32:33.440 --> 00:32:37.837 199 early 1996 and the lawyers just looked at all the patents and they 00:32:37.920 --> 00:32:42.077 distributed the patents among the various companies. They gave the the 00:32:42.160 --> 00:32:45.517 commercial net patent to NCR because they were actually selling products that 00:32:45.600 --> 00:32:49.517 used it. But nobody at NCR had any idea what a commercial net was. 00:32:49.600 --> 00:32:50.477 >> Yeah. 00:32:50.560 --> 00:32:55.037 >> Okay. So between 1996 and 2007, 00:32:55.120 --> 00:32:58.877 so there's a whole period until 2002 where I didn't actually work on machine 00:32:58.960 --> 00:33:02.957 learning or control it. I resumed working on this around 2002 00:33:03.040 --> 00:33:07.037 and between 2002 and 2007 I was working on them crossing my finger that nobody 00:33:07.120 --> 00:33:09.277 at NCR would notice and nobody noticed. 00:33:09.360 --> 00:33:12.157 >> Yeah. And I and I hope that this kind of 00:33:12.240 --> 00:33:15.517 somewhat as you said lawyers aside 00:33:15.600 --> 00:33:18.797 relative openness of the community now 00:33:18.880 --> 00:33:21.277 will continue. It accelerates the entire 00:33:21.360 --> 00:33:24.957 progress of of the industry. And 00:33:25.040 --> 00:33:29.357 you know the problems that Facebook and Google and others are 00:33:29.440 --> 00:33:34.237 facing today is not whether Facebook or Google or Microsoft or IBM or whoever is 00:33:34.320 --> 00:33:37.357 ahead of the other. It's that we don't have the technology to build the things 00:33:37.440 --> 00:33:40.717 we want to build. We want to build intelligent virtual assistants that have 00:33:40.800 --> 00:33:44.557 common sense. We don't have monopoly on good ideas for this. We don't believe we 00:33:44.640 --> 00:33:47.117 do. Maybe others do believe they do, but 00:33:47.200 --> 00:33:52.717 we don't. Okay. If a startup tells you they have the secret to, you know, human 00:33:52.800 --> 00:33:54.077 level intelligence and common sense, 00:33:54.160 --> 00:33:56.477 don't believe them. They don't. And it's 00:33:56.560 --> 00:34:00.397 going to take the entire uh work of the 00:34:00.480 --> 00:34:04.717 world research community for a while to get to the point where you can go off 00:34:04.800 --> 00:34:08.077 and in each of those companies kind of start to build things on this. We're not 00:34:08.160 --> 00:34:09.037 there yet. 00:34:09.120 --> 00:34:11.117 >> Absolutely. And this this calls to the 00:34:11.200 --> 00:34:15.197 the gap between the space of ideas and 00:34:15.280 --> 00:34:19.757 the rigorous testing of those ideas of practical application that you often 00:34:19.840 --> 00:34:24.157 speak to. You've uh written advice saying don't get fooled by people who 00:34:24.240 --> 00:34:28.317 claim to have a solution to artificial general intelligence who claim to have 00:34:28.400 --> 00:34:30.877 an AI system that work just like the 00:34:30.960 --> 00:34:32.637 human brain or who claim to have figured 00:34:32.720 --> 00:34:35.917 out how the brain works. Ask them what 00:34:36.000 --> 00:34:40.557 the error rate they get on emnest or imageet. Yeah, 00:34:40.640 --> 00:34:42.637 >> this is a little dated by the way [laughter] 00:34:42.720 --> 00:34:45.517 >> 2000 I mean five years who's counting 00:34:45.600 --> 00:34:47.997 okay but I think your opinion is still 00:34:48.080 --> 00:34:52.077 emnest and imageet uh yes maybe data 00:34:52.160 --> 00:34:54.077 there may be new benchmarks right but I 00:34:54.160 --> 00:34:56.797 think that philosophy is one you still 00:34:56.880 --> 00:35:00.397 and and somewhat hold that benchmarks 00:35:00.480 --> 00:35:04.077 and the practical testing the practical application is where you really get to 00:35:04.160 --> 00:35:06.877 test the ideas >> well it may not be completely practical 00:35:06.960 --> 00:35:10.557 like for examp you know it could be a toy data set or but but it to be some 00:35:10.640 --> 00:35:15.197 sort of task that the community as a whole has accepted as some sort of 00:35:15.280 --> 00:35:18.797 standard you know kind of benchmark if you want it doesn't need to be real. So 00:35:18.880 --> 00:35:23.837 for example many years ago here at fair um people you know chisen west on auto 00:35:23.920 --> 00:35:27.597 and board and a few others proposed the the baby tasks which were kind of a a 00:35:27.680 --> 00:35:30.477 toy problem to test the ability of 00:35:30.560 --> 00:35:34.797 machines to reason actually to access working memory and things like this and 00:35:34.880 --> 00:35:38.557 uh it was very useful even though it wasn't a real task. amnest is kind of 00:35:38.640 --> 00:35:41.757 halfway a real task. Uh so you know toy 00:35:41.840 --> 00:35:46.157 problems can be very useful. It's just that I was really struck by the fact 00:35:46.240 --> 00:35:49.997 that a lot of people particularly a lot of people with money to invest would be 00:35:50.080 --> 00:35:54.717 fooled by people telling them oh we have you know the algorithm of the cortex and 00:35:54.800 --> 00:35:56.557 you should give us 50 million. 00:35:56.640 --> 00:36:01.117 >> Yes. Absolutely. So there's a lot of people who 00:36:01.200 --> 00:36:03.437 who try to take advantage of the hype 00:36:03.520 --> 00:36:05.917 for business reasons and so on. But let 00:36:06.000 --> 00:36:08.957 me sort of talk to this idea 00:36:09.040 --> 00:36:11.837 that sort of new ideas the ideas that 00:36:11.920 --> 00:36:16.877 push the field forward may not yet have a benchmark or it may be very difficult 00:36:16.960 --> 00:36:19.597 to establish a benchmark. >> I agree that's part of the process. 00:36:19.680 --> 00:36:21.117 Establishing benchmarks is part of the 00:36:21.200 --> 00:36:24.397 process. So what are your thoughts about 00:36:24.480 --> 00:36:27.437 so we have these benchmarks on around 00:36:27.520 --> 00:36:32.557 stuff we can do with images from classification to captioning to just 00:36:32.640 --> 00:36:35.997 every kind of information you can pull off from images and the surface level 00:36:36.080 --> 00:36:37.997 there's audio data sets there's some 00:36:38.080 --> 00:36:39.997 video 00:36:40.080 --> 00:36:45.037 what can we start natural language what kind of stuff what kind of 00:36:45.120 --> 00:36:46.637 benchmarks do you see that start 00:36:46.720 --> 00:36:49.517 creeping on to more something like 00:36:49.600 --> 00:36:53.757 intelligence like reasoning like maybe 00:36:53.840 --> 00:36:58.637 you don't like the term but AGI echoes of that kind of uh 00:36:58.720 --> 00:37:01.757 >> a lot of people are working on interactive environments in which you 00:37:01.840 --> 00:37:04.397 can you can train and test uh 00:37:04.480 --> 00:37:06.637 intelligent systems so so there for 00:37:06.720 --> 00:37:10.477 example uh you know it's 00:37:10.560 --> 00:37:15.117 the classical paradigm of supervised learning is that you you have a data set 00:37:15.200 --> 00:37:16.637 you partition it into a training set 00:37:16.720 --> 00:37:18.477 validation set test set and there's a 00:37:18.560 --> 00:37:22.237 clear protocol right but What if the 00:37:22.320 --> 00:37:26.477 that assumes that the samples are stat statistically independent? You can 00:37:26.560 --> 00:37:29.277 exchange them. The order in which you see them doesn't shouldn't matter. You 00:37:29.360 --> 00:37:33.917 know, things like that. But what if the answer you give determines the next 00:37:34.000 --> 00:37:37.837 sample you see, which is the case for example in robotics, right? You robot 00:37:37.920 --> 00:37:41.997 does something and then it gets exposed to a new room and depending on where it 00:37:42.080 --> 00:37:45.037 goes, the room will be different. So that's the that creates the exploration 00:37:45.120 --> 00:37:47.277 problem. >> Mhm. 00:37:47.360 --> 00:37:50.717 the what if the samples. So that creates 00:37:50.800 --> 00:37:55.197 also a dependency between samples, right? You you if you move if you can 00:37:55.280 --> 00:37:58.637 only move in in space, the next sample you're going to see is going to be 00:37:58.720 --> 00:38:03.517 probably in the same building most likely, right? So so the all the 00:38:03.600 --> 00:38:08.877 assumptions about the validity of this training set test set hypothesis break 00:38:08.960 --> 00:38:11.757 >> whenever a machine can take an action that has an influence in the in the 00:38:11.840 --> 00:38:13.437 world and it's what it's going to see. 00:38:13.520 --> 00:38:16.157 So people are setting up artificial 00:38:16.240 --> 00:38:18.717 uh environments where where that takes 00:38:18.800 --> 00:38:22.077 place, right? The robot runs around a 3D 00:38:22.160 --> 00:38:26.237 model of a house and can interact with objects and things like this. So you do 00:38:26.320 --> 00:38:28.237 robotics by simulation. You have those, 00:38:28.320 --> 00:38:33.357 you know, open eye gym type thing or muoko kind of 00:38:33.440 --> 00:38:37.197 simulated uh robots and you have games, you know, 00:38:37.280 --> 00:38:40.877 things like that. Um so that that's where the field is going really. um this 00:38:40.960 --> 00:38:43.837 kind of environment. [snorts] Now back 00:38:43.920 --> 00:38:48.397 to the question of AGI like I don't like the term AGI 00:38:48.480 --> 00:38:52.717 because it implies that human intelligence is general 00:38:52.800 --> 00:38:58.077 >> and human intelligence is nothing like general. It's very very specialized. We 00:38:58.160 --> 00:39:00.397 think it's general. We like to think of ourselves as having general 00:39:00.480 --> 00:39:03.117 intelligence. We don't. We're very specialized. 00:39:03.200 --> 00:39:04.637 >> We're only slightly more general than 00:39:04.720 --> 00:39:07.197 >> Why does it feel general? So you kind of 00:39:07.280 --> 00:39:10.397 uh the term general. I think what's 00:39:10.480 --> 00:39:11.997 impressive about humans is this ability 00:39:12.080 --> 00:39:14.397 to learn as we were talking about 00:39:14.480 --> 00:39:17.517 learning to learn in just so many 00:39:17.600 --> 00:39:22.397 different domains. It's perhaps not arbitrarily general but just you can 00:39:22.480 --> 00:39:26.397 learn in many domains and integrate that knowledge somehow. Okay. The knowledge 00:39:26.480 --> 00:39:28.877 persists. >> So let me take a very specific example. 00:39:28.960 --> 00:39:32.237 >> Yes, >> it's not an example. It's more like a a 00:39:32.320 --> 00:39:36.077 quasi mathematical demonstration. So you have about 1 million fibers coming out 00:39:36.160 --> 00:39:39.597 of one of your eyes. Okay. 2 million total. But let's let's talk about just 00:39:39.680 --> 00:39:44.237 one of them. >> It's 1 million nerve fibers. Your optical nerve. 00:39:44.320 --> 00:39:47.837 >> Let's imagine that they are binary. So they can be active or inactive, right? 00:39:47.920 --> 00:39:51.437 So the input to your visual cortex is 1 million bits. 00:39:51.520 --> 00:39:53.997 >> Mhm. 00:39:54.080 --> 00:39:58.477 >> Now they connected to your brain in a particular way. And your brain has 00:39:58.560 --> 00:40:01.597 connections that are kind of a little bit like a conversion net. They're kind 00:40:01.680 --> 00:40:05.037 of local uh you know in space and things like this. 00:40:05.120 --> 00:40:07.997 >> Now imagine I play a trick on you. 00:40:08.080 --> 00:40:12.637 It's a pretty nasty trick, I admit. I uh I cut your optical nerve 00:40:12.720 --> 00:40:14.877 >> and I put a device that makes a random 00:40:14.960 --> 00:40:17.597 perturbation of a permutation of all the 00:40:17.680 --> 00:40:20.717 nerve fibers. So now what comes to your 00:40:20.800 --> 00:40:24.957 to your brain is a fixed but random permutation of all the pixels. 00:40:25.040 --> 00:40:27.757 >> Mhm. >> There's no way in hell that your visual 00:40:27.840 --> 00:40:30.557 cortex, even if I do this to you in 00:40:30.640 --> 00:40:33.757 infancy, will actually learn vision to 00:40:33.840 --> 00:40:37.117 the same level of quality that you can. 00:40:37.200 --> 00:40:39.917 >> Got it. And you're saying there's no way you relearn that. 00:40:40.000 --> 00:40:44.157 >> No, because now two pixels that are nearby in the world will end up in very 00:40:44.240 --> 00:40:47.997 different places in your visual cortex and your neurons there have no 00:40:48.080 --> 00:40:50.557 connections with each other because they only connected locally. 00:40:50.640 --> 00:40:55.597 >> So this whole our entire the hardware is built in many ways to support 00:40:55.680 --> 00:40:57.037 >> the locality of the real world. 00:40:57.120 --> 00:40:59.677 >> Yeah. >> Yes. That's specialization. 00:40:59.760 --> 00:41:02.557 >> Yeah. But it's still pretty damn impressive. So it's not perfect 00:41:02.640 --> 00:41:04.237 generalization. It's not even close. 00:41:04.320 --> 00:41:08.237 >> No. No. It's it's it's it's not that it's not even close. It's not at all. 00:41:08.320 --> 00:41:11.757 >> It's not specialized. >> So how many boolean functions? So let's 00:41:11.840 --> 00:41:14.477 imagine you want to train your visual 00:41:14.560 --> 00:41:17.757 system to uh you know recognize 00:41:17.840 --> 00:41:22.077 particular patterns of those 1 million bits. Okay. So that's a boolean 00:41:22.160 --> 00:41:25.437 function, right? Either the pattern is here or not here. This a two two-way 00:41:25.520 --> 00:41:30.717 classification with 1 million binary inputs. 00:41:30.800 --> 00:41:33.437 How many such boolean functions are there? 00:41:33.520 --> 00:41:35.597 Okay, you have you have two to the 1 00:41:35.680 --> 00:41:38.717 million combinations of inputs. For each 00:41:38.800 --> 00:41:41.437 of those, you have an output bit and so 00:41:41.520 --> 00:41:44.717 you have two to the two to the 1 million 00:41:44.800 --> 00:41:47.117 boolean functions of this type. Okay, 00:41:47.200 --> 00:41:50.157 which is an unimaginably large number. 00:41:50.240 --> 00:41:52.157 How many of those functions can actually 00:41:52.240 --> 00:41:54.477 be computed by your visual cortex? And 00:41:54.560 --> 00:41:59.837 the answer is a tiny tiny tiny tiny tiny tiny sliver. Like an enormously tiny 00:41:59.920 --> 00:42:05.357 sliver. >> Yeah. Yeah. So, we are ridiculously specialized, 00:42:05.440 --> 00:42:08.557 you know. Okay. >> But, [laughter] 00:42:08.640 --> 00:42:10.637 okay, that's an argument against the 00:42:10.720 --> 00:42:12.877 word general. I think there there's a 00:42:12.960 --> 00:42:16.237 >> I I there's I agree with your intuition, 00:42:16.320 --> 00:42:19.117 but I'm not sure it's it seems the bre 00:42:19.200 --> 00:42:23.757 the the brain is impressively 00:42:23.840 --> 00:42:26.637 uh capable of adjusting to things. So 00:42:26.720 --> 00:42:31.677 >> it's because we can't imagine tasks that are outside of our 00:42:31.760 --> 00:42:34.717 comprehension, >> right? So we think we think we're 00:42:34.800 --> 00:42:37.357 general because we're general of all the things that we can apprehend. 00:42:37.440 --> 00:42:40.237 >> So yeah, >> but there is a huge world out there of 00:42:40.320 --> 00:42:43.917 things that we have no idea. We call that heat, by the way. 00:42:44.000 --> 00:42:47.437 >> Heat. >> Heat. So uh at least physicists call 00:42:47.520 --> 00:42:50.477 that heat or they call it entropy which is 00:42:50.560 --> 00:42:53.517 >> you know you have a 00:42:53.600 --> 00:42:56.477 thing full of gas right 00:42:56.560 --> 00:42:57.997 >> closed system full of gas 00:42:58.080 --> 00:43:01.517 >> right closed or no closed it has uh you 00:43:01.600 --> 00:43:05.517 know pressure it has uh temperature it 00:43:05.600 --> 00:43:08.877 has you know and you can write equations 00:43:08.960 --> 00:43:12.557 PV equivalent on T you know things like that right [snorts] 00:43:12.640 --> 00:43:15.517 when you reduce a volume the temperature goes up the pressure goes up, you know, 00:43:15.600 --> 00:43:19.917 things like that, right? Uh for a perfect gas at least, those are the 00:43:20.000 --> 00:43:24.237 things you can know about that system and it's a tiny tiny number of bits 00:43:24.320 --> 00:43:28.237 compared to the complete information of the state of the entire system because 00:43:28.320 --> 00:43:29.677 the state of the entire system will give 00:43:29.760 --> 00:43:31.997 you the position and momentum of every 00:43:32.080 --> 00:43:35.517 every molecule of the gas. 00:43:35.600 --> 00:43:37.517 And what you don't know about it is the 00:43:37.600 --> 00:43:41.117 entropy and you interpret it as heat. 00:43:41.200 --> 00:43:44.237 the energy contained in that thing is is 00:43:44.320 --> 00:43:46.877 what we call heat. Now, 00:43:46.960 --> 00:43:51.357 it's very possible that in fact there is some very strong structure in how those 00:43:51.440 --> 00:43:55.277 molecules are moving. It's just that they are in a way that we're just not 00:43:55.360 --> 00:43:59.037 wired to perceive. >> Yeah, we're ignorant to it and there's 00:43:59.120 --> 00:44:02.637 in near infinite amount of things we're not wired to perceive. 00:44:02.720 --> 00:44:03.917 >> And you're right, that's a that's a nice 00:44:04.000 --> 00:44:05.677 way to put it. We're general to all the 00:44:05.760 --> 00:44:11.037 things we can imagine which is a very tiny subset of all 00:44:11.120 --> 00:44:13.517 things that are possible. >> So it's like chromograph complexity or 00:44:13.600 --> 00:44:16.877 the chromograph chhattin of complexity. 00:44:16.960 --> 00:44:19.437 >> Yeah. >> You know every 00:44:19.520 --> 00:44:22.717 bit string or every integer is random 00:44:22.800 --> 00:44:27.900 except for all the ones that you can actually write down. [laughter] 00:44:28.800 --> 00:44:31.997 >> Yeah. Okay. So beautifully put but you know so we can just call it artificial 00:44:32.080 --> 00:44:35.117 intelligence. We don't need to have a general 00:44:35.200 --> 00:44:37.997 >> or human level human level intelligence is 00:44:38.080 --> 00:44:41.117 >> you know you you'll start anytime you 00:44:41.200 --> 00:44:43.757 touch human it gets it gets interesting 00:44:43.840 --> 00:44:46.097 because um 00:44:46.180 --> 00:44:50.877 [snorts] uh you know it's it's because we attach ourselves to human and it's 00:44:50.960 --> 00:44:53.197 difficult to define what human intelligence is. 00:44:53.280 --> 00:44:54.397 >> Yeah. 00:44:54.480 --> 00:44:57.117 >> Nevertheless, my definition is uh maybe 00:44:57.200 --> 00:45:01.037 uh demo impressive intelligence. Okay. 00:45:01.120 --> 00:45:02.557 Damn impressive demonstration of 00:45:02.640 --> 00:45:04.957 intelligence whatever and so on that 00:45:05.040 --> 00:45:08.557 topic most successes in deep learning 00:45:08.640 --> 00:45:11.037 have been in supervised learning. 00:45:11.120 --> 00:45:14.637 >> What is your view on unsupervised 00:45:14.720 --> 00:45:17.757 learning? Is there hope to reduce 00:45:17.840 --> 00:45:20.877 involvement of human input and still 00:45:20.960 --> 00:45:25.437 have uh successful systems that are have practical use? 00:45:25.520 --> 00:45:26.877 >> Yeah, I mean there's definitely a hope. 00:45:26.960 --> 00:45:28.397 It's it's more than a hope actually. 00:45:28.480 --> 00:45:33.277 it's uh it's you know mounting evidence for it and that's basically all I do 00:45:33.360 --> 00:45:37.357 like the only thing I'm interested in at the moment is I call it self-supervised 00:45:37.440 --> 00:45:40.717 learning not unsupervised because unsupervised learning is is a loaded 00:45:40.800 --> 00:45:44.637 term people who know something about machine 00:45:44.720 --> 00:45:47.837 learning you know tell you so you're doing clustering or PCA 00:45:47.920 --> 00:45:49.997 >> which is not the case and the wide public you know when you say 00:45:50.080 --> 00:45:52.397 unsupervised learning oh my god you know machines are going to learn by 00:45:52.480 --> 00:45:56.477 themselves and without supervision you know they see this as 00:45:56.560 --> 00:46:00.237 >> where's where's the parents Yeah. So, so I call it self-supervised learning 00:46:00.320 --> 00:46:03.917 because in fact the underlying algorithms that are used are the same 00:46:04.000 --> 00:46:08.637 algorithms as the supervised learning algorithms. Um, except that what we 00:46:08.720 --> 00:46:10.637 train them to do is not predict a 00:46:10.720 --> 00:46:13.277 particular set of variables like the 00:46:13.360 --> 00:46:17.597 category of of an image 00:46:17.680 --> 00:46:20.957 and and not to predict a set of variables that have been provided by 00:46:21.040 --> 00:46:24.237 human labelers. But what you train the 00:46:24.320 --> 00:46:26.637 machine to do is basically reconstruct a 00:46:26.720 --> 00:46:29.437 piece of its input that it's being 00:46:29.520 --> 00:46:31.357 is being maxed maxed out essentially. 00:46:31.440 --> 00:46:36.157 You can think of it this way, right? So show a piece of video to a machine and 00:46:36.240 --> 00:46:39.597 ask it to predict what's going to happen next. And of course after a while you 00:46:39.680 --> 00:46:43.677 can show what what happens and the machine will kind of train itself to do 00:46:43.760 --> 00:46:48.477 better at that task. You can do like all the latest most 00:46:48.560 --> 00:46:53.277 successful models in natural language processing use self-supervised learning. 00:46:53.360 --> 00:46:55.197 you know, sort of BERT style systems, 00:46:55.280 --> 00:46:56.877 for example, right? You show it a window 00:46:56.960 --> 00:47:00.797 of a thousand words on a test corpus. 00:47:00.880 --> 00:47:03.757 You take out 15% of the words and then 00:47:03.840 --> 00:47:08.477 you train a machine to predict the words that are missing. 00:47:08.560 --> 00:47:11.597 >> That's self-supervised learning. It's not predicting the future. It's just, 00:47:11.680 --> 00:47:14.557 you know, predicting things in the middle, but you could have it predict 00:47:14.640 --> 00:47:16.717 the future. That's what language models do. 00:47:16.800 --> 00:47:18.797 >> So, you construct, so in an unsupervised 00:47:18.880 --> 00:47:21.277 way, you construct a model of language. 00:47:21.360 --> 00:47:25.437 Do you think >> or video or the physical world or 00:47:25.520 --> 00:47:29.757 whatever? Right. >> How far do you think that can take us? 00:47:29.840 --> 00:47:31.837 Do you think >> art 00:47:31.920 --> 00:47:35.117 >> understands anything? 00:47:35.200 --> 00:47:37.597 >> To some level it has you know a shallow 00:47:37.680 --> 00:47:41.117 understanding of uh of text but it needs 00:47:41.200 --> 00:47:44.637 to I mean to have kind of true human level intelligence I think you need to 00:47:44.720 --> 00:47:47.357 ground language in reality. So some 00:47:47.440 --> 00:47:51.757 people are attempting to do this right having systems that kind of have some 00:47:51.840 --> 00:47:55.437 visual representation of what what is being talked about which is one reason 00:47:55.520 --> 00:47:57.117 you need those interactive environments 00:47:57.200 --> 00:47:59.757 actually but this is like a huge 00:47:59.840 --> 00:48:04.877 technical problem that is not solved and that explains why self-supervised 00:48:04.960 --> 00:48:08.077 learning works in the context of natural language but does not work in the 00:48:08.160 --> 00:48:12.477 context or at least not well in the context of image recognition and video 00:48:12.560 --> 00:48:17.117 although it's making progress quickly and [snorts] the reason that reason is 00:48:17.200 --> 00:48:20.397 the fact that it's much easier to represent 00:48:20.480 --> 00:48:24.477 uncertainty in the prediction in a context of natural language than it is 00:48:24.560 --> 00:48:27.277 in the context of things like video and images. 00:48:27.360 --> 00:48:30.077 >> So for example, if I ask you to predict 00:48:30.160 --> 00:48:34.797 what words are missing, you know, 15% of the words that are taken out, 00:48:34.880 --> 00:48:38.237 >> the possibilities are small. That means >> it's small, right? There is 100,000 00:48:38.320 --> 00:48:41.277 words in the in the lexicon and what the 00:48:41.360 --> 00:48:46.077 machine spits out is a a big probability vector, right? It's a bunch of numbers 00:48:46.160 --> 00:48:49.757 between 0 and one that's some to one and we know how to do how to do this with 00:48:49.840 --> 00:48:53.677 computers. [snorts] So there representing uncertainty in the 00:48:53.760 --> 00:48:55.597 prediction is relatively easy and that's 00:48:55.680 --> 00:48:58.397 in my opinion why those techniques work 00:48:58.480 --> 00:49:02.877 for NLP for images if you ask if you 00:49:02.960 --> 00:49:06.077 block a piece of an image and you ask a system reconstruct that piece of the 00:49:06.160 --> 00:49:08.877 image there are many possible answers 00:49:08.960 --> 00:49:12.717 that are all perfectly legit right and 00:49:12.800 --> 00:49:16.957 how do you represent that this set of possible answers you you can't train a 00:49:17.040 --> 00:49:20.317 system to make one prediction you can't train a neural net to say here it is 00:49:20.400 --> 00:49:23.357 that's the image because it's there's a whole set of 00:49:23.440 --> 00:49:25.917 things that are compatible with it. So how do you get the machine to represent 00:49:26.000 --> 00:49:27.597 not a single output but a whole set of 00:49:27.680 --> 00:49:30.237 outputs 00:49:30.320 --> 00:49:33.277 and you know similarly with uh video 00:49:33.360 --> 00:49:37.277 prediction there's a lot of things that can happen in the future of a video 00:49:37.360 --> 00:49:40.717 you're looking at me right now I'm not moving my head very much but you know I 00:49:40.800 --> 00:49:43.997 might you know turn my my head to the left or to the right >> right 00:49:44.080 --> 00:49:47.437 >> if you don't have a system that can predict this 00:49:47.520 --> 00:49:50.557 >> and you train it with le square to kind of minimize the error with the 00:49:50.640 --> 00:49:54.077 prediction and what I'm doing what you get is a blurry image of myself right 00:49:54.160 --> 00:49:57.437 >> in all possible future positions that I might be in which is not a good 00:49:57.520 --> 00:49:59.837 prediction. But so there might be other 00:49:59.920 --> 00:50:02.797 ways to do the self-supervision, right, 00:50:02.880 --> 00:50:06.877 for visual scenes. >> Like what? [laughter] 00:50:06.960 --> 00:50:09.117 >> If I to I mean if I knew I wouldn't tell 00:50:09.200 --> 00:50:12.637 you I publish it first. I don't know. I 00:50:12.720 --> 00:50:17.597 >> No, there might be. [laughter] >> So I mean these are kind of uh there 00:50:17.680 --> 00:50:20.877 might be artificial ways of like selfplay in games to where you can 00:50:20.960 --> 00:50:22.877 simulate part of the environment. You can 00:50:22.960 --> 00:50:24.157 >> Oh, that doesn't solve the problem. It's 00:50:24.240 --> 00:50:27.677 just a way of generating data. 00:50:27.760 --> 00:50:32.317 But because you have more of a control mean you can control. Yeah. It's a way 00:50:32.400 --> 00:50:36.157 to generate data. That's right. And because you can do huge amounts of data 00:50:36.240 --> 00:50:38.477 generation that doesn't you're right. 00:50:38.560 --> 00:50:43.517 This well it's it's a creeps up on the problem from the side of data and you 00:50:43.600 --> 00:50:44.957 don't think that's the right way to creep up. 00:50:45.040 --> 00:50:47.837 >> It doesn't solve this problem of handling uncertainty in the world. 00:50:47.920 --> 00:50:51.037 Right. So if you if you have a machine 00:50:51.120 --> 00:50:55.757 learn a predictive model of the world in a game that is deterministic or quasi 00:50:55.840 --> 00:50:57.517 deterministic 00:50:57.600 --> 00:51:00.637 uh it's easy right just you know give a 00:51:00.720 --> 00:51:05.117 few frames of the game to a connet put a bunch of layers and then have the game 00:51:05.200 --> 00:51:06.877 generates the next few frames 00:51:06.960 --> 00:51:09.117 >> and and if the game is deterministic it 00:51:09.200 --> 00:51:13.197 works fine. Um, and that includes, you 00:51:13.280 --> 00:51:18.397 know, feeding the system with the action that your little character is is going 00:51:18.480 --> 00:51:23.277 to take the problem comes from the fact that the 00:51:23.360 --> 00:51:27.197 real world and certain most games are not entirely predictable. And so there 00:51:27.280 --> 00:51:31.357 you get those blurry predictions and you can't do planning with blurry predictions. 00:51:31.440 --> 00:51:33.117 Right? So if you have a perfect model of 00:51:33.200 --> 00:51:37.117 the world, you can in your head run this 00:51:37.200 --> 00:51:42.157 model with a hypothesis for a sequence of actions. Are you going to predict the 00:51:42.240 --> 00:51:43.997 outcome of that sequence of actions? 00:51:44.080 --> 00:51:48.397 >> Mhm. But if your model is imperfect, 00:51:48.480 --> 00:51:51.917 how can you plan? >> Yeah. It quickly explodes. 00:51:52.000 --> 00:51:56.477 What are your thoughts on the extension of this which topic I'm super excited 00:51:56.560 --> 00:51:59.917 about? It's connected to something you were talking about in terms of robotics 00:52:00.000 --> 00:52:02.797 is active learning. So as opposed to 00:52:02.880 --> 00:52:05.437 sort of unsup completely unsupervised or 00:52:05.520 --> 00:52:08.237 self-supervised learning, 00:52:08.320 --> 00:52:11.997 you ask the system for human help for 00:52:12.080 --> 00:52:14.637 for selecting parts you want annotated 00:52:14.720 --> 00:52:16.637 next. So if you think about a robot 00:52:16.720 --> 00:52:19.037 exploring a space or a baby exploring a 00:52:19.120 --> 00:52:24.397 space or a system exploring a data set every once in a while asking for human 00:52:24.480 --> 00:52:25.197 input, 00:52:25.280 --> 00:52:29.277 >> do you see value in that kind of work? 00:52:29.360 --> 00:52:33.277 >> I don't see transformative value. It it's going to make things that we can 00:52:33.360 --> 00:52:36.717 already do >> more efficient or they will learn 00:52:36.800 --> 00:52:39.597 slightly more efficiently, but it's not going to make machines sort of 00:52:39.680 --> 00:52:42.297 significantly more intelligent, I think. 00:52:42.380 --> 00:52:42.557 >> [snorts] 00:52:42.640 --> 00:52:45.517 >> And I and and by the way there is no 00:52:45.600 --> 00:52:50.237 opposition there's no um uh conflict 00:52:50.320 --> 00:52:52.957 between self-supervised learning reinforcement learning and supervised 00:52:53.040 --> 00:52:56.157 learning or imitation learning or active learning. 00:52:56.240 --> 00:52:58.877 >> I see self-supervised learning as a as a 00:52:58.960 --> 00:53:01.037 preliminary to all of the above. 00:53:01.120 --> 00:53:03.677 >> Yes. So 00:53:03.760 --> 00:53:05.677 the example I use very often is how is 00:53:05.760 --> 00:53:10.077 it that so if you use 00:53:10.160 --> 00:53:11.917 classical reinforcement learning deep 00:53:12.000 --> 00:53:14.717 reinforcement learning if you want 00:53:14.800 --> 00:53:18.477 the best methods today 00:53:18.560 --> 00:53:23.037 soal model free reinforcement learning to learn to play Atari games uh take 00:53:23.120 --> 00:53:26.317 about 80 hours of training to reach the level that any human can reach in about 00:53:26.400 --> 00:53:28.477 15 minutes. Mhm. 00:53:28.560 --> 00:53:31.037 >> Um they get better than humans, but it 00:53:31.120 --> 00:53:34.477 takes them a long time. 00:53:34.560 --> 00:53:38.477 Alpha star. Okay. the, you know, Oreo 00:53:38.560 --> 00:53:41.437 vin and and his teams the system to play 00:53:41.520 --> 00:53:45.197 um to um to play Starcraft 00:53:45.280 --> 00:53:48.557 plays, you know, a single map, a single 00:53:48.640 --> 00:53:51.677 type of uh player 00:53:51.760 --> 00:53:54.157 and can reach 00:53:54.240 --> 00:53:57.197 better than human level with about the 00:53:57.280 --> 00:54:00.477 equivalent of 200 years of training 00:54:00.560 --> 00:54:05.117 playing against itself. It's 200 years, right? It's not something that no no 00:54:05.200 --> 00:54:08.557 human can could ever do. >> I mean, I'm not sure what lesson to take 00:54:08.640 --> 00:54:12.077 away from that. >> Okay. Now, take those algorithms, the 00:54:12.160 --> 00:54:15.037 best RL algorithms we have today. To 00:54:15.120 --> 00:54:19.917 train a a car to drive itself, >> it would probably have to drive millions 00:54:20.000 --> 00:54:23.757 of hours. It will have to kill thousands of pedestrians. It will have to run into 00:54:23.840 --> 00:54:26.157 thousands of trees. It will have to run off cliffs. 00:54:26.240 --> 00:54:28.557 >> Yeah. >> And it had to run the cliff multiple 00:54:28.640 --> 00:54:32.317 times before it figures out that it's a bad idea, first of all. Yeah. 00:54:32.400 --> 00:54:33.917 >> And second of all, before it figures out 00:54:34.000 --> 00:54:38.877 how not to do it and so I mean this type of learning obviously does not reflect 00:54:38.960 --> 00:54:42.477 the kind of learning that animals and humans do. There is something missing 00:54:42.560 --> 00:54:44.557 that's really really important there. 00:54:44.640 --> 00:54:48.877 And my hypothesis which I've been advocating for like five years now is 00:54:48.960 --> 00:54:53.037 that we have predictive models of the world 00:54:53.120 --> 00:54:57.677 that include the ability to predict under uncertainty. [snorts] And what 00:54:57.760 --> 00:55:02.477 allows us to not run off a cliff when we learn to 00:55:02.560 --> 00:55:06.077 drive, most of us can learn to drive in about 20 or 30 hours of training without 00:55:06.160 --> 00:55:09.277 ever crashing, causing any accident. And 00:55:09.360 --> 00:55:13.437 if we drive next to a cliff, we know that if we turn the wheel to the right, 00:55:13.520 --> 00:55:16.797 the car is going to run off the cliff and nothing good is going to come out of 00:55:16.880 --> 00:55:19.437 this because we have a pretty good model of intuitive physics that tells us, you 00:55:19.520 --> 00:55:23.037 know, the car is going to fall. We know we know about gravity. Babies learned 00:55:23.120 --> 00:55:24.637 this around the age of eight or nine 00:55:24.720 --> 00:55:27.357 months that objects don't float, they 00:55:27.440 --> 00:55:32.077 they fall. Um, and you know, we have a pretty good idea of the effect of 00:55:32.160 --> 00:55:34.957 turning the wheel on the car and you know, we know we need to stay on the 00:55:35.040 --> 00:55:38.077 road. So, there's a lot of things that we bring to the table, which is 00:55:38.160 --> 00:55:39.437 basically our predictive model of the 00:55:39.520 --> 00:55:42.477 world. [snorts] And that model allows us 00:55:42.560 --> 00:55:45.117 to not do stupid things and to basically 00:55:45.200 --> 00:55:49.597 stay within the context of uh things we need to do. we still face, you know, 00:55:49.680 --> 00:55:53.837 unpredictable situations and that's how we learn. But that allows us to learn 00:55:53.920 --> 00:55:57.597 really, really, really quickly. So that's called model based reinforcement 00:55:57.680 --> 00:56:01.117 learning. There's some imitation and supervised 00:56:01.200 --> 00:56:04.397 learning because we have a driving instructor that tells us occasionally 00:56:04.480 --> 00:56:10.157 what to do. Uh but most of the learning >> is model based >> is learning the model. 00:56:10.240 --> 00:56:12.637 >> Yeah. >> Learning physics that we've done since 00:56:12.720 --> 00:56:15.357 we were babies. That's where all almost all the learning 00:56:15.440 --> 00:56:17.277 >> and the physics is somewhat transferable 00:56:17.360 --> 00:56:21.917 from is transferable from scene to scene. Stupid things are the same 00:56:22.000 --> 00:56:25.197 everywhere. >> Yeah. I mean if you you know you have an 00:56:25.280 --> 00:56:29.037 experience of the world you don't need to be a particularly from a particularly 00:56:29.120 --> 00:56:31.517 intelligent species to know that if you 00:56:31.600 --> 00:56:34.877 spill water from a container the you 00:56:34.960 --> 00:56:37.197 know the the rest is going to get wet. 00:56:37.280 --> 00:56:40.397 You might get wet. Um so you know cats 00:56:40.480 --> 00:56:44.317 know this right? Yeah. >> So the main problem we need to solve is 00:56:44.400 --> 00:56:47.117 how do we learn models of the world? 00:56:47.200 --> 00:56:48.397 That's and that's what I'm interested. 00:56:48.480 --> 00:56:51.197 That's what self-supervised learning is all about. 00:56:51.280 --> 00:56:52.877 >> If you were to try to construct a 00:56:52.960 --> 00:56:56.477 benchmark for um 00:56:56.560 --> 00:56:58.877 let's let's look at emnest. I love that 00:56:58.960 --> 00:57:02.077 data set. But if do you think it's 00:57:02.160 --> 00:57:06.957 useful interesting slash possible to uh 00:57:07.040 --> 00:57:11.117 perform well on emnest with just one example of each digit? 00:57:11.200 --> 00:57:15.597 And how would we solve that problem? 00:57:15.680 --> 00:57:17.037 >> The answer is probably yes. The question 00:57:17.120 --> 00:57:21.277 is what other type of learning are you allowed to do? So if what you're allowed 00:57:21.360 --> 00:57:25.117 to do is train on some gigantic data set of labeled digit, that's called transfer 00:57:25.200 --> 00:57:28.317 learning. >> And we know that works. Okay. 00:57:28.400 --> 00:57:32.397 >> Uh we do this at Facebook like in production, right? We we train large 00:57:32.480 --> 00:57:35.517 convolutional nest to predict hashtags that people type on Instagram and we 00:57:35.600 --> 00:57:39.597 train on billions of images, literally billions. And uh and then we chop off 00:57:39.680 --> 00:57:43.437 the last layer and fine-tune on whatever task we want. That works really well. 00:57:43.520 --> 00:57:47.117 You can beat you know the imageet record with this. We actually open source the 00:57:47.200 --> 00:57:48.957 whole thing like a few weeks ago. 00:57:49.040 --> 00:57:50.557 >> Yeah, that's that's still pretty cool. 00:57:50.640 --> 00:57:54.717 But yeah, so what in what would be impressive and what's useful and 00:57:54.800 --> 00:57:57.437 impressive? What kind of transfer learning would be useful and impressive? 00:57:57.520 --> 00:58:00.637 Is it Wikipedia? That kind of thing? 00:58:00.720 --> 00:58:04.157 >> I don't think transfer learning is really where we should focus. We should 00:58:04.240 --> 00:58:07.037 try to do 00:58:07.120 --> 00:58:08.877 you know have a kind of scenario for a 00:58:08.960 --> 00:58:13.037 benchmark where you have unlabelled data 00:58:13.120 --> 00:58:15.597 and you can and it's very large number 00:58:15.680 --> 00:58:18.237 of unlabelled data. Uh it could be video 00:58:18.320 --> 00:58:21.357 clips, it could be where you do you know 00:58:21.440 --> 00:58:23.197 frame prediction, it could be images 00:58:23.280 --> 00:58:28.397 where you could choose to you know mask a piece of it 00:58:28.480 --> 00:58:30.157 could be whatever but they're unlabeled 00:58:30.240 --> 00:58:32.877 and you're not allowed to label them. 00:58:32.960 --> 00:58:37.517 So you do some training on this and then 00:58:37.600 --> 00:58:41.517 you train on a particular supervised 00:58:41.600 --> 00:58:45.277 task imageet or nest and you measure how 00:58:45.360 --> 00:58:48.797 your test error decreases or validation error 00:58:48.880 --> 00:58:51.837 decreases as you increase the number of label training samples. 00:58:51.920 --> 00:58:55.757 >> Mhm. >> Okay. And 00:58:55.840 --> 00:58:58.477 and what what you'd like to see is is 00:58:58.560 --> 00:59:01.357 that you know your your error decreases 00:59:01.440 --> 00:59:05.677 much faster than if you train from scratch from random weights. 00:59:05.760 --> 00:59:09.197 So that to reach the same level of performance that a completely supervised 00:59:09.280 --> 00:59:13.757 uh purely supervised system would reach you would need way fewer samples. So 00:59:13.840 --> 00:59:17.357 that's the crucial question because it will answer the question to like you 00:59:17.440 --> 00:59:19.197 know people interested in medical image 00:59:19.280 --> 00:59:21.037 analysis. Okay, you know, if I want to 00:59:21.120 --> 00:59:24.637 get to a particular uh level of uh of 00:59:24.720 --> 00:59:27.197 error rate for this task, I know I need 00:59:27.280 --> 00:59:31.197 um a million samples, can I do, you 00:59:31.280 --> 00:59:34.797 know, self-supervised pre-training to reduce this to about 100 or something? 00:59:34.880 --> 00:59:37.997 >> And you think the answer there is self-supervised pre-training? 00:59:38.080 --> 00:59:42.237 >> Yeah, some form some for some form of it. 00:59:42.320 --> 00:59:45.837 >> Telling you active learning, but you disagree. 00:59:45.920 --> 00:59:49.837 >> No, it's it's not useless. It's just not going to lead to a quantum leap. It's 00:59:49.920 --> 00:59:51.197 just going to make things that we already do. 00:59:51.280 --> 00:59:55.437 >> So, you're way smarter than me. I just disagree with you and but I don't have 00:59:55.520 --> 00:59:59.037 anything to back that. It's just intuition. So, I worked with a lot of 00:59:59.120 --> 01:00:00.637 large scale data sets and there's 01:00:00.720 --> 01:00:03.437 something there might be magic in active 01:00:03.520 --> 01:00:07.917 learning. But, okay, at least I said it publicly. [laughter] 01:00:08.000 --> 01:00:09.757 At least I'm being an idiot publicly. 01:00:09.840 --> 01:00:13.437 Okay. It's not being an idiot. It's, you know, working with the data you have. I 01:00:13.520 --> 01:00:17.757 mean I mean certainly people are doing things like okay I have 3,000 hours of 01:00:17.840 --> 01:00:21.357 you know imitation running for self-driving car but most of those are 01:00:21.440 --> 01:00:25.597 incredibly boring. What I like is select you know 10% of them that are kind of 01:00:25.680 --> 01:00:27.757 the most informative and with just that 01:00:27.840 --> 01:00:33.117 I would probably reach the same. So it's a weak form of uh of active learning if 01:00:33.200 --> 01:00:37.357 you want. >> Yes. But there might be a a much stronger version. 01:00:37.440 --> 01:00:39.357 >> Yeah that's right. >> That's what and that's an open question 01:00:39.440 --> 01:00:41.917 if it exists. The the question is how 01:00:42.000 --> 01:00:44.877 much stronger we can get. Elon Musk is 01:00:44.960 --> 01:00:49.437 confident. Talked to him recently. Uh he's confident that large scale data and 01:00:49.520 --> 01:00:53.197 deep learning can solve the autonomous driving problem. 01:00:53.280 --> 01:00:57.037 What are your thoughts on the limits possibilities of deep learning in this 01:00:57.120 --> 01:01:00.237 space? >> It's obviously part of the solution. I 01:01:00.320 --> 01:01:03.757 mean I don't think we'll ever have a self-driving system or at least not in 01:01:03.840 --> 01:01:05.517 the foreseeable future that does not use 01:01:05.600 --> 01:01:07.677 deep learning. Let me put it this way. 01:01:07.760 --> 01:01:11.277 Now how much of it? So in the history of 01:01:11.360 --> 01:01:15.837 sort of engineering particularly sort of 01:01:15.920 --> 01:01:19.437 sort of AI like systems there's generally a first phase where everything 01:01:19.520 --> 01:01:22.477 is built by hand and there is a second phase and that was the case for 01:01:22.560 --> 01:01:27.357 autonomous driving you know 20 30 years ago there's a phase where there's a 01:01:27.440 --> 01:01:30.477 little bit of learning is used but there's a lot of engineering that's 01:01:30.560 --> 01:01:35.437 involved in kind of you know taking care of corner cases and and putting limits 01:01:35.520 --> 01:01:38.957 etc because the learning system is not perfect and then as technology 01:01:39.040 --> 01:01:44.157 progresses, we end up relying more and more on learning. That's the history of 01:01:44.240 --> 01:01:47.277 character recognition, the history of speech recognition, now computer vision, 01:01:47.360 --> 01:01:49.037 natural language processing. And I think 01:01:49.120 --> 01:01:52.397 the same is going to happen with uh with 01:01:52.480 --> 01:01:55.917 autonomous driving that currently the 01:01:56.000 --> 01:02:00.637 u the methods that are closest to providing some level of autonomy, some 01:02:00.720 --> 01:02:04.077 you know decent level of autonomy where you don't expect a driver to kind of do 01:02:04.160 --> 01:02:08.397 anything is where you constrain the world. So you 01:02:08.480 --> 01:02:12.397 only run within, you know, 100 square kilometers or square miles in Phoenix, 01:02:12.480 --> 01:02:14.637 but the weather is nice and the roads 01:02:14.720 --> 01:02:17.437 are wide, which is what Whimo is doing. 01:02:17.520 --> 01:02:20.317 You uh completely overengineer the car 01:02:20.400 --> 01:02:24.157 with tons of light and sophisticated 01:02:24.240 --> 01:02:27.517 sensors that are too expensive for consumer cars, but they're fine if you 01:02:27.600 --> 01:02:32.157 just run a fleet. And you engineer the thing uh the hell 01:02:32.240 --> 01:02:36.077 out of the everything else. You you map the entire world, so you have complete 01:02:36.160 --> 01:02:39.597 3D model of everything. So the only thing that the perception system has to 01:02:39.680 --> 01:02:41.757 take care of is moving objects and and 01:02:41.840 --> 01:02:46.867 and construction and sort of you know things that that weren't in your map 01:02:46.950 --> 01:02:49.357 [snorts] and you can engineer a good you know 01:02:49.440 --> 01:02:52.957 slam system and all that stuff right so so that's kind of the current approach 01:02:53.040 --> 01:02:56.077 that's closest to some level of autonomy but I think eventually the long-term 01:02:56.160 --> 01:03:00.077 solution is going to rely more and more on 01:03:00.160 --> 01:03:03.197 learning and possibly using a combination of self-supervised learning 01:03:03.280 --> 01:03:05.837 and model based 01:03:05.920 --> 01:03:11.117 reinforcement or something like that >> but ultimately learning will be at not 01:03:11.200 --> 01:03:14.397 just at the core but really the fundamental part of the system. 01:03:14.480 --> 01:03:17.517 >> Yeah, it already is but it'll become more and more. 01:03:17.600 --> 01:03:21.277 >> What do you think it takes to build a system with human level intelligence? 01:03:21.360 --> 01:03:23.517 You talked about the AI system in the 01:03:23.600 --> 01:03:28.477 movie her being way out of reach our current reach. This might be outdated as 01:03:28.560 --> 01:03:30.557 well but uh >> it's still way out of reach. 01:03:30.640 --> 01:03:33.197 >> It's still way out of reach. Uh what 01:03:33.280 --> 01:03:38.157 would it take to build her, do you think? Um, so I can tell you the first 01:03:38.240 --> 01:03:41.357 two obstacles that we have to clear, but I don't know how many obstacles there 01:03:41.440 --> 01:03:44.637 are after this. So the image I usually use is that there's a bunch of mountains 01:03:44.720 --> 01:03:47.677 that we have to climb and we can see the first one, but we don't know if there 01:03:47.760 --> 01:03:50.237 are 50 mountains behind it or not. And 01:03:50.320 --> 01:03:52.077 this might be a good sort of metaphor 01:03:52.160 --> 01:03:55.677 for why AI researchers in the past have 01:03:55.760 --> 01:03:57.517 been overly optimistic about the result 01:03:57.600 --> 01:04:00.557 of of AI. 01:04:00.640 --> 01:04:03.757 You know for example um Newan Simon 01:04:03.840 --> 01:04:08.397 right wrote the general problem solver and they called it the general problem solver 01:04:08.480 --> 01:04:11.357 >> general problem solver >> okay and of course the first thing you 01:04:11.440 --> 01:04:13.997 realize is that all the problems you want to solve are exponential and so you 01:04:14.080 --> 01:04:17.277 can't actually use it for anything useful >> but you know 01:04:17.360 --> 01:04:21.117 >> yeah so yeah all you see is the first peak so what are the first couple of 01:04:21.200 --> 01:04:24.477 peaks for her >> um so the first peak which is precisely 01:04:24.560 --> 01:04:27.837 what I'm working on is uh self-s supervised learning how how do we get 01:04:27.920 --> 01:04:29.677 machines to learn models of the world by 01:04:29.760 --> 01:04:32.077 observation kind of like babies and like 01:04:32.160 --> 01:04:35.117 young animals. 01:04:35.200 --> 01:04:38.317 So we've been working with um you know 01:04:38.400 --> 01:04:40.797 cognitive scientists. 01:04:40.880 --> 01:04:43.437 So this Emanuel Dup who's at fair in 01:04:43.520 --> 01:04:46.317 Paris is uh 01:04:46.400 --> 01:04:49.037 halftime he's also a researcher in 01:04:49.120 --> 01:04:52.477 French university and he he um he has 01:04:52.560 --> 01:04:55.517 this chart that shows that which how 01:04:55.600 --> 01:04:58.077 many months of of life baby humans can 01:04:58.160 --> 01:05:00.477 learn different concepts and you can you 01:05:00.560 --> 01:05:03.997 can measure this sort of various ways. 01:05:04.080 --> 01:05:08.717 So things like um um distinguishing animate objects from 01:05:08.800 --> 01:05:12.317 inanimate inanimate object you can you can tell the difference at age two three 01:05:12.400 --> 01:05:15.837 months whether an object is going to stay 01:05:15.920 --> 01:05:18.797 stable is going to fall you know about 01:05:18.880 --> 01:05:23.517 four months you can tell you know there various things like this and then things 01:05:23.600 --> 01:05:26.957 like gravity the fact that objects are not supposed to float in the air but are 01:05:27.040 --> 01:05:30.957 supposed to fall you learn this around the age of eight or nine months if you 01:05:31.040 --> 01:05:35.037 look at a lot of you know eight-month old babies You give them a bunch of toys 01:05:35.120 --> 01:05:37.837 on their high chair. First thing they do is they throw them on the ground and 01:05:37.920 --> 01:05:40.157 they look at them. >> It's because, you know, they're learning 01:05:40.240 --> 01:05:43.357 about actively learning about gravity. 01:05:43.440 --> 01:05:46.637 >> Gravity. Yeah. >> Okay. So, they're not trying to annoy 01:05:46.720 --> 01:05:49.677 you, but they, you know, they need to do the experiment, right? 01:05:49.760 --> 01:05:51.917 >> Yeah. >> Uh, so, you know, how do we get machines 01:05:52.000 --> 01:05:55.277 to learn like babies? mostly by observation with a little bit of 01:05:55.360 --> 01:05:58.877 interaction and learning those those those models of the world because I 01:05:58.960 --> 01:06:03.677 think that's really a crucial piece of an intelligent autonomous system. So if 01:06:03.760 --> 01:06:07.037 you think about the architecture of an intelligent autonomous system, it needs 01:06:07.120 --> 01:06:08.557 to have a predictive model of the world. 01:06:08.640 --> 01:06:12.237 So something that says here is a world at time t here is a set of world at time 01:06:12.320 --> 01:06:16.797 t plus one if I take this action >> and it's not a single answer it can be a 01:06:16.880 --> 01:06:18.477 >> yeah it can be distribution. Yeah. 01:06:18.560 --> 01:06:21.037 >> Yeah. Well we don't know how to represent distributions in high 01:06:21.120 --> 01:06:24.397 dimension spaces. So it's got to be something weaker than that. Okay. 01:06:24.480 --> 01:06:28.157 >> But with some some representation of uncertainty. 01:06:28.240 --> 01:06:31.917 >> If you have that then you can do what optimal control theorist call model 01:06:32.000 --> 01:06:36.157 predictive control which means that you can run your model with a hypothesis for 01:06:36.240 --> 01:06:39.997 a sequence of action and then see the result. Now what you need the other 01:06:40.080 --> 01:06:43.917 thing you need is some sort of objective that you want to optimize. Am I reaching 01:06:44.000 --> 01:06:48.397 the goal of grabbing this object? Am I minimizing energy? Am I whatever? Right? 01:06:48.480 --> 01:06:53.037 So there is some sort of objective that you have to minimize and so in your head 01:06:53.120 --> 01:06:55.757 if you have this model you can figure out the sequence of action that will 01:06:55.840 --> 01:07:00.077 optimize your objective. That objective is something that 01:07:00.160 --> 01:07:03.757 ultimately is rooted in your basel ganglia at least in the human brain. 01:07:03.840 --> 01:07:08.797 That's that's what it is. Basel ganglia computes your level of contentment or or 01:07:08.880 --> 01:07:12.717 miscontentment. I don't know if that's a word. Uh unhappiness. Okay. 01:07:12.800 --> 01:07:14.477 >> Yeah. Yeah. >> Discontentment. 01:07:14.560 --> 01:07:17.117 >> Discontentment maybe. Um and so your 01:07:17.200 --> 01:07:20.397 entire behavior is driven towards 01:07:20.480 --> 01:07:22.317 kind of minimizing that objective which 01:07:22.400 --> 01:07:25.037 is maximizing your contentment computed 01:07:25.120 --> 01:07:30.077 by your your bas ganglia. >> And what you have is an objective 01:07:30.160 --> 01:07:33.197 function which is basically a predictor of what your bas ganglia is going to 01:07:33.280 --> 01:07:35.917 tell you. >> So you're not going to put your hand on 01:07:36.000 --> 01:07:40.237 fire because you know it's going to you know it's going to burn and you're going 01:07:40.320 --> 01:07:43.517 to get hurt. And you're predicting this because of your model of the world and 01:07:43.600 --> 01:07:45.757 your your sort of predictor of this 01:07:45.840 --> 01:07:50.557 objective, right? So you if you have those you have those three components, 01:07:50.640 --> 01:07:52.637 you have the four components. You have 01:07:52.720 --> 01:07:54.797 the the hardwired 01:07:54.880 --> 01:07:59.037 uh contentment objective um computer if 01:07:59.120 --> 01:08:02.637 you want calculator [snorts] and then you have the three components. One is 01:08:02.720 --> 01:08:04.397 the objective predictor which basically 01:08:04.480 --> 01:08:06.317 predicts your level of contentment. one 01:08:06.400 --> 01:08:10.077 is the model of the world and there's a 01:08:10.160 --> 01:08:12.317 third module I didn't mention which is a 01:08:12.400 --> 01:08:14.877 module that will figure out the best 01:08:14.960 --> 01:08:19.917 course of action to optimize an objective given your model 01:08:20.000 --> 01:08:21.517 >> okay 01:08:21.600 --> 01:08:24.557 >> call this a policy policy network or 01:08:24.640 --> 01:08:29.677 something like that right [snorts] now you need those three components to act 01:08:29.760 --> 01:08:33.357 autonomously intelligently and you can be stupid in three different ways you 01:08:33.440 --> 01:08:36.157 can be stupid because your model of your of the world is wrong. 01:08:36.240 --> 01:08:38.397 >> Mhm. >> You can be stupid because your objective 01:08:38.480 --> 01:08:42.237 is not aligned with what you actually want to achieve. 01:08:42.320 --> 01:08:46.317 >> Okay. Uh >> in humans that would be a psychopath, 01:08:46.400 --> 01:08:49.757 >> right? >> And then the the third thing you the 01:08:49.840 --> 01:08:52.797 third way you can be stupid is that you have the right model, you have the right 01:08:52.880 --> 01:08:56.717 objective, but you're unable to figure out a course of action to optimize your 01:08:56.800 --> 01:08:58.477 objective given your model. >> Right? 01:08:58.560 --> 01:09:02.637 >> Okay. Um some people who are in charge of big 01:09:02.720 --> 01:09:04.157 countries actually have all three that 01:09:04.240 --> 01:09:07.357 are wrong. All right. [laughter] 01:09:07.440 --> 01:09:09.757 Uh, which countries? I don't know. Okay. 01:09:09.840 --> 01:09:12.317 So, uh, if we think about this this 01:09:12.400 --> 01:09:16.957 agent, if we think about the movie her, you've 01:09:17.040 --> 01:09:20.957 criticized the art project that is Sophia the 01:09:21.040 --> 01:09:23.837 robot. And what that project essentially 01:09:23.920 --> 01:09:26.797 does is uses our natural inclination to 01:09:26.880 --> 01:09:30.877 anthropomorphize things that look like human and give 01:09:30.960 --> 01:09:32.957 them more. Do you think that could be 01:09:33.040 --> 01:09:35.677 used by AI systems like in the movie 01:09:35.760 --> 01:09:40.797 Her? So, do you think that body is needed to 01:09:40.880 --> 01:09:43.997 create a feeling of intelligence? 01:09:44.080 --> 01:09:47.437 >> Uh, well, if Sophia was just an art piece, I would have no problem with it, 01:09:47.520 --> 01:09:50.157 but it's presented as something else. 01:09:50.240 --> 01:09:55.117 >> Let me add that comment real quick. If creators of Sophia could change 01:09:55.200 --> 01:09:58.637 something about their marketing or behavior in general, what would it be? 01:09:58.720 --> 01:10:01.787 What What's >> Oh, just about everything. 01:10:01.870 --> 01:10:04.877 [laughter] I mean, 01:10:04.960 --> 01:10:07.197 don't you think here's a tough question. 01:10:07.280 --> 01:10:09.837 Let me So, I agree with you. So, Sophia 01:10:09.920 --> 01:10:15.517 is not the general public feels that Sophia can do way more than 01:10:15.600 --> 01:10:17.437 she actually can. >> That's right. 01:10:17.520 --> 01:10:20.237 >> And the people who created Sophia are 01:10:20.320 --> 01:10:22.077 not 01:10:22.160 --> 01:10:25.517 honestly publicly communicating uh 01:10:25.600 --> 01:10:28.717 trying to teach the public. But 01:10:28.800 --> 01:10:31.917 >> here's a tough question. Don't you think 01:10:32.000 --> 01:10:36.957 uh this the same thing is uh scientists 01:10:37.040 --> 01:10:41.037 in industry and research are taking advantage of the same mis 01:10:41.120 --> 01:10:43.117 misunderstanding in the public when they 01:10:43.200 --> 01:10:47.117 create AI companies or publish stuff. um 01:10:47.200 --> 01:10:49.197 >> some companies yes I mean there is no 01:10:49.280 --> 01:10:54.317 sense of there's no desire to delude there's no desire to to kind of 01:10:54.400 --> 01:10:57.757 overclaim what something is done right you publish a paper on AI that you know 01:10:57.840 --> 01:11:01.517 has this result on imageet you know it's pretty clear I mean it's not not even 01:11:01.600 --> 01:11:06.237 interesting anymore but >> you know I I don't think there is that 01:11:06.320 --> 01:11:08.317 I mean the reviewers are generally not 01:11:08.400 --> 01:11:10.877 very forgiving of 01:11:10.960 --> 01:11:13.357 >> of you know unsupported claims of of 01:11:13.440 --> 01:11:15.757 this type and but there are certainly 01:11:15.840 --> 01:11:18.557 quite a few startups that have had a 01:11:18.640 --> 01:11:21.197 huge amount of hype around this that uh 01:11:21.280 --> 01:11:25.597 I find extremely damaging and I've been calling it out when I've seen it. Um so 01:11:25.680 --> 01:11:28.637 yeah, but to go back to your original question like 01:11:28.720 --> 01:11:32.957 >> the necessity of embodiment. I think I don't think embodiment is necessary. I 01:11:33.040 --> 01:11:36.637 think grounding is necessary. So I don't think we're going to get machines that 01:11:36.720 --> 01:11:39.517 really understand language without some level of grounding in the real world. 01:11:39.600 --> 01:11:43.517 And it's not clear to me that language is a high enough bandwidth medium to 01:11:43.600 --> 01:11:46.077 communicate how the real world works. 01:11:46.160 --> 01:11:49.357 Um I think for this >> can you talk to ground what grounding means? 01:11:49.440 --> 01:11:51.837 >> So grounding means that so there is this 01:11:51.920 --> 01:11:54.237 classic problem of common sense 01:11:54.320 --> 01:11:56.717 reasoning you know the the winner winner 01:11:56.800 --> 01:12:02.077 schema right and so I tell you the the trophy doesn't fit in the suitcase 01:12:02.160 --> 01:12:04.877 because it's too it's too big or the trophy doesn't fit in the suitcase 01:12:04.960 --> 01:12:09.117 because it's too small and the it in the first case refers to the trophy in the 01:12:09.200 --> 01:12:12.397 second case to the suitcase. And the reason you can figure this out is 01:12:12.480 --> 01:12:14.877 because you know what the trophy and the suitcase are. you know one is supposed 01:12:14.960 --> 01:12:18.957 to fit in the other one and you know the notion of size and a big object doesn't 01:12:19.040 --> 01:12:22.477 fit in a small object unless it's a TARDIS you know things like that right 01:12:22.560 --> 01:12:25.117 so you have this this knowledge of how 01:12:25.200 --> 01:12:27.357 the world works of geometry and things 01:12:27.440 --> 01:12:30.477 like that um I don't believe you can 01:12:30.560 --> 01:12:32.157 learn everything about the world by just 01:12:32.240 --> 01:12:34.877 being told in language how the world 01:12:34.960 --> 01:12:39.437 works I think you need some uh low-level perception of the world you know be it 01:12:39.520 --> 01:12:41.437 visual touch you know whatever but some 01:12:41.520 --> 01:12:46.557 higher bandwidth perception of the So by reading all the world's text, you still 01:12:46.640 --> 01:12:48.317 may not have enough information. 01:12:48.400 --> 01:12:50.557 >> That's right. There's a lot of things 01:12:50.640 --> 01:12:52.717 that just will never appear in text and 01:12:52.800 --> 01:12:55.757 that you can't really infer. So I think 01:12:55.840 --> 01:13:00.077 >> uh common sense will emerge from >> you know certainly a lot of language 01:13:00.160 --> 01:13:02.317 interaction but also with watching 01:13:02.400 --> 01:13:07.357 videos or perhaps even interacting in in virtual environments and possibly you 01:13:07.440 --> 01:13:10.717 know robot interacting in the real world but I don't actually believe necessarily 01:13:10.800 --> 01:13:14.637 that this last one is absolutely necessary but I think there's a the need 01:13:14.720 --> 01:13:19.677 for some grounding >> but the final product doesn't 01:13:19.760 --> 01:13:23.197 necessarily need to be embodied you're saying it just needs to have an 01:13:23.280 --> 01:13:25.117 awareness a grounding to 01:13:25.200 --> 01:13:28.157 >> but he needs to know how the world works 01:13:28.240 --> 01:13:30.317 to have you know to not be frustrated 01:13:30.400 --> 01:13:35.197 frustrating to talk to >> and uh you talked about emotions being 01:13:35.280 --> 01:13:38.957 important that's that's a whole another topic 01:13:39.040 --> 01:13:41.437 >> well so you know I talked about this uh 01:13:41.520 --> 01:13:45.197 the basil ganglia ganglia as the 01:13:45.280 --> 01:13:48.557 you your your you know the thing that calc you know calculates your level of 01:13:48.640 --> 01:13:52.557 mix contentment contentment and then there is this other module that sort of 01:13:52.640 --> 01:13:55.677 tries to do a prediction of whether you're going to be content or not. 01:13:55.760 --> 01:13:57.597 That's the source of some emotion. So 01:13:57.680 --> 01:14:03.117 fear, for example, is an anticipation of bad things that can happen to you, 01:14:03.200 --> 01:14:05.917 right? >> Uh you have this inkling that there is 01:14:06.000 --> 01:14:08.637 some chance that something really bad is going to happen to you and that creates 01:14:08.720 --> 01:14:10.637 fear. >> When you know for sure that something 01:14:10.720 --> 01:14:14.557 bad is going to happen to you, you kind of give up, right? It's not fear anymore. 01:14:14.640 --> 01:14:16.637 >> It's uncertainty that creates fear. 01:14:16.720 --> 01:14:19.597 >> So So the punch line is yes, we're not going to have autonomous intelligence 01:14:19.680 --> 01:14:24.157 without emotions. Okay. [sighs and gasps] 01:14:24.240 --> 01:14:28.157 >> Whatever the heck emotions are. So you mentioned very practical things of fear, 01:14:28.240 --> 01:14:30.557 but there's a lot of other mess around it. 01:14:30.640 --> 01:14:33.597 >> But there are kind of the results of, you know, drives. 01:14:33.680 --> 01:14:36.077 >> Yeah. There's deeper uh biological stuff 01:14:36.160 --> 01:14:40.237 going on and I've talked to a few folks on this. There's uh there's fascinating 01:14:40.320 --> 01:14:42.877 stuff that ultimately connects to our to 01:14:42.960 --> 01:14:48.077 to our brain. If we create an AGI system, sorry, 01:14:48.160 --> 01:14:51.517 >> human level intelligence, >> human level intelligence system 01:14:51.600 --> 01:14:54.237 and you get to ask her one question, 01:14:54.320 --> 01:14:56.957 what would that question be? 01:14:57.040 --> 01:15:01.277 >> You know, I think the the first one we'll create will probably not be not be 01:15:01.360 --> 01:15:04.637 that smart. They be like a four-year-old. Okay. 01:15:04.720 --> 01:15:09.997 >> So, you would have to ask her a question to know she's not that smart. 01:15:10.080 --> 01:15:13.757 >> Yeah. Well, what's a good question to ask, you 01:15:13.840 --> 01:15:17.917 know, to be >> the cause of wind? 01:15:18.000 --> 01:15:20.637 And if she answers, oh, it's because the 01:15:20.720 --> 01:15:25.677 leaves of the tree are moving and that creates wind, she's on to something. 01:15:25.760 --> 01:15:29.837 >> And if she says, that's a stupid question, she's really on to something. 01:15:29.920 --> 01:15:32.317 >> No. And then you tell her actually, you 01:15:32.400 --> 01:15:35.197 know, here is the the real thing. And 01:15:35.280 --> 01:15:37.677 she says, oh yeah, that makes sense. 01:15:37.760 --> 01:15:42.877 So questions that that uh reveal the ability to do common sense reasoning 01:15:42.960 --> 01:15:47.197 about the physical world. >> Yeah. And you'll sum it up with a causal inference. 01:15:47.280 --> 01:15:51.597 >> Causal inference. Well, >> it was a huge honor. Congratulations on 01:15:51.680 --> 01:15:54.397 the touring award, Yan. Thank you so much for talking today. 01:15:54.480 --> 01:15:56.807 >> Thank you. >> Appreciate it. 01:15:56.890 --> 01:15:58.910 [music] 01:16:04.090 --> 01:16:06.110 >> [music]