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

[0] The following is a conversation with Jan Lacoon.

[1] 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.

[2] He's a professor in New York University, a vice president and chief AI scientists of Facebook, and co -recipient of the Turing Award for his work on deep learning.

[3] He's probably best known as the founding father of convolutional neural network.

[4] In particular, their application to optical character recognition and the famed M. Nist dataset.

[5] He is also an outspoken personality, unafraid to speak his mind in a distinctive French accent, and explore provocative ideas both in the rigorous medium of academic research and the somewhat less rigorous medium of Twitter and Facebook.

[6] This is the Artificial Intelligence podcast.

[7] 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 Friedman, spelled F -R -I -D -M -A -N.

[8] And now here's my conversation with Jan Lacoon.

[9] You said that 2001 Space Odyssey is one of your favorite movies.

[10] Hal 9 ,000 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.

[11] Do you see Hal as flawed in some fundamental way or even evil?

[12] Or did he do the right thing?

[13] Neither.

[14] There's no notion of evil in that context other than the fact that people die.

[15] But it was an example of what people call value misalignment, right?

[16] You give an objective to a machine, and the machine strives to achieve this objective.

[17] And if you don't put any constraints on this objective, like don't kill people and don't do things like this, the machine given the power will do stupid things just to achieve this objective or damaging things to achieve its objective.

[18] It's a little bit like, I mean, we are used to this in the context of human society.

[19] We put in place laws to prevent people from doing bad things because spontaneously they would do those bad things.

[20] So we have to shape their cost function the 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 that point how you know there's a mission there's a this fuzziness around the ambiguity around what the actual mission is but you know do you think that there will be a 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.

[21] Well, that's the trick.

[22] I mean, eventually we'll have to figure out how to do this.

[23] And again, we're not starting from scratch because we've been doing this with humans for millennia.

[24] So designing objective functions for people is something that we know how to do.

[25] And we don't do it by programming things, although the legal code is called code.

[26] So that tells you something.

[27] And it's actually the design of an objective function.

[28] That's really what legal code is.

[29] It tells you here is what you can do.

[30] Here is what you can't do.

[31] If you do it, you pay that much.

[32] That's an objective function.

[33] 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.

[34] But no, we've been writing laws for millennia, and that's exactly what it is.

[35] So that's where, you know, the science of lawmaking and computer science will come together.

[36] We'll come together.

[37] So it's nothing, there's nothing special about how or AI systems is just the continuation of tools used to make some of these difficult ethical judgments that laws make.

[38] Yeah, and we have systems like this already that, you know, make many decisions for ourselves in society that, you know, need to be designed in a way that.

[39] that they, like, you know, rules about things that sometimes have bad side effects.

[40] 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.

[41] So you don't see this on the camera here, but all the decoration in this room is all pictures from 2001 in Space Odyssey.

[42] Wow.

[43] Is that by accident?

[44] Or is there a lot?

[45] It's not by accident.

[46] It's by design.

[47] Oh, wow.

[48] So if you were, if you were to build how.

[49] 10 ,000.

[50] So an improvement of Hal 9 ,000.

[51] What would you improve?

[52] Well, first of all, I wouldn't ask you to hold secrets and tell lies because that's really what breaks it in the end.

[53] That's the fact that it's asking itself questions about the purpose of the mission.

[54] 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 the discovery on the lunar surface that really was kept secret.

[55] And And one part of Hal'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.

[56] So you think there's never should be a set of things that an AI system should not be allowed, like a set of facts that should not be shared with the human operators?

[57] Well, I think, no, I think it should be a bit like in the design of autonomous AI systems.

[58] There should be the equivalent of, you know, the oath that hypocrite oaths that doctors sign up to, right?

[59] So there's certain things, certain rules that you have to abide by.

[60] And we can sort of hardwire this into our machines to kind of make sure they don't go.

[61] So I'm not, you know, an advocate of the three laws of robotics, you know, the Asimov kind of thing, because I don't think it's practical, but, but, you know, some level of limits.

[62] But to be clear, 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.

[63] We don't, we don't have autonomous intelligent machines.

[64] We have intelligent machines.

[65] Some are intelligent machines that are very specialized.

[66] but they don't really sort of satisfy an objective.

[67] They're just, you know, kind of trained to do one thing.

[68] So until we have some idea for design of a full -fledged autonomous intelligent system, asking the question of how we design this objective, I think, is a little too abstract.

[69] It's a little to abstract.

[70] There's useful elements to it in that it helps us understand our own ethical codes, humans.

[71] So even just as a thought experiment, if you imagine, that an AGI system is here today, how would we program it?

[72] Is a kind of nice thought experiment of constructing how should we have a law, have a system of laws for us humans.

[73] It's just a nice practical tool.

[74] 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.

[75] These things start creeping in that they're worth thinking about.

[76] but certainly they shouldn't be framed as how.

[77] 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?

[78] Sort of personally where you said back and just had this kind of, oh, that's pretty cool moment.

[79] That's nice, that's surprising.

[80] I don't know if it's an idea rather than a sort of empirical fact.

[81] The fact that you can build gigantic neural nets, trend them on, you know, relatively small amounts of data, relatively with stochastic grid and descent, and that it actually works, breaks everything you read in every textbook, right?

[82] Every pre -deep learning textbook that told you, you need to have fewer parameters and you have data samples.

[83] 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.

[84] A huge number of parameters, non -convex, and somehow it's just very relative to the number of parameters data, it's able to learn anything.

[85] Right.

[86] Does that still surprise you today?

[87] Well, it was kind of obvious to me before I knew anything that this is a good idea.

[88] and then it became surprising that it worked because I started reading those textbooks.

[89] Okay.

[90] So can you talk through the intuition of why it was obvious to you if you remember?

[91] Well, okay.

[92] So the intuition was it's sort of like, you know, those people in the late 19th century who proved that heavier than air flight was impossible, right?

[93] And of course you have birds, right?

[94] They do fly.

[95] And so on the face of it, it's obviously wrong as an empirical question, right?

[96] And so we have the same kind of thing that, you know, we know that the brain works.

[97] We don't know how, but we know it works.

[98] And we know it's a large network of neurons in interaction and that learning takes place by changing the connections.

[99] So kind of getting this level of inspiration without copying 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.

[100] There's also the idea somehow that I've been convinced of since I was an undergrad that even before.

[101] that intelligence is inseparable from learning.

[102] So the idea somehow that you can create an intelligent machine by basically programming, for me it was a non -starter from the start.

[103] Every intelligent entity that we know about arrives at this intelligence through learning.

[104] So learning, you know, machine learning was a completely obvious path.

[105] Also because I'm lazy, so, you know, You automate basically everything, and learning is the automation of intelligence.

[106] So do you think, so what is learning then?

[107] What falls under learning?

[108] Because do you think of reasoning as learning?

[109] Well, reasoning is certainly a consequence of learning as well, just like other functions of the brain.

[110] The big question about reasoning is, how do you make reasoning compatible?

[111] with gradient -based learning.

[112] Do you think neural networks can be made to reason?

[113] Yes, there is no question about that.

[114] Again, we have a good example, right?

[115] The question is how?

[116] 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?

[117] Another question is all of our kind of model of what reasoning is that are based on logic are discrete and are therefore incompatified.

[118] with gradient -based learning.

[119] And I'm a very strong believer in this idea of gradient -based learning.

[120] I don't believe that are the types of learning that don't use kind of gradient information if you want.

[121] So you don't like discrete mathematics?

[122] You don't like anything discrete?

[123] Well, it's not that I don't like it.

[124] It's just that it's incompatible with learning and I'm a big fan of learning, right?

[125] 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.

[126] The math that you use for deep learning, you know, it kind of has more to do with, you know, cybernetics, the kind of math you do in electrical engineering than the kind of math you do in computer science.

[127] And, you know, nothing in machine learning is exact, right?

[128] Computer science is all about sort of, you know, obviously, compulsive attention to details of like, you know, every index has to be right.

[129] And you can prove that an algorithm is correct, right?

[130] Machine learning is the science of sloppiness.

[131] really 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 works with continuous functions that's able to do build knowledge however we think about reasoning build on previous knowledge build on extra knowledge, create new knowledge, generalized outside of any training set ever built.

[132] What does that look like?

[133] If, yeah, maybe do you have inklings of thoughts of what that might look like?

[134] Yeah, I mean, yes and no. If I had precise ideas about this, I think, you know, we'd be building it right now.

[135] And there are people working on this, though, whose main research interest is actually exactly that, right?

[136] So what you need to have is a working memory.

[137] So you need to have some device, if you want, some subsystem that can store a relatively large number of factual episodic information for a reasonable amount of time.

[138] So in the brain, for example, there are kind of three main types of memory.

[139] One is the sort of memory of the state of your cortex.

[140] and that sort of disappears within 20 seconds.

[141] You can't remember things for more than about 20 seconds or a minute if you don't have any other form of memory.

[142] The second type of memory, which is longer term but still short term, is the hippocampus.

[143] So you came into this building.

[144] You remember where the exit is, where the elevators are.

[145] You have some map of that building that's stored in your hippocampus.

[146] You might remember something about what I said a few minutes ago.

[147] I forgot it all already.

[148] Of course, it's been erased.

[149] But, you know, but that would be in your hippocampus.

[150] And then the longer term memory is in the synapse, the synapses, right?

[151] So what you need if you want a system that's capable reasoning is that you want the hippocampus -like thing, right?

[152] And that's what people have tried to do with memory networks and, you know, neurotary machines and stuff like that, right?

[153] And now with transformers, which have sort of a memory in their kind of self -attention system.

[154] You can think of it this way.

[155] So that's one element you need.

[156] 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 then do this iteratively multiple times.

[157] Because a chain of reasoning is a process by which you update your knowledge about the set of the world, about, you know, what's going to happen, et cetera.

[158] And that has to be this sort of recurrent operation, basically.

[159] And you think that kind of, if we think about a transformer, so that seems to be too small to contain the knowledge that's, to represent the knowledge that's contained Wikipedia, for example.

[160] But a transformer doesn't have this idea of recurrence.

[161] It's got a fixed number of layers, and that's a number of steps that, you know, limits, basically, it's representation.

[162] But recurrence would build on the knowledge.

[163] somehow.

[164] I mean, it would evolve the knowledge and expand the amount of information, perhaps, or useful information within that knowledge.

[165] But is this something that just can emerge with size?

[166] Because it seems like everything we have now is too small.

[167] No, it's not clear.

[168] I mean, how you access and write into an associative memory inefficient way.

[169] I mean, so the original memory network maybe had something like the right architecture, but if you try to scale up a memory network, so that the memory contains all the Wikipedia, it doesn't quite work.

[170] So there's a need for new ideas there.

[171] But it's not the only form of reasoning.

[172] So there's another form of reasoning, which is very classical also in some types of AI.

[173] And it's based on, let's call it energy minimization.

[174] So you have some sort of objective, some energy function that represents the quality, or the negative quality, okay?

[175] Energy goes up when things get bad and they get low when things get good.

[176] So let's say you want to figure out, you know, what gestures do I need to do to grab an object or walk out the door?

[177] If you have a good model of your own body, a good model of the environment, using this kind of energy minimization, you can do planning.

[178] And it's in optimal control, it's called model predictive control.

[179] You have a model of what's going to happen in the world as a consequence of your actions.

[180] And that allows you to buy energy minimization, figure out a sequence of action that optimizes a particular objective function, which measures, you know, minimizes a number of times you're going to hit something and the energy you're going to spend doing the gesture and etc. So that's a form of reasoning.

[181] Planning is a form of reasoning.

[182] And perhaps what led to the ability of humans to reason is the fact that or, you know, 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.

[183] And so, you know, it's the same capacity that you need to have.

[184] So in your intuition is, if we look at expert systems and encoding knowledge as logic systems, as graphs in this kind of way is not a useful way to think about knowledge?

[185] Graphs are a little brittle or logic representation.

[186] So basically variables that have values and then constraint between them that are represented by rules is a little too rigid and too brittle.

[187] So some of the early efforts in that respect were to put probabilities on them.

[188] So a rule you know, if you have this and that symptom, you know, you have this disease with that probability and you should prescribe that antibiotic with that probability, right?

[189] That's the mycine system from the 70s.

[190] And that's what that branch of AI led to, you know, based on networks and graphical models and causal inference and variational, you know, method.

[191] So there is, I mean, certainly a lot of interesting work going on in this area.

[192] issue with this is knowledge acquisition.

[193] How do you reduce a bunch of data to a graph of this type?

[194] Yeah, relies on the expert on a human being to encode, to add knowledge.

[195] And that's essentially impractical.

[196] Yeah.

[197] So that's a big question.

[198] The second question is, do you want to represent knowledge as symbols and do you want to manipulate them with logic?

[199] And again, that's incompatible we're learning.

[200] So one suggestion with Jeffington has been advocating for many decades is replace symbols by vectors.

[201] Think of it as pattern of activities in a bunch of neurons or units or whatever you want to call them.

[202] And replace logic by continuous functions.

[203] And that becomes now compatible.

[204] There's a very good set of ideas by written in a paper about 10 years ago by Leon Boutou on, who is here at Facebook.

[205] The title of the paper is for machine learning to machine reasoning.

[206] And his idea is that a learning system should be able to manipulate objects that are in a space and then put the result back in the same space.

[207] So it's this idea of working memory, basically.

[208] And it's very enlightening.

[209] And in a sense, that might learn something like the simple expert systems.

[210] I mean, you can learn basic logic operations there.

[211] Yeah, quite possibly.

[212] There's a big debate on sort of how much prior structure you have to put in for this kind of stuff to emerge.

[213] That's the debate I have with Gary Marcus and people like that.

[214] Yeah, yeah.

[215] So, and the other person, so I just talked to Judea Pearl, when you mentioned causal inference world.

[216] So his worry is that the current neural networks are not able to learn what causes what causal inference between things?

[217] So I think he's right and wrong about this.

[218] If he's talking about the sort of classic type of neural nets, people sort of didn't worry too much about this.

[219] But there's a lot of people now working on causal inference, and there's a paper that just came out last week by Leone Boutou, among others, David Lopez -Baz and bunch of other people, exactly on that problem of how do you kind of, you know, get an neural net to sort of pay attention to real causal relationships, which may also solve issues of bias in data and things like this.

[220] I'd like to read that paper because that ultimately the challenge is 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 all the all the basic clause of microphysics are time reversible so there is no causality the arrow of time is not real right yeah it's as soon as you start looking at macroscopic systems where there is unpredictable randomness where there is clearly an hour of time but it's a big mystery in physics actually well how that emerges is it emerging or is it part of the fundamental fabric of reality?

[221] Or is it a bias of intelligent systems that, you know, because of the second law of thermodynamics, we perceive a particular hour of time, but in fact, it's kind of arbitrary, right?

[222] So yeah, physicists, mathematicians, they don't care about, I mean, the math doesn't care about the flow of time.

[223] Well, certainly microphysics doesn't.

[224] People themselves are not very good at establishing causal relationships.

[225] If you ask this, I think it was in one of Seymour Papert's book on, like, children learning.

[226] You know, he studied with Jean Piaget, you know.

[227] He's the guy who co -authored the book Perceptron with Marvin Minsky that kind of killed the first wave of neural nets.

[228] But he was actually a learning person.

[229] He, in the sense of studying learning in humans and machines, that's why he got interested in the perceptron.

[230] And he wrote that, if you ask a little kid, about what is the cause of the wind.

[231] A lot of kids will say, they will think for a while and they'll say, oh, it's the branches and the trees.

[232] They move and that creates wind, right?

[233] So they get the causal relationship backwards.

[234] And it's because their understanding of the world and intuitive physics is not that great, right?

[235] I mean, these are like, you know, four or five -year -old kids.

[236] You know, it gets better, and then you understand that this can't be, right?

[237] But there are many things which we can, I mean, because of our common sense understanding of things, what people call common sense, and with understanding of physics, we can, there's a lot of stuff that we can figure out causality.

[238] Even with diseases, we can figure out what's not causing what often.

[239] There's a lot of mystery, of course, but the idea is that you should be able to encode that into systems, because it seems unlikely to be able to figure that out themselves.

[240] 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 you to some deity, some divinity, right?

[241] And that's a cop -out.

[242] That's a way of saying, like, I don't know the cause, so, you know, God did it, right?

[243] So you mentioned Marvin Minsky and the irony of, you know, maybe causing the first day I winter.

[244] You were there in the 90s, You were there in the 80s, of course.

[245] In the 90s, why do you think people lost faith in deep learning in the 90s and found it again a decade later, over a decade later?

[246] Yeah, it wasn't called deep learning yet.

[247] It was just called neural nets.

[248] Yeah, the lost interests.

[249] I mean, I think I would put that around 1995, at least the machine learning community.

[250] There was always a neural net community, but it became kind of disconnected from sort of mainstream machine learning.

[251] you want.

[252] There were, it was basically electrical engineering that kept at it.

[253] Right.

[254] And computer science.

[255] Just gave up.

[256] Gave up on neural nets.

[257] I don't, I don't know.

[258] You know, I was too close to it to really sort of analyze it with sort of an unbiased eye if you want.

[259] But I would, I would make a few guesses.

[260] 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 backprop in your favorite language.

[261] And that favorite language was not Python.

[262] It was not Matt Lab.

[263] It was not any of those things because they didn't exist, right?

[264] You had to write it in Fortyno C or something like this, right?

[265] So you would experiment with it.

[266] You would probably make some very basic mistakes, like, you know, badly initialize your weights, make the network too small because you read in the textbook, you know, you don't want too many right?

[267] And of course, you know, and you would train on XOR because you didn't have any other data set to trade on.

[268] And of course, you know, it works half the time.

[269] So we would say, I give up.

[270] Also, we would train it with batch gradient, which, you know, isn't less efficient.

[271] So there was a lot of, there was a bag of tricks that you had to know to make those things work or you had to reinvent.

[272] And a lot of people just didn't and they just couldn't make it work.

[273] So that's one thing.

[274] The investment in software platform to be able to kind of, you know, display things, figure out why things don't work, kind of get a good intuition for how to get them to work, have enough flexibility so you can create, you know, network architectures like convolutional nets and stuff like that.

[275] It was hard.

[276] I mean, you had to write everything from scratch.

[277] And again, you didn't have any Python or Matlab or anything, right?

[278] I read that, sorry to interrupt, but I read that you wrote in Lisp, your first versions of Linette with convolution neural networks, which, by the way, one of my favorite languages, that's how I knew you were legit.

[279] JET, touring award, whatever.

[280] You programmed in LISP.

[281] It's still my favorite language.

[282] But it's not that we programmed in LISP, it's that we had to write our LISP interpreter.

[283] Okay, because it's not like if we used one that existed.

[284] So we wrote a LISP interpreter that we hooked up to, you know, a backend library that we wrote also for sort of neural -nate computation.

[285] And then after a few years around 1991, we invented this idea of basically having modules that know how to forward -propagate and back - propagate gradients, and then interconnecting those modules in a graph.

[286] Leon Boutou had made proposals on this, about this in the late 80s, and we're able to implement this using our list system.

[287] Eventually, we wanted to use that system to make build production code for character recognition at Bell Labs, so we actually wrote a compiler for that Lisp interpreter, so that Petrisimard, who is not Microsoft, kind of did the bulk of it with Leone and me, and so we could write our system in Lisp and then compile to C. and then we'll have a self -contained complete system that could kind of do the entire thing.

[288] Neither PyTorch North and Sanfluf can do this today.

[289] Yeah, okay.

[290] It's coming.

[291] Yeah.

[292] I mean, there's something like that in PyTorch called, you know, TorScript.

[293] And so, you know, we had to write to release metropolis.

[294] We had to write to release compiler.

[295] We had to invest a huge amount of effort to do this.

[296] And not everybody, if you don't completely believe in the concept, you're not going to invest a time to do this.

[297] this.

[298] Now, at the time also, you know, or today, this would turn into torch or pie torch or transfer or whatever.

[299] We'd put it in open source.

[300] Everybody would use it and, you know, realize it's good.

[301] Back before 1995, working at AT &T, there's no way the lawyers would let you release anything in open source of this nature.

[302] And so we could not distribute our code, really.

[303] And on that point, and sorry to go on a million tangents, but on that point, I also read that there There was some almost patent on Kamolishu and your networks at Bell Labs.

[304] So that, first of all, I mean, just two actually.

[305] That ran out.

[306] Thankfully, in 2007.

[307] In 2007.

[308] What, can we, can we just talk about that first?

[309] I know you're Facebook, but you also at NYU.

[310] And what does it mean to patent ideas like these?

[311] software ideas essentially or what are mathematical ideas or what are they okay so they're not mathematical ideas so there are you know algorithms and there was a period where the u .s. patent office would allow the patent of software as long as it was embodied 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 believe believed in this, but I don't believe in this kind of patent.

[312] Facebook basically doesn't believe in this kind of patent.

[313] Google files patents because they've been burned with Apple, and so now they do this for defensive purpose, but usually they say, we're not going to see you if you infringe.

[314] Facebook has a similar policy.

[315] They say, you know, we have a patent on certain things for defensive purpose.

[316] We're not going to see you if you infringe, unless you through us.

[317] so the industry does not believe in patents they are there because of you know the legal landscape and and various things but but i don't really believe in patterns for this kind of stuff okay so that's that's a great thing so i i'll tell you a worse story actually so what happens was the the first the first patent about convolution net was about kind of the early version of convolutional net that didn't have separate pulling layers it had you know convolutional layers was tried more than one, if you want, right?

[318] And then there was a second one on convolution nets with separate pulling layers trained with backprop.

[319] And there were files filed in 89 and 1990 or something like this.

[320] At the time, the life of a patent was 17 years.

[321] So here's what happened over the next few years is that we started developing character recognition technology around convolution nets.

[322] And in 1994, a check.

[323] reading system was deployed in ATM machines.

[324] In 1995, it was for large check reading machines in back offices, et cetera.

[325] 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.

[326] Now, AT &T split up in 1996, early 1996.

[327] And the lawyers just looked at all the patents and they distributed the patents among the various companies.

[328] They gave the Cornuconnet pattern to NCR because they were actually selling products that used it.

[329] But nobody had NCR had any idea what a commercial net was.

[330] Yeah.

[331] So between 1996 and 2007, so there's a whole period until 2002 where I didn't actually work on machine learning or commercial net.

[332] I resumed working on this around 2002.

[333] And between 2002 and 2007, I was working on them, crossing my finger that nobody had NCR would notice.

[334] And nobody noticed.

[335] Yeah, and I hope that this kind of somewhat, as you said, lawyers aside, relative openness of the community now will continue.

[336] It accelerates the entire progress of the industry.

[337] And, you know, the problems that Facebook and Google and others are fitting today is not whether Facebook or Google or Microsoft or IBM or whoever is ahead of the other, is that we don't have the technology to build these things we want to build.

[338] we want to build intelligent virtual assistance that have common sense.

[339] We don't have monopoly on good ideas for this.

[340] We don't believe we do.

[341] Maybe others believe they do, but we don't.

[342] If a startup tells you they have the secret to, you know, human level intelligence and common sense, don't believe them.

[343] They don't.

[344] And it's going to take the entire work of the world research community for a while to get to the point where you can go off and in each of those companies kind of start to build things on this.

[345] we're not there yet.

[346] Absolutely.

[347] And this calls to the gap between the space of ideas and the rigorous testing of those ideas of practical application that you often speak to.

[348] You've written advice saying, don't get fooled by people who claim to have a solution to artificial general intelligence, who claim to have an AI system that work just like the human brain or who claim to have figured out how the brain works.

[349] Ask them what the error rate they get on MNIST or ImageNet.

[350] Yeah, this is a little dated, by the way.

[351] I mean, five years, who's counting?

[352] Okay.

[353] But I think your opinion is the MNIST and ImageNet, yes, may be dated.

[354] There may be new benchmarks, right?

[355] But I think that philosophy is when you still somewhat hold that benchmarks and the practical testing, the practical application is where you really get to test the ideas.

[356] Well, it may not be completely practical.

[357] Like, for example, you know, it could be a toy dataset, but it has 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.

[358] It doesn't need to be real.

[359] So, for example, many years ago here at Fair, people, you know, Cheson West and Antoine Bonn and a few others proposed the Babi tasks, which were kind of a toy problem to test the ability of machines to reason, actually, to access working memory and things like this.

[360] And it was very useful, even though it wasn't a real task.

[361] Amnist is kind of halfway a real task.

[362] So, you know, toy problems can be very useful.

[363] 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.

[364] Yes, absolutely.

[365] So there's a lot of people who try to take advantage of the hype for business reasons and so on.

[366] But let me sort of talk to this idea that new ideas, the ideas that push the field forward may not yet have a benchmark.

[367] Or it may be very difficult to establish a benchmark.

[368] I agree.

[369] That's part of the process.

[370] Establishing benchmarks is part of the process.

[371] 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.

[372] There's audio data sets, there's some video.

[373] What can we start natural language?

[374] What kind of stuff, what kind of benchmarks do you see that start creeping on to more something like intelligence, like reasoning, like maybe you don't like the term, but AGI echoes of that kind of formulation?

[375] A lot of people are working on interactive environments in which you can train and test intelligent systems.

[376] So there, for example, you know, it's the classical paradigm of supervised learning is that you have a data set, you partition it into a training set, validation set, test set, and there's a clear protocol, right?

[377] But what if the, that assumes that the samples are statistically independent.

[378] You can exchange them, the order in which you see them doesn't, It shouldn't matter, you know, things like that.

[379] But what if the answer you give determines the next sample you see, which is the case, for example, in robotics, right?

[380] You robot does something and then it gets exposed to a new room, and depending on where it goes, the room would be different.

[381] So that's the, that creates the exploration problem.

[382] The, what if the samples, so that creates also a dependency between samples, right?

[383] If you move, if you can only move in space, the next sample you're going to see is going to be probably in the same building, most likely.

[384] So all the assumptions about the validity of this training set test set hypothesis break.

[385] Whenever a machine can take an action that has an influence in the world, and it's what is going to see.

[386] So people are setting up artificial environments where that takes place, right?

[387] The robot runs around a 3D model.

[388] over house and can interact with objects and things like this.

[389] So you do robotics by simulation, you have those, you know, opening eye gym type thing or Mujoko kind of simulated robots and you have games, you know, things like that.

[390] So that's where the field is going, really, this kind of environment.

[391] Now, back to the question of AGI.

[392] Like, I don't like the term AGI because it implies that human intelligence is general.

[393] And human intelligence is nothing like general.

[394] It's very, very specialized.

[395] We think it's general.

[396] We like to think of ourselves as having general intelligence.

[397] We don't.

[398] We're very specialized.

[399] We're only slightly more general than...

[400] Why does it feel general?

[401] So you're kind of the term general.

[402] I think what's impressive about humans is the ability to learn, as we were talking about learning, to learn in just so many different domains.

[403] It's perhaps not arbitrarily general, but just you can learn in many domains and integrate that knowledge somehow.

[404] Okay.

[405] The knowledge persists.

[406] So let me take a very specific example.

[407] Yes.

[408] It's not an example.

[409] It's more like a quasi -mathematical demonstration.

[410] So you have about one million fibers coming out of one of your eyes, okay, two million total.

[411] But let's talk about just one of them.

[412] It's one million nerve fibers, your optical nerve.

[413] Let's imagine that they are binary, so they can be active or inactive, right?

[414] So the input to your visual cortex is 1 million bits.

[415] Now, they're connected to your brain in a particular way, and your brain has connections that are kind of a little bit like a convolutional net.

[416] They're kind of local, you know, in space and things like this.

[417] Now, imagine I play a trick on you.

[418] It's a pretty nasty trick, I admit.

[419] I cut your optical nerve, and I put a device that makes a random perturbation of permutation of all the nerve fibers.

[420] So now what comes to your brain is a fixed.

[421] but random permutation of all the pixels.

[422] 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.

[423] Got it.

[424] And you're saying there's no way you've relearn that.

[425] No, because now two pixels that are nearby in the world will end up in very different places in your visual cortex.

[426] And your neurons there have no connections with each other because they only connect it locally.

[427] So this whole, our entire, the hard word, is built in many ways to support the locality of the real world.

[428] Yeah.

[429] Yes, that's specialization.

[430] Yeah, but it's still pretty damn impressive.

[431] So it's not perfect generalization.

[432] It's not even close.

[433] No, no, it's not that it's not even close.

[434] It's not at all.

[435] Yeah, it's specialized, yeah.

[436] So how many Boolean functions?

[437] So let's imagine you want to train your visual system to, you know, recognize particular patterns of those 1 million bits.

[438] Okay, so that's a Boolean function, right?

[439] Either the pattern is here or not here.

[440] It's a two -way classification with 1 million binary inputs.

[441] How many such Boolean functions are there?

[442] You have 2 to the 1 million combinations of inputs.

[443] For each of those, you have an output bit.

[444] And so you have 2 to the 1 million bullion functions of this type, which is an unimaginably large number.

[445] How many of those functions can actually be computed your original cortex.

[446] And the answer is a tiny, tiny, tiny, tiny, tiny, tiny, tiny, tiny, sliver, like an enormously tiny sliver.

[447] Yeah.

[448] Yeah.

[449] So we are ridiculously specialized.

[450] Okay.

[451] But, okay, that's an argument against the word general.

[452] I think there's a, I, I agree with your intuition, but I'm not sure it's, it seems the brain is impressively, um, capable of adjusting to things.

[453] It's because we can't imagine tasks that are outside of our comprehension.

[454] So we think we think we're general because we're general of all the things that we can apprehend.

[455] But there is a huge world out there of things that we have no idea.

[456] We call that heat, by the way.

[457] Heat.

[458] So, at least physicists, it's heat.

[459] Or they call it entropy, which is kind of.

[460] You have a thing full of gas, right?

[461] Close system for gas.

[462] Right?

[463] Close or not closed?

[464] It has pressure, it has temperature, it has, you know, and you can write equations PV equal NRT, you know, things like that, right?

[465] When you reduce the volume, the temperature goes up, the pressure goes up, you know, things like that, right?

[466] For perfect gas, at least.

[467] Those are the things you can know about that system.

[468] And it's a tiny, tiny number of bits compared to the complete information of the state of the entire system.

[469] Because the state of the entire system will give you the position and momentum of every molecule of the gas.

[470] And what you don't know about it is the entropy, and you interpret it as heat.

[471] The energy containing that thing is what we call heat.

[472] Now, it's very possible that, in fact, there is some very strong structure in how those molecules are moving.

[473] It's just that they are in a way that we are just.

[474] just not wired to perceive.

[475] Yeah, we're ignorant to it.

[476] And there's, in your infinite amount of things we're not wired to perceive.

[477] And you're right, that's a nice way to put it.

[478] We're general to all the things we can imagine, which is a very tiny subset of all things that are possible.

[479] It's like comagraph complexity or the combative chat in someone of complexity.

[480] You know, every bit string or every integer is random, except for all the ones that you can actually write down.

[481] Yeah, okay, so beautiful people.

[482] So we can just call it artificial intelligence.

[483] We don't need to have a general.

[484] Or human level.

[485] Human level intelligence is good.

[486] You'll start, anytime you touch human, it gets interesting because, you know, it's because we attach ourselves to human and it's difficult to define what human.

[487] intelligences.

[488] Nevertheless, my definition is maybe a dem -impressive intelligence.

[489] Okay, damn -impressive demonstration of intelligence, whatever.

[490] And so on that topic, most successes in deep learning have been in supervised learning.

[491] What is your view on unsupervised learning?

[492] Is there a hope to reduce involvement of human input and still have successful systems that have practically use?

[493] Yeah, I mean, there's definitely a hope.

[494] It's more than a hope, actually.

[495] It's, you know, mounting evidence for it.

[496] And that's basically all I do, like, the only thing I'm interested in at the moment is, I call it self -supervised learning, not unsupervised learning, because unsupervised learning is a loaded term.

[497] People who know something about machine learning, you know, tell you, so you're doing clustering or PCA, which is not the case.

[498] And the way public, you know, when you say unsupervised learning, oh, my God, you know, machines are going to learn, themselves and without supervision.

[499] You know, they see this as...

[500] Where's the parents?

[501] Yeah.

[502] So I call itself supervised learning because, in fact, the underlying algorithms that are used are the same algorithms as the supervised learning algorithms, except that what we train them to do is not predict a particular set of variables, like the category of an image, and not to predict a set of variables that have been provided by human label.

[503] But what you train the machine to do is basically reconstruct a piece of its input, that it's being masked out, essentially.

[504] You can think of it this way, right?

[505] So show a piece of a video to a machine and ask it to predict what's going to happen next.

[506] And of course, after a while, you can show what happens, and the machine will kind of train itself to do better at that task.

[507] You can do, like all the latest most successful models in natural language processing, use self -supervised learning.

[508] you know, sort of birth -style systems, for example, right?

[509] You show it a window of a thousand words on a text corpus.

[510] You take out 15 % of the words, and then you train the machine to predict the words that are missing.

[511] That's self -supervise learning.

[512] It's not predicting the future.

[513] It's just, you know, predicting things in the middle.

[514] But you could have you predict the future.

[515] That's what language models do.

[516] So you construct, so in an unsupervised way, you construct a model of language.

[517] do you think...

[518] Or video or the physical world or whatever, right?

[519] How far do you think that can take us?

[520] Do you think Bart understands anything?

[521] To some level, it has a shallow understanding of text, but it needs to, I mean, to have kind of true human level intelligence, I think you need to ground language in reality.

[522] So some people are attempting to do this, right?

[523] having the systems that kind of have some visual representation of what is being talked about, which is one reason you need those interactive environments, actually.

[524] But there's like a huge technical problem that is not solved, and that explains why self -supervasionally 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.

[525] And the reason, that reason is the fact that it's much easier to work, represent uncertainty in the prediction in the context of natural language than it is in the context of things like video and images.

[526] So, for example, if I ask you to predict what words are missing, you know, 15 % of the words that have taken out.

[527] The possibility is small.

[528] It's small, right?

[529] There is 100 ,000 words in the lexicon, and what the machine spits out is a big probability vector, right?

[530] It's a bunch of numbers between zero and one that's on to one.

[531] and we know how to do this with computers.

[532] So there, representing uncertainty in the prediction is relatively easy, and that's, in my opinion, why those techniques work for NLP.

[533] For images, 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?

[534] And how do you represent this set of possible answers?

[535] You can't train a system to make one prediction.

[536] You can train a neural net to say, here it is.

[537] That's the image because there's a whole set of things that are compatible with it.

[538] So how do you get the machine to represent not a single output, but a whole set of outputs?

[539] And, you know, similarly with video prediction, there's a lot of things that can happen in the future of a video.

[540] You're looking at me right now.

[541] I'm not moving my head very much, but, you know, I might, you know, turn my head to the left or to the right.

[542] If you don't have a system that can predict this and you train it with Lee Square to kind of minimize the error with a prediction and what I'm doing, what you get is a blurry image of myself in all possible future positions that I might be in, which is not a good prediction.

[543] So there might be other ways to do the self -supervision, right, for visual scenes.

[544] Like what?

[545] If I knew I wouldn't tell you, I didn't publish it first.

[546] I don't know.

[547] No, there might be.

[548] It might be.

[549] So, I mean, these are kind of, there might be artificial ways of like self -play in games to where you can simulate part of the environment.

[550] You can...

[551] Oh, that doesn't solve the problem.

[552] It's just a way of generating data.

[553] But because you have more of a control, like, maybe you can control, yeah, it's a way to generate data.

[554] That's right.

[555] And because you can do huge amounts of data generation, that doesn't, you're right.

[556] Well, it's a creeps up on the problem from the side of data.

[557] And you don't think that's the right way to creep up.

[558] It doesn't solve this problem of handling uncertainty in the world, right?

[559] So if you have a machine, learn a predictive model of the world in a game that is deterministic or quasi -deterministic, it's easy.

[560] Right?

[561] Just, you know, give a few frames of the game to a conventette, put a bunch of layers, and then half the game generates the next few frames.

[562] And if the game is deterministic, it works fine.

[563] And that includes, you know, feeding the system with the action that your little character is going to take.

[564] The problem comes from the fact that the real world and most games are not entirely predictable.

[565] And so there you get those blurry predictions, and you can't do planning with blurry predictions.

[566] 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, and you're going to predict the outcome of that sequence of actions?

[567] But if your model is imperfect, how can you plan?

[568] Yeah, it quickly explodes.

[569] What are your thoughts on the extension of this, which topic I'm super excited about?

[570] It's connected to something you were talking about in terms of robotics, is active learning.

[571] So as opposed to completely unsupervised or self -supervised learning, you asked the system, for human help, for selecting parts you want to annotated next.

[572] 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.

[573] Do you see value in that kind of work?

[574] I don't see transformative value.

[575] It's going to make things that we can already do more efficient, or they will learn slightly more efficiently.

[576] But it's not going to make machines sort of significantly more intelligent.

[577] I think, and by the way, there is no opposition, there's no conflict between self -supervised learning, reinforcement learning and supervised learning, or imitation learning, or active learning.

[578] I see self -supervice learning as a preliminary to all of the above.

[579] Yes.

[580] 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.

[581] The best methods today, so -called model -free reinforcement learning, to learn to play Atari games, take about 80 hours of training to reach the level that any human can reach in about 15 minutes.

[582] They get better than humans, but it takes them a long time.

[583] Alpha star, okay, the, you know, are your vinyas, and his team's system to play StarCraft plays, you know, a single map, a single type of player and can reach better than human level with about the equivalent of 200 years of training playing against itself.

[584] It's 200 years, right?

[585] It's not something that no human can could ever do.

[586] I mean, I'm not sure what less than take away from that.

[587] Okay.

[588] Now, take those algorithms, the best oral algorithms we have today, to train a car to drive itself.

[589] It would probably have to drive millions of hours.

[590] It will have to kill thousands of pedestrians.

[591] It will have to run into thousands of trees.

[592] It will have to run off cliffs.

[593] And it had to run off cliff multiple times before it figures out that it's a bad idea, first of all.

[594] And second of all, before it figures out how not to do it.

[595] And so, I mean, this type of learning obviously does not reflect the kind of learning that animals and humans do.

[596] There is something missing that's really, really important there.

[597] 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.

[598] 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.

[599] 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.

[600] Because we have a pretty good model of intuitive physics that tells us, you know, the car is going to fall.

[601] We know about gravity.

[602] Babies learn this around the age of eight or nine months that objects don't float.

[603] They fall.

[604] 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 stand the road.

[605] So there's a lot of things that we bring to the table, which is basically, or predictive model of the world.

[606] And that model allows us to not do stupid things and to basically stay within the context of things we need to do.

[607] We still face unpredictable situations, and that's how we learn.

[608] But that allows us to learn really, really, really quickly.

[609] So that's called model -based reinforcement learning.

[610] There's some imitation and supervised learning because we have a driving instructor that tells us occasionally what to do.

[611] But most of the learning is model -based.

[612] is learning the model, learning physics that we've done since we were babies.

[613] That's where almost all are learning.

[614] And the physics is somewhat transferable from, it's transferable from scene to scene.

[615] Stupid things are the same everywhere.

[616] Yeah.

[617] I mean, if you, you know, you have an experience of the world.

[618] You don't need to be particularly from a particularly intelligent species to know that if you spill water from a container, you know, the rest is going to get wet.

[619] You might get wet.

[620] So, you know, cats know this, right?

[621] Yeah.

[622] So the main problem we need to solve is how do we learn models of the world?

[623] And that's what I'm interested.

[624] That's what self -supervising is all about.

[625] If you were to try to construct a benchmark for, let's look at MNIS.

[626] I love that data set.

[627] Do you think it's useful, interesting, slash possible, to perform well on MNIS?

[628] with just one example of each digit.

[629] And how would we solve that problem?

[630] The answer is probably yes.

[631] The question is what other type of learning are you allowed to do?

[632] So if what you're allowed to do is train on some gigantic data set of labeled digit that's called transfer learning.

[633] And we know that works, okay?

[634] We do this at Facebook, like in production, right?

[635] We train large convolutionness to predict hashtags that people type on Instagram and we train on billions of images, literally billions.

[636] And then we chop off the last layer and fine tune on whatever task we want.

[637] That works really well.

[638] You can beat, you know, the ImageNet record with us.

[639] We actually open source the whole thing like a few weeks ago.

[640] Yeah, that's still pretty cool.

[641] But yeah, so what would be impressive?

[642] What's useful and impressive?

[643] What kind of transfer learning would be useful and impressive?

[644] Is it Wikipedia, that kind of thing?

[645] No, no. So I don't think transfer learning is really where we should focus.

[646] We should try to do, you know, have a kind of scenario for a benchmark where you have unlabeled data, and you can, and it's a very large number of unlabel data.

[647] It could be video clips, it could be where you do frame prediction, it could be images, where you could choose to, you know, mask a piece of it.

[648] It could be whatever, but they're unlabeled and you're not allowed to label them.

[649] So you do some training on this, and then you train on a particular supervised task, ImageNet or Enlist, and you measure how your test error, decrease, or validation error, decreases as you increase the number of label training samples.

[650] Okay, what you'd like to see is that your error decreases much faster than if you train from scratch, from random weights.

[651] So that to reach the same level of performance and a completely supervised, purely supervised system would reach, you would need way fewer samples.

[652] So that's the crucial question, because it will answer the question to, like, you know, people are interested in medical image analysis.

[653] Okay, you know, if I want to get to a particular level of error rate for this task, I know I need a million samples.

[654] Can I do, you know, self -supervised pre -training to reduce this to about 100 or something?

[655] And you think the answer there is self -supervised pre -training?

[656] Yeah.

[657] some form, some form of it.

[658] Telling you, active learning, but you disagree.

[659] No, it's not useless.

[660] It's just not going to lead to a quantum leap.

[661] It's just going to make things that we already do.

[662] So you're way smarter than me. I just disagree with you.

[663] But I don't have anything to back that.

[664] It's just intuition.

[665] So I worked with a lot of large -scale data sets, and there's something that might be magic in active learning.

[666] But, okay, at least I said it publicly.

[667] 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 learning for a start driving car but most of those are incredibly 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 active learning if you want yes but there might be a much stronger version.

[668] Yeah, that's right.

[669] And that's an open question if it exists.

[670] The question is how much stronger it can get.

[671] Elon Musk is confident.

[672] I talked to him recently.

[673] He's confident that large -scale data and deep learning can solve the autonomous driving problem.

[674] What are your thoughts on the limit's possibilities of deep learning in this space?

[675] It's obviously part of the solution.

[676] I mean, I don't think we'll ever have a set driving system, or at least not in the foreseeable future that does not use deep learning.

[677] Maybe put it this way.

[678] Now, how much of it?

[679] So in the history of sort of engineering, particularly sort of AI -like systems, there's generally a first phase where everything is built by hand, and there is a second phase, and that was the case for autonomous driving, you know, 20, 30 years ago.

[680] 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, and, you know, putting limits, et cetera, because the learning system is not perfect.

[681] And then as technology progresses, we end up relying more and more on learning.

[682] That's the history of character recognition, the history of speech recognition, now computer vision, natural language processing.

[683] And I think the same is going to happen with autonomous driving, that currently the methods that are closest to providing some level of autonomy, some, you know, decent level of autonomy, where you don't expect a driver to kind of do anything.

[684] is where you constrain the world.

[685] So you only run within 100 square kilometers or square miles in Phoenix.

[686] But the weather is nice and the roads are wide, which is what Waymo is doing.

[687] You completely over -engineer the car with tons of LIDARs and sophisticated sensors that are too expensive for consumer cars, but they're fine if you just run a fleet.

[688] And you engineer the hell out of everything else.

[689] you map the entire world, so you have a complete 3D model of everything.

[690] So the only thing that the perception system has to take care of is moving objects and construction and sort of things that weren't in your map.

[691] And you can engineer a good slam system and all that stuff, right?

[692] So that's kind of the current approach that's closest to some level of autonomy.

[693] 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.

[694] But ultimately, learning will be not just at the core, but really the fundamental part of the system.

[695] Yeah, it already is, but it'll become more and more.

[696] What do you think it takes to build a system with human -level intelligence?

[697] You talked about the AI system in the movie, her being way out of reach, our current reach.

[698] This might be outdated as well, but...

[699] It's still way out of reach.

[700] It's still way out of reach.

[701] what would it take to build her, do you think?

[702] 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.

[703] So the image I usually use is that there is a bunch of mountains that we have to climb and we can see the first one, but we don't know if there are 50 mountains behind it or not.

[704] And this might be a good sort of metaphor for why AI researchers in the past have been overly optimistic about the result of AI.

[705] you know for example new one and simon wrote the general problem solver and they called it the 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 in general what are the first couple of peaks for her so the first peak which is precisely what i'm working on is self -supervasioning how do we get machines to learn models of the world by observation, kind of like babies and like young animals.

[706] So we've been working with, you know, cognitive scientists.

[707] So there's Emmanuel Dupu, who is at fair in Paris, is half -time.

[708] He's also a researcher in French university.

[709] And he has this chart that shows at which how many months of life baby humans kind of learn different concepts.

[710] And you can measure this in various ways.

[711] So things like distinguishing animate objects from inanimate object, you can tell the difference at age two, three months.

[712] Whether an object is going to stay stable is going to fall, you know, about four months you can tell.

[713] You know, there are various things like this.

[714] And then things like gravity, the fact that objects are not supposed to float in the but are supposed to fall, you run this around the age of eight or nine months.

[715] If you look at a lot of, you know, eight -month -old babies, you give them a bunch of toys on their high chair.

[716] First thing they do is they throw them on the ground that they look at them.

[717] It's because, you know, they're learning about, actively learning, about gravity.

[718] Gravity, yeah.

[719] Okay.

[720] So they're not trying to annoy you, but they, you know, they need to do the experiment, right?

[721] So, you know, how do we get machines to learn like babies, mostly by observation with a little bit of interaction?

[722] and learning those models of the world, because I think that's really a crucial piece of an intelligent autonomous system.

[723] So if you think about the architecture of an intelligent autonomous system, it needs to have a predictive model of the world.

[724] So something that says, here is the world at time T, here is a set of the world at time T plus one if I take this action.

[725] And it's not a single answer.

[726] It can be a bunch of things.

[727] Well, we don't know how to represent distributions in high -dimensional continuous spaces.

[728] So it's going to be something weaker than that, okay.

[729] But with some representation, uncertainty.

[730] If you have that, then you can do what optimal control theory is called model predictive control, which means that you can run your model with a hypothesis for a sequence of action and then see the result.

[731] Now, what you need, the other thing you need is some sort of objective that you want to optimize.

[732] Am I reaching the goal of grabbing this object?

[733] Am I minimizing energy?

[734] Am I whatever, right?

[735] So there is some sort of objective that you have to minimize.

[736] And so in your head, if you have this model, you can figure out the sequence of action that will optimize your objective.

[737] That objective is something that ultimately is rooted in your basal ganglia, at least in the human brain.

[738] That's what it's.

[739] Bezel ganglia computes your level of contentment or miscontentment.

[740] I don't know if that's a word.

[741] Unhappiness, okay?

[742] Discontentment.

[743] Discontentment, maybe.

[744] And so your entire behavior is driven towards kind of minimizing that objective, which is maximizing your contentment, computed by your basoganglia.

[745] And what you have is an objective function which is basically a predictor of what your basogglia is going to tell you.

[746] 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.

[747] And you're predicting this because of your model of the world and your sort of predictor of this objective, right?

[748] So if you have those, you have those three components, you have four components, you have the the hardwired contentment objective computer, if you want, calculator.

[749] And then you have the three components.

[750] One is the objective predictor, which basically predicts your level of contentment.

[751] One is the model of the world.

[752] 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.

[753] Okay?

[754] Yeah.

[755] Cool, this policy, policy network, or something like that, right?

[756] Now, you need those three components to act autonomously, intelligently, and you can be stupid in three different ways.

[757] You can be stupid because your model of the world is wrong.

[758] You can be stupid because your objective is not aligned with what you actually want to achieve.

[759] Okay.

[760] In humans, that would be a psychopath.

[761] And then the third thing, the third way you can be stupid.

[762] 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.

[763] Okay.

[764] Some people who are in charge of big countries actually have all three that are wrong.

[765] All right.

[766] Which countries?

[767] I don't know.

[768] Okay.

[769] So if we think about this agent, if you think about the movie, Her, you've criticized the art project that is Sophia, the robot.

[770] And what that project essentially does is uses our natural inclination to anthropomorphize things that look like human and give them more.

[771] Do you think that could be used by AI systems, like in the movie Her?

[772] So do you think that body is needed to create a feeling of intelligence?

[773] 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 just about everything 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 not honestly publicly communicating trying to teach the public.

[774] But here's a tough question.

[775] Don't you think the same thing is scientists in industry and research are taking advantage of the same misunderstanding in the public when they create AI companies or publish stuff?

[776] Some companies, yes.

[777] I mean, there is no sense of, there's no desire to delude.

[778] There's no desire to kind of overclaim with something it's done, right?

[779] You publish a paper on AI that, you know, has this result on ImageNet, you know, it's pretty clear.

[780] I mean, it's not even interesting anymore.

[781] But, you know, I don't think there is that, I mean, the reviewers are generally not very forgiving of, you know, unsupported claims of this type.

[782] And, but there are certainly quite a few startups that have had a huge.

[783] huge amount of hype around this that I find extremely damaging and I've been calling it out when I've seen it.

[784] So, yeah, but to go back to your original question, like the necessity of embodiment, I think, I don't think embodiment is necessary.

[785] I think grounding is necessary.

[786] So I don't think we're going to get machines that I really understand language without some level of grounding in the real world.

[787] And it's not clear to me that language is a high enough bandwidth medium to communicate how the real world works.

[788] I think for this.

[789] Can you talk to what grounding means?

[790] So grounding means that, so there is this classic problem of common sense reasoning, you know, the winograd schema, right?

[791] And so I tell you the trophy doesn't fit in the suitcase because it's too big, or the trophy doesn't fit in the suitcase because it's too small.

[792] And the it in the first case refers to the trophy in the second case to the suitcase.

[793] And the reason you can figure this out is because you know what the trophy in the suitcase are.

[794] You know, one is supposed to fit in the other one, and you know the notion of size and the big object doesn't fit in a small object unless it's a tortoise, you know, things like that, right?

[795] So you have this knowledge of how the world works, of geometry and things like that.

[796] I don't believe you can learn everything about the world by just being told in language how the world works.

[797] I think you need some low -level perception of the world, be it visual touch, you know, whatever, but some higher bandwidth perception of the world.

[798] So by reading all the world's text, you still may not have enough information.

[799] That's right.

[800] There's a lot of things that just will never appear in text and that you can't really infer.

[801] So I think common sense will emerge from, you know, certainly a lot of language interaction, but also with watching videos or perhaps even interacting in virtual environments and possibly, you know, robot interacting in the real world.

[802] But I don't actually believe necessarily that this last one is absolutely necessary.

[803] But I think there's a need for some grounding.

[804] But the final product doesn't necessarily need to be embodied, you're saying.

[805] It just needs to have an awareness, a grounding.

[806] Right.

[807] But it needs to know how the word works to have, you know, to not be frustrated, frustrating to talk to.

[808] And you talked about emotions being important.

[809] That's a whole other topic.

[810] Well, so, you know, I talked about this, the basal ganglia as the, you know, the thing that calculates your level of mixed 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.

[811] That's the source of some emotion.

[812] So fear, for example, is an anticipation of bad things that can happen to you, right?

[813] You have this inkling that there is some chance that something really bad is going to happen to you and that creates fear.

[814] When you know for sure that something bad is going to happen to you, you kind of give up, right?

[815] It must not be there anymore.

[816] It's uncertainty that creates fear.

[817] So the punchline is, we're not going to have autonomous intelligence without emotions.

[818] Whatever the heck emotions are.

[819] So you mentioned very practical things of fear, but there's a lot of other mess around this.

[820] But there are kind of the results of, you know, drives.

[821] Yeah, there's deeper biological stuff going on.

[822] And I've talked to a few folks on this.

[823] There's fascinating stuff that ultimately connects to our brain.

[824] if we create an AGI system, sorry, human level intelligence system, and you get to ask her one question, what would that question be?

[825] You know, I think the first one will create will probably not be that smart.

[826] They'd be like a four -year -old.

[827] So you would have to ask her a question to know she's not that smart.

[828] Yeah.

[829] Well, what's a good question to ask?

[830] You know, to be impressed.

[831] What is the cause of wind?

[832] And if she answers, oh, it's because the leaves of the tree are moving and that creates wind, she's on to something.

[833] And if she says that's a stupid question, she's really on to something.

[834] No, and then you tell her, actually, you know, here is the real thing.

[835] And she says, oh, yeah, that makes sense.

[836] So questions that reveal the ability to do common sense reasoning about the physical world.

[837] Yeah, and you'll summon over to calls on your friends.

[838] Calls on your friends.

[839] Well, it was a huge honor.

[840] Congratulations on the Touring Award.

[841] Thank you so much for talking today.

[842] Thank you.

[843] Appreciate it.