Yann LeCun: Deep Learning, Convolutional Neural Networks, and Self-Supervised Learning

HAL 9000 and value misalignment

00:00The following is a conversation with Yan Lakun. He's considered to be one of the fathers of deep learning, which if you've been hiding under a rock, is the recent revolution in AI that has captivated the world with the possibility of what machines can learn from data. He's a professor at New York University, a vice president and chief AI scientist at Facebook, and co-reient of the touring award for his work on deep learning. He's probably best known as the founding father of convolutional neural networks, in particular their application to optical character recognition and the famed emnest data set. He is also an outspoken personality unafraid to speak his mind in a distinctive French accent and explore

00:44provocative ideas both in the rigorous medium of academic research and the somewhat less rigorous medium of Twitter and Facebook. This is the artificial intelligence podcast. If you enjoy it, subscribe on YouTube, give it five stars on iTunes, support it on Patreon, or simply connect with me on Twitter at Lex Freedman, spelled F R I D M. And now, here's my conversation with Yan Lun.

01:21[music]

01:28You said that 2001 Space Odyssey is one of your favorite movies. Hell 9000 decides to get rid of the astronauts for people who haven't seen the movie. Spoiler alert, because he it she believes that the astronauts they will interfere with the mission. Do you see Hal as flawed in some fundamental way or even evil or did he do the right thing? >> Neither. There's no notion of evil in that in that context other than the fact that people die. But it was an example of what people call uh value misalignment, right? And you give an objective to a machine and the machine strives to achieve this objective and if you don't put any constraints on this objective like don't kill people and don't do things like this.

02:19The machine given the power will do stupid things just to achieve this this objective or damaging things to achieve this objective. It's a little bit like I mean we are used to this in the context of human society. We we put in place laws to prevent people from doing bad things because spontaneously they would do those bad things, right? So we have to shape their cost function, their objective function if you want through laws to kind of correct and education obviously to sort of correct for for those. So maybe just pushing a little further on on that point, Hal, you know, there's a mission. there's a this fuzziness around the the ambiguity around what the actual mission is but you know do you think that there will be

03:11a time from a utilitarian perspective where an AI system where it is not misalignment where it is alignment for the greater good of society that an AI system will make decisions that are difficult >> well that's the trick I mean eventually we'll have to figure out how to do this and again we're not starting from scratch because we've been doing this with humans for for millennia. >> So designing objective functions for people is something that we know how to do and we don't do it by um you know programming things although the legal code is called code. Um so that tells you something and it's actually the design of an objective function. That's really what legal code is right. It tells you here is what you

03:53can do here is what you can't do. If you do it you pay that much. That's that's an objective function. So there is this idea somehow that it's a new thing for people to try to design objective functions that are aligned with the common good. But no, we've been writing laws for millennia and that's exactly what it is. Um so this that's where you know uh the science of uh law making and uh and computer science will >> come together >> will come together. >> So it's nothing there's nothing special about hell or AI systems. It's just a continuation of tools used to make some of these difficult ethical judgments that laws make. >> Yeah. And we and we have systems like this already that you know uh make many

04:35decisions for for ourselves in society that you know need to be designed in a way that they like you know rules about things that sometime sometimes have bad side effects and we have to be flexible enough about those rules so that they can be broken when it's obvious that they shouldn't be applied. So, you don't see this on the camera here, but all the decoration in this room is all pictures from 2001 of Space Odyssey. [laughter] >> Wow. Is that by accident or is there a lot? >> It's not by accident. It's by design. >> Oh, wow. So, if you were um if you were to build HAL 10,000, so an improvement of HAL 9,000, what would you improve? Well, first of all, I wouldn't ask it to hold secrets and tell lies because that's really what

05:20breaks it in the end. That's the the fact that it's asking itself questions about the purpose of the mission and it's you know pieces things together that it's heard you know all the secrecy of the preparation of the mission and the fact that it was discovery on on the lunar surface that really was kept secret and and one part of how's memory knows this and the other part is does not know it and is supposed to not tell anyone and that creates internal conflict. So you think there's never should be a set of things that an AI system should not be allowed like a a set of facts that should not be shared with the human operators. Well, I think no. I think that I think it should be a bit like in the design of

06:06autonomous AI systems. Um, there should be the equivalent of, you know, the the the the oath that hypocratic oath Yeah. >> that uh doctors sign up to, right? So, there's certain things, certain rules that that that you have to abide by and we can sort of hardwire this into into our into our machines to kind of make sure they don't go. So I'm not you know an advocate of the the three law three laws of robotics you know the azimov kind of thing because I don't think it's practical but but you know some some level of uh of limits but but to be clear uh this is not these are not questions that are kind of really worth asking today because we just don't have the technology to do this. We don't we don't have autonomous intelligent

06:53machines. We have intelligent machines, semi- intelligent machines that are very specialized, but they don't they don't really sort of satisfy an objective. They're just, you know, kind of trained to do one thing. >> So, until we have some idea for design of a full-fledged autonomous intelligent system, asking the question of how how we design is objective, I think is a little a little too abstract. >> It's a little too abstract. There's useful elements to it in that it helps us understand our own ethical codes, humans. So even just as a thought experiment, if you imagine that an AGI system is here today, h how would we program it is a kind of nice thought experiment of constructing how should we

The most surprising fact in deep learning

07:37have a law have a system of laws for us humans. It's it's just a nice practical tool. And I think there's echoes of that idea too in the AI systems we have today that don't have to be that intelligent like autonomous vehicles. These things start creeping in that they're worth thinking about but certainly they shouldn't be framed as as hell. Yeah. >> Looking back, what is the most I'm sorry if it's a silly question, but what is the most beautiful or surprising idea in deep learning or AI in general that you've ever come across sort of personally where you sat back and and just had this kind of a that's pretty cool moment. That's nice surprising. >> I don't know if it's an idea rather than a sort of empirical

08:27uh fact. The fact that you can build gigantic neural nets train them on, you know, relatively small amounts of data relatively with stoastic gradient and descent and that it actually works. Uh breaks everything you read in every textbook, right? Every pre deep learning textbook that told you you need to have fewer parameters than you have data samples. Um you know if you have a non-convex objective function, you have no guarantee of convergence. You know all those things that you read in textbook and they tell you stay away from this and they're all wrong. >> Huge number of parameters non-convex and somehow which just vary relative to the number of uh parameters data it's able to learn anything

09:12>> right >> does that still surprise you today? >> Well it it was kind of obvious to me before I knew anything that that this is a good idea and then it became surprising that it worked because I started reading those textbooks. >> Okay. Wait. So, okay. So, can you talk through the intuition of why it was obvious to you if you remember? >> Well, okay. So, the intuition was it's it's sort of like, you know, those people in the late 19th century who proved that heavier than than air flight was impossible, right? And of course, you have birds, right? They do fly. And so, on the face of it, it it's obviously wrong as an empirical question, right? And so, we have the same kind of thing that, you know, the we know that the

09:54brain works. We don't know how but we know it works and we know it's a large network of neurons in interaction and that learning takes place by changing the connections. So kind of getting this level of inspiration without covering the details but sort of trying to derive basic principles you know that kind of gives you a clue as to which direction to go. There's also the idea somehow that I' I've been convinced of since I was in undergrad that even before that intelligence is inseparable from learning. So you the idea somehow that you can create an intelligent machine by basically programming for me was a non-starter you know from the start every intelligent entity that we know about arrives at this intelligence through

10:40learning so learning you know machine learning was completely obvious path uh also because I'm lazy so you know [laughter] >> you automate basically everything and learning is the automation of >> exactly >> intelligence, >> right? >> So do you think so what is learning then? What falls under learning? Because do you think of reasoning is learning? >> Well, reasoning is certainly a a a consequence of of learning as well just like other functions of uh of the brain. The big question about reasoning is how do you make reasoning compatible with gradient-based learning? >> Do you think neural networks can be made to reason? Yes, there is no question about that. Again, we have a good example, right?

Reasoning, memory, and gradient-based learning

11:28The question is is how? So, the question is how much prior structure do you have to put in the neural net so that something like human reasoning will emerge from it you know from learning. Uh another question is all of our kind of model of what reasoning is that are based on logic are discrete and and and are therefore incompatible with gradient based learning. Mhm. >> And I'm a very strong believer in this idea gradient based learning. I don't believe that uh other types of learning that don't use kind of gradient information if you want. >> So you don't like discrete mathematics. You don't like anything discrete. >> Well, that's it's not that I don't like it. It's just that it's it's incompatible with learning and I'm I'm a

12:08big fan of learning, right? So in fact, that's perhaps one reason why deep learning has been kind of looked at with suspicion by a lot of computer scientists because the math is very different. the math that uh you use for deep learning, you know, kind of has more to do with, you know, cybernetics, uh the kind of math you do in electrical engineering than the kind of math you do in computer science. [snorts] And and you know, nothing in in machine learning is exact, right? Computer science is all about sort of, you know, obviously compulsive attention to details of like, you know, every index has to be right and you can prove that an algorithm is correct, right? Uh machine learning is the science of sloppiness. Really?

12:47[laughter] >> That's beautiful. So okay, maybe let's feel around in the dark of what is a neural network that reasons or a system that is works with continuous functions that's able to do uh build knowledge. However we think about reasoning, build on previous knowledge, build on extra knowledge, create new knowledge, generalize outside of any training set ever built. What does that look like? If uh yeah, maybe do you have inklings of thoughts of what that might look like? >> Well, yeah. I mean, yes and no. If I had precise ideas about this, I think, you know, we'd be building it right now. But and there are people working on this or whose main research interest is actually exactly that right. So what you need to

13:41have is a working memory. So you need to have some device if you want uh some subsystem that can store uh a relatively large number of factual episodic information for you know a reasonable amount of time. So you you know in the in the brain for example there are kind of three main types of memory. one is the the sort of memory of the the the state of your cortex and that sort of disappears within 20 seconds. You can't remember things for more than about 20 seconds or a minute if if if you don't have any other form of memory. Uh [snorts] the second type of memory which is longer term but still shortterm is the hippoc campus. So you can you know you came into this building you remember where the where the the

14:27exit is where the uh elevators are. Um, you have some map of that building that's stored in your hippoc campus. >> You might remember something about what I said, you know, a few minutes ago. >> I forgot it all already. >> It's been erased, but you know, but that that would be in your in your hypoc campus. >> Uh, and then the the longer term memory is in the synapse, the synapses, right? >> Um, so what you need if you want a system that's capable of reasoning is that you want the hypocampus like thing, right? And [snorts] that's what people have tried to do with memory networks and you know neural machines and stuff like that right and and now with transformers which have sort of a a memory in their

15:07kind of self attention system you can you can think of it this way so so that's one element you need another thing you need is some sort of network that can access this memory get an information back and then kind of crunch on it and do this iteratively multiple times because a chain of reasoning is is a process by which you you you kind of you update your knowledge about the state of the world about you know what's going to happen etc and that that has to be this sort of recurrent operation basically >> and you think that kind of if if we think about a transformer so that seems to be too small to contain the knowledge that's that's uh to represent the knowledge that's contained in Wikipedia media for example

15:56>> but transformer doesn't have this idea of recurrence it's got a fixed number of layers and that's the number of steps that you know limits basically it representation >> but recurrence would build on the knowledge somehow I mean it would it would evolve the knowledge and expand the amount of information perhaps or useful information within that knowledge >> but is is this something that just can emerge with size because it seems like everything we have now is too small no it's not It's not it's not clear I mean how how you access and write into an associative memory in efficient way. I mean sort of the original memory network maybe had something like the right architecture but uh if you try to scale

16:36up a memory network so that the memory contains all Wikipedia it doesn't quite work >> right >> so so there's there's a need for new ideas there okay but it's not the only form of reasoning so there's another form of reasoning which is through which is very classical also in some types of AI uh and it's based on u let's call it energy minimization okay so you have uh some sort of objective, some energy function that represents the the the um the quality or the negative quality. Okay, energy goes up when things get bad and they get low when things get good. So, let's say you you want to figure out, you know, what gestures do I need to to do to grab an object or walk out the door. [snorts] If you have a good model of uh your own

Causality and the limits of human intuition

17:27body, a good model of the environment using this kind of energy minimization, you can make a you can make you can do planning and it's uh in optimal control is called it's called model model predictive control. You have a model of what's going to happen in the world as a consequence of your actions and that allows you to by energy minimization figure out a sequence of action that optimizes a particular objective function which measures you know minimizes number of times you're going to hit something and the energy you're going to spend doing the gesture and etc. So so that's a form of reasoning. Planning is a form of reasoning. And perhaps what led to the ability of humans to reason is the fact that or you

18:07know species you know that appear before us had to do some sort of planning to be able to hunt and survive and survive the winter in particular and so you know it's the same capacity that that you need to have. So in your intuition is um if we look at expert systems and encoding knowledge as logic systems as graphs or in this kind of way is not a useful way to think about knowledge. >> Graphs are a little brittle or or logic uh representation. So basically you know variables that that have values and then constraint between them that are represented by rules uh is a little too rigid and too brittle right. So one of the you know some of the early efforts in that respect um were were to put probabilities on

18:58them. So a rule you know you know if you have this and that symptom you know you have this uh disease with that probability and you should prescribe that antibiotic with that probability right this the m system from the from the 70s um and that that's what that branch of AI led to uh you know basics and graphical models and causal inference and viral you know >> method so so there there is I mean certainly a lot of interesting work going on in this area. The main issue with this is is knowledge acquisition. How do you uh reduce a bunch of data to a graph of this type? >> Yeah. It relies on the expert to on a human being to encode to add knowledge >> and that's essentially impractical. >> Yeah, it's not scalable.

19:46>> That's that's a big question. The second question is do you want to represent knowledge as symbols and do you want to manipulate them with logic? And again that's incompatible with learning. So uh one suggestion which Jeffington has been advocating for many decades is replace symbols by uh vectors. Think of it as pattern of activities in a bunch of neurons or units or whatever you want to call them and replace logic by continuous functions. >> Mhm. >> Okay. And that becomes now compatible. There's a very good set of ideas by uh written in a paper about 10 years ago by Leon Bou on who is here at Facebook. Um um the title of the paper is from machine learning to machine reasoning and his idea is that uh a learning

20:36learning system should be able to manipulate objects that are in the same spa in a space and then put the result back in the same space. So it's this idea of working memory basically [snorts] and it's uh it's very enlightening >> and in a sense that might learn something like the simple expert systems. I mean it's you can learn basic logic operations there. >> Yeah, quite possibly. Yeah, there's a big debate on sort of how much prior structure you have to put in for this kind of stuff to emerge. That's the debate I have with Gary Marcus and people like that. >> Yeah. Yeah. So and the other person so I just talked to Judea Pearl >> from you mentioned causal inference world. So his worry is that the current

21:20neural networks are not able to learn what causes what causal inference between things. >> So I think I think he's right and wrong about this. uh if he's talking about the sort of classic uh type of neural nets, people sort of didn't worry too much about this, but there's a lot of people now working on calls on inference and there's a paper that just came out last week by Leonu among others, David Lopez and a bunch of other people exactly on that problem of um how do you kind of you know get a neural net to sort of pay attention to real causal relationships which may also solve issues of bias. in data and things like this. So >> I' I'd like to read that paper because uh that ultimately the challenge there

22:09is also seems to fall back on the human expert to ultimately decide causality between things. >> People are not very good at establishing causality first of all. So first of all you talk to physicists and physicists actually don't believe in causality because look at the >> all the basic laws of microfysics are time reversible. So there is no causality. >> The arrow of time is not real. >> It's it's as soon as you start looking at microscopic systems where there is unpredictable randomness where there is clearly an arrow of time but it's a big mystery in physics actually how that emerges. Um >> is it emergent or is it part of the fundamental fabric of reality? Yeah. Or is it a bias of intelligent systems that

22:54you know because of the second law of thermodynamics we perceive a particular hour of time but in fact it's kind of arbitrary right >> so yeah physicists mathematicians they don't care about I mean the math doesn't care about the flow of time >> well certainly certainly micro physics doesn't people themselves are not very good at establishing causal causal relationships if you ask this I think it was in one of simar paper's book on on um like children learning you know he studied with Jean Pia you know he's the guy who co-authored [snorts] the book perceptron with Marvin Minsky that kind of killed the first wave of neural nets but but he was actually a a learning person he uh in the sense of studying

23:37learning in humans and and machines that's why he got interested in perceptron and uh and he he he wrote that if you ask a little kid uh about what is the cause of of the wind a lot of kids will say they will think for a while and they'll say oh it's the the branches in the trees they move and that creates wind right so they get the causal relationship backwards and it's because their understanding of the world and intuitive physics is not that great right I mean these are like you know four or five year old kids >> you know it gets better and then you understand that this it can't be right but there are many things which we can because of our common sense understanding of things what people call

24:19common sense >> and with understanding of physics we can >> there's a lot of stuff that we can figure out causality even with diseases we can figure out what's not causing what >> often there's a lot of mystery of course but the idea is that you should be able to encode that into systems it seems unlikely they'd be able to figure that out themselves >> well whenever we can do intervention but you know all of humanity has been completely deluded for millennia probably since existence about a very very wrong causal relationship where whatever you can explain you attribute it to you know some deity some divinity right >> um and that's a copout that's a way of saying like I don't know the cause so

Why neural nets were abandoned in the 1990s

24:59you know God did it right so you mentioned Marvin Minsky and uh the irony of um uh you know maybe causing the first AI winter you were there in the '9s you were there in the 80s of course uh in the '9s why do you think people lost faith in deep learning in the '9s and found it again a decade later, over a decade later. >> Yeah, it wasn't called deep learning yet. It was just called neural nets, but um >> yeah, they lost interest. I mean, I think I would put that around 1995, at least the machine learning community. There was always a neural net community, but it but became kind of disconnected from sort of mainstream machine learning if you want. There were it was basically electrical engineering that kept at it,

25:49>> right? >> Uh and computer science >> just gave up >> gave up on uh on neural nets. I don't I don't know, you know, I was too close to it to really sort of analyze it with sort of a unbiased eye if you want. But I I would I would I would I would make a few guesses. So the first one is at the time neural nets were it was very hard to make them work in the sense that you would you know implement back prop in your favorite language um and that favorite language was not Python it was not matlab it was not any of those things because they didn't exist right you had to write it in forc or something like this right um so you would uh experiment with it you would probably make some very basic mistakes takes like you know badly

26:39initialize your weights make the network too small because you read in the textbook you know you don't want too many parameters right and of course you know and you would train on exor because you didn't have any other data set to train it on and of course you know it works half the time so you' say I give up also you train it with batch gradient which you know isn't efficient so there's a lot of there's bag of tricks that you had to know to make those things work or you you had to reinvent and a lot of people just didn't and they just couldn't make it work. Um, so that's one thing. The investment in software platform to be able to kind of, you know, display things, figure out why things don't work, kind of get a good

27:17intuition for how to get them to work, have enough flexibility so you can create, you know, network architectures like convolutional nets and stuff like that. It was hard. I mean, you had to write everything from scratch and again you didn't have any Python or MATLAB or anything, right? >> [snorts] >> I read that, sorry to interrupt, but I read that you wrote in in lisp the your first uh versions of uh Lynette with convolutional networks, which by the way, one of my favorite languages. >> That's how I knew you were legit. Uh touring award, whatever. This what you programmed in Lisp. That's >> it's still my favorite language. But >> it's not that we programmed in Lisp, it's that we had to write a list

27:53interpreter. >> Okay? Because it's not like we used one that existed. So >> we wrote a list interpreter that we hooked up to you know a backend library that we wrote also for sort of neural net computation. And then after a few years around 1991 we invented this idea of uh basically having modules that know how to forward propagate and back propagate gradients and then interconnecting those modules in a graph. [snorts] Leotu had made proposals on this about this in the late 80s and we're able to implement this using a list system. Eventually we wanted to use that system to make build production code for character recognition at Bell Lab. So we actually wrote a compiler for that list interpreter so that Patricy Mart who's

28:36now Microsoft kind of did did the bulk of it with Leon and and me and uh and so we could write our system in list and then compile to C and then we'll have a self-contained complete system that could kind of do the entire thing. Neither PyTorch nor transform can do this today yet. Okay, it's coming. >> Yeah, >> I mean there's something like that in PyTorch called you know to script >> and so you know we had to write all this interpreter. to this compiler. where to invest a huge amount of effort to do this and not everybody if if you don't completely believe in the concept you're not going to invest the time to do this right >> now at the time also you know or today >> this would turn into torture pietor

29:19torture tensorflow or whatever we put it in open source everybody would use it and you know realize it's good back before 1995 working at AT&T there's no way the lawyers would let you release anything in open source of this nature And so we could not distribute our code really. >> And on that point and sorry to go on a million tangents but on that point I also read that there was some almost pat like a patent on convolution networks labs. >> [sighs and gasps] >> So that >> first of all I mean just >> there's two actually [laughter] >> uh that ran out uh the that >> thankfully in 2007 >> in 2007 I what can we can we just talk about that for I know you're a Facebook but you're also at NYU and and

30:10what does it mean to patent ideas like these software ideas essentially or what or mathematical ical ideas or what are they? >> Okay, so they're not mathematical ideas. There are, you know, algorithms and there was a period where the US patent office would allow the patent of software as long as it was embodied. Uh the Europeans are very different. They don't they don't quite accept that. They have a different concept. But you know, I don't I no longer I mean I never actually strongly believed in this, but I I don't believe in this kind of patent. Facebook basically doesn't believe in this kind of patent. Google fights patents because they've been burned with Apple. And so now they do this for defensive purpose. But usually

30:59they say we're not going to sue you if you infringe. >> Facebook has a has a similar policy. They say, you know, we patent on certain things for defensive purpose. We're not going to see you if you infringe unless you sue us. Um so the the industry does not believe in in patents. They're there because of you know the legal landscape and and and various things but but I don't really believe in patents for this kind of stuff. >> Okay. So that's that's a great thing. So I I >> I'll tell you a worst story actually. So what happens was the the first the first patent about convolutional net was about kind of the early version of convolutional net that didn't have separate pooling layers. It had convolutional layers which tried more

31:40than one if you want right >> and then there was a second one on convolutional nets with separate pooling layers uh train with back prop and they were filed filed in ' 89 and 1990 something like this at the time the life life of a patent was 17 years so here's what what happened over the next few years is that we started developing um character recognition technology around conviction nets [snorts] and uh in 1994 a check reading system was deployed in ATM machines. In 1995 it was for large check reading machines in back offices etc. And those systems were developed by an engineering group that we were collaborating with at AT&T and they were commercialized by NCR which at the time was a subsidiary of AT&T. Mhm.

32:28>> Now, AT&T split up in 1996, 199 early 1996 and the lawyers just looked at all the patents and they distributed the patents among the various companies. They gave the the commercial net patent to NCR because they were actually selling products that used it. But nobody at NCR had any idea what a commercial net was. >> Yeah. >> Okay. So between 1996 and 2007, so there's a whole period until 2002 where I didn't actually work on machine learning or control it. I resumed working on this around 2002 and between 2002 and 2007 I was working on them crossing my finger that nobody at NCR would notice and nobody noticed. >> Yeah. And I and I hope that this kind of somewhat as you said lawyers aside relative openness of the community now

33:18will continue. It accelerates the entire progress of of the industry. And you know the problems that Facebook and Google and others are facing today is not whether Facebook or Google or Microsoft or IBM or whoever is ahead of the other. It's that we don't have the technology to build the things we want to build. We want to build intelligent virtual assistants that have common sense. We don't have monopoly on good ideas for this. We don't believe we do. Maybe others do believe they do, but we don't. Okay. If a startup tells you they have the secret to, you know, human level intelligence and common sense, don't believe them. They don't. And it's going to take the entire uh work of the world research community for a while to

Patents, AT&T, and the convolutional net story

34:02get to the point where you can go off and in each of those companies kind of start to build things on this. We're not there yet. >> Absolutely. And this this calls to the the gap between the space of ideas and the rigorous testing of those ideas of practical application that you often speak to. You've uh written advice saying don't get fooled by people who claim to have a solution to artificial general intelligence who claim to have an AI system that work just like the human brain or who claim to have figured out how the brain works. Ask them what the error rate they get on emnest or imageet. Yeah, >> this is a little dated by the way [laughter] >> 2000 I mean five years who's counting okay but I think your opinion is still

34:48emnest and imageet uh yes maybe data there may be new benchmarks right but I think that philosophy is one you still and and somewhat hold that benchmarks and the practical testing the practical application is where you really get to test the ideas >> well it may not be completely practical like for examp you know it could be a toy data set or but but it to be some sort of task that the community as a whole has accepted as some sort of standard you know kind of benchmark if you want it doesn't need to be real. So for example many years ago here at fair um people you know chisen west on auto and board and a few others proposed the the baby tasks which were kind of a a toy problem to test the ability of

35:30machines to reason actually to access working memory and things like this and uh it was very useful even though it wasn't a real task. amnest is kind of halfway a real task. Uh so you know toy problems can be very useful. It's just that I was really struck by the fact that a lot of people particularly a lot of people with money to invest would be fooled by people telling them oh we have you know the algorithm of the cortex and you should give us 50 million. >> Yes. Absolutely. So there's a lot of people who who try to take advantage of the hype for business reasons and so on. But let me sort of talk to this idea that sort of new ideas the ideas that push the field forward may not yet have a benchmark or it may be very difficult

36:16to establish a benchmark. >> I agree that's part of the process. Establishing benchmarks is part of the process. So what are your thoughts about so we have these benchmarks on around stuff we can do with images from classification to captioning to just every kind of information you can pull off from images and the surface level there's audio data sets there's some video what can we start natural language what kind of stuff what kind of benchmarks do you see that start creeping on to more something like intelligence like reasoning like maybe you don't like the term but AGI echoes of that kind of uh >> a lot of people are working on interactive environments in which you can you can train and test uh

Benchmarks and honest claims

37:04intelligent systems so so there for example uh you know it's the classical paradigm of supervised learning is that you you have a data set you partition it into a training set validation set test set and there's a clear protocol right but What if the that assumes that the samples are stat statistically independent? You can exchange them. The order in which you see them doesn't shouldn't matter. You know, things like that. But what if the answer you give determines the next sample you see, which is the case for example in robotics, right? You robot does something and then it gets exposed to a new room and depending on where it goes, the room will be different. So that's the that creates the exploration

37:45problem. >> Mhm. the what if the samples. So that creates also a dependency between samples, right? You you if you move if you can only move in in space, the next sample you're going to see is going to be probably in the same building most likely, right? So so the all the assumptions about the validity of this training set test set hypothesis break >> whenever a machine can take an action that has an influence in the in the world and it's what it's going to see. So people are setting up artificial uh environments where where that takes place, right? The robot runs around a 3D model of a house and can interact with objects and things like this. So you do robotics by simulation. You have those,

38:28you know, open eye gym type thing or muoko kind of simulated uh robots and you have games, you know, things like that. Um so that that's where the field is going really. um this kind of environment. [snorts] Now back to the question of AGI like I don't like the term AGI because it implies that human intelligence is general >> and human intelligence is nothing like general. It's very very specialized. We think it's general. We like to think of ourselves as having general intelligence. We don't. We're very specialized. >> We're only slightly more general than >> Why does it feel general? So you kind of uh the term general. I think what's impressive about humans is this ability to learn as we were talking about

Is human intelligence general?

39:14learning to learn in just so many different domains. It's perhaps not arbitrarily general but just you can learn in many domains and integrate that knowledge somehow. Okay. The knowledge persists. >> So let me take a very specific example. >> Yes, >> it's not an example. It's more like a a quasi mathematical demonstration. So you have about 1 million fibers coming out of one of your eyes. Okay. 2 million total. But let's let's talk about just one of them. >> It's 1 million nerve fibers. Your optical nerve. >> Let's imagine that they are binary. So they can be active or inactive, right? So the input to your visual cortex is 1 million bits. >> Mhm. >> Now they connected to your brain in a particular way. And your brain has

39:58connections that are kind of a little bit like a conversion net. They're kind of local uh you know in space and things like this. >> Now imagine I play a trick on you. It's a pretty nasty trick, I admit. I uh I cut your optical nerve >> and I put a device that makes a random perturbation of a permutation of all the nerve fibers. So now what comes to your to your brain is a fixed but random permutation of all the pixels. >> Mhm. >> There's no way in hell that your visual cortex, even if I do this to you in infancy, will actually learn vision to the same level of quality that you can. >> Got it. And you're saying there's no way you relearn that. >> No, because now two pixels that are nearby in the world will end up in very

40:44different places in your visual cortex and your neurons there have no connections with each other because they only connected locally. >> So this whole our entire the hardware is built in many ways to support >> the locality of the real world. >> Yeah. >> Yes. That's specialization. >> Yeah. But it's still pretty damn impressive. So it's not perfect generalization. It's not even close. >> No. No. It's it's it's it's not that it's not even close. It's not at all. >> It's not specialized. >> So how many boolean functions? So let's imagine you want to train your visual system to uh you know recognize particular patterns of those 1 million bits. Okay. So that's a boolean function, right? Either the pattern is

41:23here or not here. This a two two-way classification with 1 million binary inputs. How many such boolean functions are there? Okay, you have you have two to the 1 million combinations of inputs. For each of those, you have an output bit and so you have two to the two to the 1 million boolean functions of this type. Okay, which is an unimaginably large number. How many of those functions can actually be computed by your visual cortex? And the answer is a tiny tiny tiny tiny tiny tiny sliver. Like an enormously tiny sliver. >> Yeah. Yeah. So, we are ridiculously specialized, you know. Okay. >> But, [laughter] okay, that's an argument against the word general. I think there there's a >> I I there's I agree with your intuition,

42:16but I'm not sure it's it seems the bre the the brain is impressively uh capable of adjusting to things. So >> it's because we can't imagine tasks that are outside of our comprehension, >> right? So we think we think we're general because we're general of all the things that we can apprehend. >> So yeah, >> but there is a huge world out there of things that we have no idea. We call that heat, by the way. >> Heat. >> Heat. So uh at least physicists call that heat or they call it entropy which is >> you know you have a thing full of gas right >> closed system full of gas >> right closed or no closed it has uh you know pressure it has uh temperature it has you know and you can write equations PV equivalent on T you know things like

43:10that right [snorts] when you reduce a volume the temperature goes up the pressure goes up, you know, things like that, right? Uh for a perfect gas at least, those are the things you can know about that system and it's a tiny tiny number of bits compared to the complete information of the state of the entire system because the state of the entire system will give you the position and momentum of every every molecule of the gas. And what you don't know about it is the entropy and you interpret it as heat. the energy contained in that thing is is what we call heat. Now, it's very possible that in fact there is some very strong structure in how those molecules are moving. It's just that they are in a way that we're just not

43:55wired to perceive. >> Yeah, we're ignorant to it and there's in near infinite amount of things we're not wired to perceive. >> And you're right, that's a that's a nice way to put it. We're general to all the things we can imagine which is a very tiny subset of all things that are possible. >> So it's like chromograph complexity or the chromograph chhattin of complexity. >> Yeah. >> You know every bit string or every integer is random except for all the ones that you can actually write down. [laughter] >> Yeah. Okay. So beautifully put but you know so we can just call it artificial intelligence. We don't need to have a general >> or human level human level intelligence is >> you know you you'll start anytime you

44:41touch human it gets it gets interesting because um [snorts] uh you know it's it's because we attach ourselves to human and it's difficult to define what human intelligence is. >> Yeah. >> Nevertheless, my definition is uh maybe uh demo impressive intelligence. Okay. Damn impressive demonstration of intelligence whatever and so on that topic most successes in deep learning have been in supervised learning. >> What is your view on unsupervised learning? Is there hope to reduce involvement of human input and still have uh successful systems that are have practical use? >> Yeah, I mean there's definitely a hope. It's it's more than a hope actually. it's uh it's you know mounting evidence for it and that's basically all I do

Self-supervised learning explained

45:33like the only thing I'm interested in at the moment is I call it self-supervised learning not unsupervised because unsupervised learning is is a loaded term people who know something about machine learning you know tell you so you're doing clustering or PCA >> which is not the case and the wide public you know when you say unsupervised learning oh my god you know machines are going to learn by themselves and without supervision you know they see this as >> where's where's the parents Yeah. So, so I call it self-supervised learning because in fact the underlying algorithms that are used are the same algorithms as the supervised learning algorithms. Um, except that what we train them to do is not predict a

46:10particular set of variables like the category of of an image and and not to predict a set of variables that have been provided by human labelers. But what you train the machine to do is basically reconstruct a piece of its input that it's being is being maxed maxed out essentially. You can think of it this way, right? So show a piece of video to a machine and ask it to predict what's going to happen next. And of course after a while you can show what what happens and the machine will kind of train itself to do better at that task. You can do like all the latest most successful models in natural language processing use self-supervised learning. you know, sort of BERT style systems, for example, right? You show it a window

46:56of a thousand words on a test corpus. You take out 15% of the words and then you train a machine to predict the words that are missing. >> That's self-supervised learning. It's not predicting the future. It's just, you know, predicting things in the middle, but you could have it predict the future. That's what language models do. >> So, you construct, so in an unsupervised way, you construct a model of language. Do you think >> or video or the physical world or whatever? Right. >> How far do you think that can take us? Do you think >> art >> understands anything? >> To some level it has you know a shallow understanding of uh of text but it needs to I mean to have kind of true human level intelligence I think you need to

47:44ground language in reality. So some people are attempting to do this right having systems that kind of have some visual representation of what what is being talked about which is one reason you need those interactive environments actually but this is like a huge technical problem that is not solved and that explains why self-supervised learning works in the context of natural language but does not work in the context or at least not well in the context of image recognition and video although it's making progress quickly and [snorts] the reason that reason is the fact that it's much easier to represent uncertainty in the prediction in a context of natural language than it is in the context of things like video and

48:26images. >> So for example, if I ask you to predict what words are missing, you know, 15% of the words that are taken out, >> the possibilities are small. That means >> it's small, right? There is 100,000 words in the in the lexicon and what the machine spits out is a a big probability vector, right? It's a bunch of numbers between 0 and one that's some to one and we know how to do how to do this with computers. [snorts] So there representing uncertainty in the prediction is relatively easy and that's in my opinion why those techniques work for NLP for images if you ask if you block a piece of an image and you ask a system reconstruct that piece of the image there are many possible answers that are all perfectly legit right and

49:12how do you represent that this set of possible answers you you can't train a system to make one prediction you can't train a neural net to say here it is that's the image because it's there's a whole set of things that are compatible with it. So how do you get the machine to represent not a single output but a whole set of outputs and you know similarly with uh video prediction there's a lot of things that can happen in the future of a video you're looking at me right now I'm not moving my head very much but you know I might you know turn my my head to the left or to the right >> right >> if you don't have a system that can predict this >> and you train it with le square to kind of minimize the error with the

49:50prediction and what I'm doing what you get is a blurry image of myself right >> in all possible future positions that I might be in which is not a good prediction. But so there might be other ways to do the self-supervision, right, for visual scenes. >> Like what? [laughter] >> If I to I mean if I knew I wouldn't tell you I publish it first. I don't know. I >> No, there might be. [laughter] >> So I mean these are kind of uh there might be artificial ways of like selfplay in games to where you can simulate part of the environment. You can >> Oh, that doesn't solve the problem. It's just a way of generating data. But because you have more of a control mean you can control. Yeah. It's a way to generate data. That's right. And

50:34because you can do huge amounts of data generation that doesn't you're right. This well it's it's a creeps up on the problem from the side of data and you don't think that's the right way to creep up. >> It doesn't solve this problem of handling uncertainty in the world. Right. So if you if you have a machine learn a predictive model of the world in a game that is deterministic or quasi deterministic uh it's easy right just you know give a few frames of the game to a connet put a bunch of layers and then have the game generates the next few frames >> and and if the game is deterministic it works fine. Um, and that includes, you know, feeding the system with the action that your little character is is going

51:18to take the problem comes from the fact that the real world and certain most games are not entirely predictable. And so there you get those blurry predictions and you can't do planning with blurry predictions. Right? So if you have a perfect model of the world, you can in your head run this model with a hypothesis for a sequence of actions. Are you going to predict the outcome of that sequence of actions? >> Mhm. But if your model is imperfect, how can you plan? >> Yeah. It quickly explodes. What are your thoughts on the extension of this which topic I'm super excited about? It's connected to something you were talking about in terms of robotics is active learning. So as opposed to sort of unsup completely unsupervised or

Reinforcement learning inefficiency and world models

52:05self-supervised learning, you ask the system for human help for for selecting parts you want annotated next. So if you think about a robot exploring a space or a baby exploring a space or a system exploring a data set every once in a while asking for human input, >> do you see value in that kind of work? >> I don't see transformative value. It it's going to make things that we can already do >> more efficient or they will learn slightly more efficiently, but it's not going to make machines sort of significantly more intelligent, I think. >> [snorts] >> And I and and by the way there is no opposition there's no um uh conflict between self-supervised learning reinforcement learning and supervised

52:53learning or imitation learning or active learning. >> I see self-supervised learning as a as a preliminary to all of the above. >> Yes. So the example I use very often is how is it that so if you use classical reinforcement learning deep reinforcement learning if you want the best methods today soal model free reinforcement learning to learn to play Atari games uh take about 80 hours of training to reach the level that any human can reach in about 15 minutes. Mhm. >> Um they get better than humans, but it takes them a long time. Alpha star. Okay. the, you know, Oreo vin and and his teams the system to play um to um to play Starcraft plays, you know, a single map, a single type of uh player and can reach

53:54better than human level with about the equivalent of 200 years of training playing against itself. It's 200 years, right? It's not something that no no human can could ever do. >> I mean, I'm not sure what lesson to take away from that. >> Okay. Now, take those algorithms, the best RL algorithms we have today. To train a a car to drive itself, >> it would probably have to drive millions of hours. It will have to kill thousands of pedestrians. It will have to run into thousands of trees. It will have to run off cliffs. >> Yeah. >> And it had to run the cliff multiple times before it figures out that it's a bad idea, first of all. Yeah. >> And second of all, before it figures out how not to do it and so I mean this type

54:36of learning obviously does not reflect the kind of learning that animals and humans do. There is something missing that's really really important there. And my hypothesis which I've been advocating for like five years now is that we have predictive models of the world that include the ability to predict under uncertainty. [snorts] And what allows us to not run off a cliff when we learn to drive, most of us can learn to drive in about 20 or 30 hours of training without ever crashing, causing any accident. And if we drive next to a cliff, we know that if we turn the wheel to the right, the car is going to run off the cliff and nothing good is going to come out of this because we have a pretty good model

55:18of intuitive physics that tells us, you know, the car is going to fall. We know we know about gravity. Babies learned this around the age of eight or nine months that objects don't float, they they fall. Um, and you know, we have a pretty good idea of the effect of turning the wheel on the car and you know, we know we need to stay on the road. So, there's a lot of things that we bring to the table, which is basically our predictive model of the world. [snorts] And that model allows us to not do stupid things and to basically stay within the context of uh things we need to do. we still face, you know, unpredictable situations and that's how we learn. But that allows us to learn really, really, really quickly. So

55:56that's called model based reinforcement learning. There's some imitation and supervised learning because we have a driving instructor that tells us occasionally what to do. Uh but most of the learning >> is model based >> is learning the model. >> Yeah. >> Learning physics that we've done since we were babies. That's where all almost all the learning >> and the physics is somewhat transferable from is transferable from scene to scene. Stupid things are the same everywhere. >> Yeah. I mean if you you know you have an experience of the world you don't need to be a particularly from a particularly intelligent species to know that if you spill water from a container the you know the the rest is going to get wet.

56:37You might get wet. Um so you know cats know this right? Yeah. >> So the main problem we need to solve is how do we learn models of the world? That's and that's what I'm interested. That's what self-supervised learning is all about. >> If you were to try to construct a benchmark for um let's let's look at emnest. I love that data set. But if do you think it's useful interesting slash possible to uh perform well on emnest with just one example of each digit? And how would we solve that problem? >> The answer is probably yes. The question is what other type of learning are you allowed to do? So if what you're allowed to do is train on some gigantic data set of labeled digit, that's called transfer

57:25learning. >> And we know that works. Okay. >> Uh we do this at Facebook like in production, right? We we train large convolutional nest to predict hashtags that people type on Instagram and we train on billions of images, literally billions. And uh and then we chop off the last layer and fine-tune on whatever task we want. That works really well. You can beat you know the imageet record with this. We actually open source the whole thing like a few weeks ago. >> Yeah, that's that's still pretty cool. But yeah, so what in what would be impressive and what's useful and impressive? What kind of transfer learning would be useful and impressive? Is it Wikipedia? That kind of thing? >> I don't think transfer learning is

58:02really where we should focus. We should try to do you know have a kind of scenario for a benchmark where you have unlabelled data and you can and it's very large number of unlabelled data. Uh it could be video clips, it could be where you do you know frame prediction, it could be images where you could choose to you know mask a piece of it could be whatever but they're unlabeled and you're not allowed to label them. So you do some training on this and then you train on a particular supervised task imageet or nest and you measure how your test error decreases or validation error decreases as you increase the number of label training samples. >> Mhm. >> Okay. And and what what you'd like to see is is

58:58that you know your your error decreases much faster than if you train from scratch from random weights. So that to reach the same level of performance that a completely supervised uh purely supervised system would reach you would need way fewer samples. So that's the crucial question because it will answer the question to like you know people interested in medical image analysis. Okay, you know, if I want to get to a particular uh level of uh of error rate for this task, I know I need um a million samples, can I do, you know, self-supervised pre-training to reduce this to about 100 or something? >> And you think the answer there is self-supervised pre-training? >> Yeah, some form some for some form of

59:41it. >> Telling you active learning, but you disagree. >> No, it's it's not useless. It's just not going to lead to a quantum leap. It's just going to make things that we already do. >> So, you're way smarter than me. I just disagree with you and but I don't have anything to back that. It's just intuition. So, I worked with a lot of large scale data sets and there's something there might be magic in active learning. But, okay, at least I said it publicly. [laughter] At least I'm being an idiot publicly. Okay. It's not being an idiot. It's, you know, working with the data you have. I mean I mean certainly people are doing things like okay I have 3,000 hours of you know imitation running for self-driving car but most of those are

1:00:21incredibly boring. What I like is select you know 10% of them that are kind of the most informative and with just that I would probably reach the same. So it's a weak form of uh of active learning if you want. >> Yes. But there might be a a much stronger version. >> Yeah that's right. >> That's what and that's an open question if it exists. The the question is how much stronger we can get. Elon Musk is confident. Talked to him recently. Uh he's confident that large scale data and deep learning can solve the autonomous driving problem. What are your thoughts on the limits possibilities of deep learning in this space? >> It's obviously part of the solution. I mean I don't think we'll ever have a

Autonomous driving and deep learning

1:01:02self-driving system or at least not in the foreseeable future that does not use deep learning. Let me put it this way. Now how much of it? So in the history of sort of engineering particularly sort of sort of AI like systems there's generally a first phase where everything is built by hand and there is a second phase and that was the case for autonomous driving you know 20 30 years ago there's a phase where there's a little bit of learning is used but there's a lot of engineering that's involved in kind of you know taking care of corner cases and and putting limits etc because the learning system is not perfect and then as technology progresses, we end up relying more and more on learning. That's the history of

1:01:44character recognition, the history of speech recognition, now computer vision, natural language processing. And I think the same is going to happen with uh with autonomous driving that currently the u the methods that are closest to providing some level of autonomy, some you know decent level of autonomy where you don't expect a driver to kind of do anything is where you constrain the world. So you only run within, you know, 100 square kilometers or square miles in Phoenix, but the weather is nice and the roads are wide, which is what Whimo is doing. You uh completely overengineer the car with tons of light and sophisticated sensors that are too expensive for consumer cars, but they're fine if you

1:02:27just run a fleet. And you engineer the thing uh the hell out of the everything else. You you map the entire world, so you have complete 3D model of everything. So the only thing that the perception system has to take care of is moving objects and and and construction and sort of you know things that that weren't in your map [snorts] and you can engineer a good you know slam system and all that stuff right so so that's kind of the current approach that's closest to some level of autonomy but I think eventually the long-term solution is going to rely more and more on learning and possibly using a combination of self-supervised learning and model based reinforcement or something like that >> but ultimately learning will be at not

1:03:11just at the core but really the fundamental part of the system. >> Yeah, it already is but it'll become more and more. >> What do you think it takes to build a system with human level intelligence? You talked about the AI system in the movie her being way out of reach our current reach. This might be outdated as well but uh >> it's still way out of reach. >> It's still way out of reach. Uh what would it take to build her, do you think? Um, so I can tell you the first two obstacles that we have to clear, but I don't know how many obstacles there are after this. So the image I usually use is that there's a bunch of mountains that we have to climb and we can see the first one, but we don't know if there

Building human-level intelligence: obstacles and architecture

1:03:47are 50 mountains behind it or not. And this might be a good sort of metaphor for why AI researchers in the past have been overly optimistic about the result of of AI. You know for example um Newan Simon right wrote the general problem solver and they called it the general problem solver >> general problem solver >> okay and of course the first thing you realize is that all the problems you want to solve are exponential and so you can't actually use it for anything useful >> but you know >> yeah so yeah all you see is the first peak so what are the first couple of peaks for her >> um so the first peak which is precisely what I'm working on is uh self-s supervised learning how how do we get machines to learn models of the world by

1:04:29observation kind of like babies and like young animals. So we've been working with um you know cognitive scientists. So this Emanuel Dup who's at fair in Paris is uh halftime he's also a researcher in French university and he he um he has this chart that shows that which how many months of of life baby humans can learn different concepts and you can you can measure this sort of various ways. So things like um um distinguishing animate objects from inanimate inanimate object you can you can tell the difference at age two three months whether an object is going to stay stable is going to fall you know about four months you can tell you know there various things like this and then things like gravity the fact that objects are

1:05:25not supposed to float in the air but are supposed to fall you learn this around the age of eight or nine months if you look at a lot of you know eight-month old babies You give them a bunch of toys on their high chair. First thing they do is they throw them on the ground and they look at them. >> It's because, you know, they're learning about actively learning about gravity. >> Gravity. Yeah. >> Okay. So, they're not trying to annoy you, but they, you know, they need to do the experiment, right? >> Yeah. >> Uh, so, you know, how do we get machines to learn like babies? mostly by observation with a little bit of interaction and learning those those those models of the world because I think that's really a crucial piece of

1:06:01an intelligent autonomous system. So if you think about the architecture of an intelligent autonomous system, it needs to have a predictive model of the world. So something that says here is a world at time t here is a set of world at time t plus one if I take this action >> and it's not a single answer it can be a >> yeah it can be distribution. Yeah. >> Yeah. Well we don't know how to represent distributions in high dimension spaces. So it's got to be something weaker than that. Okay. >> But with some some representation of uncertainty. >> If you have that then you can do what optimal control theorist call model predictive control which means that you can run your model with a hypothesis for

1:06:36a sequence of action and then see the result. Now what you need the other thing you need is some sort of objective that you want to optimize. Am I reaching the goal of grabbing this object? Am I minimizing energy? Am I whatever? Right? So there is some sort of objective that you have to minimize and so in your head if you have this model you can figure out the sequence of action that will optimize your objective. That objective is something that ultimately is rooted in your basel ganglia at least in the human brain. That's that's what it is. Basel ganglia computes your level of contentment or or miscontentment. I don't know if that's a word. Uh unhappiness. Okay. >> Yeah. Yeah. >> Discontentment.

1:07:14>> Discontentment maybe. Um and so your entire behavior is driven towards kind of minimizing that objective which is maximizing your contentment computed by your your bas ganglia. >> And what you have is an objective function which is basically a predictor of what your bas ganglia is going to tell you. >> So you're not going to put your hand on fire because you know it's going to you know it's going to burn and you're going to get hurt. And you're predicting this because of your model of the world and your your sort of predictor of this objective, right? So you if you have those you have those three components, you have the four components. You have the the hardwired uh contentment objective um computer if

1:07:59you want calculator [snorts] and then you have the three components. One is the objective predictor which basically predicts your level of contentment. one is the model of the world and there's a third module I didn't mention which is a module that will figure out the best course of action to optimize an objective given your model >> okay >> call this a policy policy network or something like that right [snorts] now you need those three components to act autonomously intelligently and you can be stupid in three different ways you can be stupid because your model of your of the world is wrong. >> Mhm. >> You can be stupid because your objective is not aligned with what you actually want to achieve.

1:08:42>> Okay. Uh >> in humans that would be a psychopath, >> right? >> And then the the third thing you the third way you can be stupid is that you have the right model, you have the right objective, but you're unable to figure out a course of action to optimize your objective given your model. >> Right? >> Okay. Um some people who are in charge of big countries actually have all three that are wrong. All right. [laughter] Uh, which countries? I don't know. Okay. So, uh, if we think about this this agent, if we think about the movie her, you've criticized the art project that is Sophia the robot. And what that project essentially does is uses our natural inclination to anthropomorphize things that look like human and give

Sophia, embodiment, grounding, and emotions

1:09:30them more. Do you think that could be used by AI systems like in the movie Her? So, do you think that body is needed to create a feeling of intelligence? >> Uh, well, if Sophia was just an art piece, I would have no problem with it, but it's presented as something else. >> Let me add that comment real quick. If creators of Sophia could change something about their marketing or behavior in general, what would it be? What What's >> Oh, just about everything. [laughter] I mean, don't you think here's a tough question. Let me So, I agree with you. So, Sophia is not the general public feels that Sophia can do way more than she actually can. >> That's right. >> And the people who created Sophia are

1:10:20not honestly publicly communicating uh trying to teach the public. But >> here's a tough question. Don't you think uh this the same thing is uh scientists in industry and research are taking advantage of the same mis misunderstanding in the public when they create AI companies or publish stuff. um >> some companies yes I mean there is no sense of there's no desire to delude there's no desire to to kind of overclaim what something is done right you publish a paper on AI that you know has this result on imageet you know it's pretty clear I mean it's not not even interesting anymore but >> you know I I don't think there is that I mean the reviewers are generally not very forgiving of >> of you know unsupported claims of of

1:11:13this type and but there are certainly quite a few startups that have had a huge amount of hype around this that uh I find extremely damaging and I've been calling it out when I've seen it. Um so yeah, but to go back to your original question like >> the necessity of embodiment. I think I don't think embodiment is necessary. I think grounding is necessary. So I don't think we're going to get machines that really understand language without some level of grounding in the real world. And it's not clear to me that language is a high enough bandwidth medium to communicate how the real world works. Um I think for this >> can you talk to ground what grounding means? >> So grounding means that so there is this

1:11:51classic problem of common sense reasoning you know the the winner winner schema right and so I tell you the the trophy doesn't fit in the suitcase because it's too it's too big or the trophy doesn't fit in the suitcase because it's too small and the it in the first case refers to the trophy in the second case to the suitcase. And the reason you can figure this out is because you know what the trophy and the suitcase are. you know one is supposed to fit in the other one and you know the notion of size and a big object doesn't fit in a small object unless it's a TARDIS you know things like that right so you have this this knowledge of how the world works of geometry and things like that um I don't believe you can

1:12:30learn everything about the world by just being told in language how the world works I think you need some uh low-level perception of the world you know be it visual touch you know whatever but some higher bandwidth perception of the So by reading all the world's text, you still may not have enough information. >> That's right. There's a lot of things that just will never appear in text and that you can't really infer. So I think >> uh common sense will emerge from >> you know certainly a lot of language interaction but also with watching videos or perhaps even interacting in in virtual environments and possibly you know robot interacting in the real world but I don't actually believe necessarily

1:13:10that this last one is absolutely necessary but I think there's a the need for some grounding >> but the final product doesn't necessarily need to be embodied you're saying it just needs to have an awareness a grounding to >> but he needs to know how the world works to have you know to not be frustrated frustrating to talk to >> and uh you talked about emotions being important that's that's a whole another topic >> well so you know I talked about this uh the basil ganglia ganglia as the you your your you know the thing that calc you know calculates your level of mix contentment contentment and then there is this other module that sort of tries to do a prediction of whether you're going to be content or not.

1:13:55That's the source of some emotion. So fear, for example, is an anticipation of bad things that can happen to you, right? >> Uh you have this inkling that there is some chance that something really bad is going to happen to you and that creates fear. >> When you know for sure that something bad is going to happen to you, you kind of give up, right? It's not fear anymore. >> It's uncertainty that creates fear. >> So So the punch line is yes, we're not going to have autonomous intelligence without emotions. Okay. [sighs and gasps] >> Whatever the heck emotions are. So you mentioned very practical things of fear, but there's a lot of other mess around it. >> But there are kind of the results of, you know, drives.

1:14:33>> Yeah. There's deeper uh biological stuff going on and I've talked to a few folks on this. There's uh there's fascinating stuff that ultimately connects to our to to our brain. If we create an AGI system, sorry, >> human level intelligence, >> human level intelligence system and you get to ask her one question, what would that question be? >> You know, I think the the first one we'll create will probably not be not be that smart. They be like a four-year-old. Okay. >> So, you would have to ask her a question to know she's not that smart. >> Yeah. Well, what's a good question to ask, you know, to be >> the cause of wind? And if she answers, oh, it's because the leaves of the tree are moving and that

1:15:22creates wind, she's on to something. >> And if she says, that's a stupid question, she's really on to something. >> No. And then you tell her actually, you know, here is the the real thing. And she says, oh yeah, that makes sense. So questions that that uh reveal the ability to do common sense reasoning about the physical world. >> Yeah. And you'll sum it up with a causal inference. >> Causal inference. Well, >> it was a huge honor. Congratulations on the touring award, Yan. Thank you so much for talking today. >> Thank you. >> Appreciate it. [music]

1:16:04>> [music]