Season 1 Ep. 3 Yann LeCun explains why Facebook would crumble without AI

[Music] today with us we have young lacoon jan is a french-born computer scientist often described as one of the godfathers of ai in deep learning he's chief ai scientist at facebook he's professor at the nyu center for data science and the koran institute of mathematical science and he's the winner of the 2018 touring award also known as the nobel prize of computer science welcome yam really glad to have you on the podcast here really glad to be here with you so as i talk about connecting with people um i'm kind of thinking back to what i thought for me at least was a pretty pivotal moment in the ai world we talked about 2012 the imagenet moment but then soon thereafter 2013 i remember it's it's europe's again the big conference we all go to and we go there to nerd out we talk about equations programs we write ideas that i would say until then pretty much nobody cared about except for the maybe a few thousand of us showing up at that conference and but in 2013 is this rumor and it's like oh mark zuckerberg is at the conference founder ceo facebook and it's like wait what is this what what is founder ceo of facebook doing at this conference he he's not a machine learning researcher this is machine learning researchers here and and then i see him on on the schedule he's on the schedule for the deep learning workshop q a with mark zuckerberg it says that one of the next days it was two days later and then i hear oh um apparently he's already here and last night he was hanging out with young lacoon and a few other people and i'm just what's going on here and and obviously you were hanging out with him and i'm kind of curious you must not have met him the first time at the conference there must have been some build up towards that and and what was that that whole transition from deep learning first of all being largely ignored then in 2012 being widely accepted as the promising path forward and everybody's starting to work on it after you've worked in it for so long also everybody's following your your path not just you but you and several others of course and but then all of a sudden industry gets interested how did how did you experience that industry right so funnily enough there are a few companies that actually were interested in deep learning and computational nets before 2012. uh one of them in particular was nec where i worked briefly uh in 2002 2003 and i worked on things like like phase detection pedestrian detection things like this and and when after i left uh another person kind of joined called caillou who basically picked up on a lot of those projects and sort of became a big fan of those conventional nets and he got an ec to actually commercialize some products that were based on accomplishments like vending machines in japan that would like recognize your age and gender so they could you know so this was in the uh that's flown under the radar largely i don't think yeah no that was like uh pretty much under the radar that happened around 2000 uh like the commercialization of those things around 2010 11 or so um and but they've been working on this like for for longer there was even a spin-off of nec that builds like you know surveillance like video surveillance systems based on you know pedestrian you know personal detection using commercial nets um but i was disconnected from this uh afterwards um uh but what happened in 2012 with the imagenet uh competition is that it attracted the attention of a lot of people engineers uh in uh in industry and there was a particular uh group of engineers at facebook who was working on computer vision and you know it was mostly exploratory because they didn't know what they could use computer vision for really at facebook um but they started experimenting with commercial nets and we're getting really good results for things like you know classifying images figuring out if they were like you know offensive material or not or doing face recognition stuff like that and and you know mark was really impressed by the result and at the same time he was going through uh kind of a thought you know this was a relatively short time after facebook uh you know became you know became a publicly traded company and uh you know it was you know turning 10 years old pretty soon uh and and you know google plus was on the way out so they were not like fighting for their survival anymore they were basically very well established uh in the social network uh thing so so you know mark and and shrek you know mike shepherd the the cto were thinking about the next 10 years like what what are going to be the big challenges for us over the next 10 years and ai sort of came on top basically that was like it was going to be a key technology for you know getting facebook to uh to move forward uh so what happened did they just called email you or yeah so i met uh you know some of those uh some of those guys working on this at cvpr and then i got a video call with uh with schrepp the cto of the company he wanted to kind of know uh first of all actually with the manager of that group uh sunivas and then uh with shrek uh in the summer roughly of 2013. we wanted to kind of get the line like what are the cool things and you know what are the things we should be watching you know looking for you know how do we get into that field like you know who should we hire you know stuff like that right they're really looking for advice more at the time yeah it was more advice that's right um and you know then then i had a call with with mark uh mike zuckerberg this was maybe oh before that before that uh i i get a i get an email a call from a former student of mine who was working at google at the time uh marcarelli around zetto who whom you you know uh and he tells me i'm being recruited by facebook and i tell him like why would you want to go to facebook they don't have a research lab you know you're a research scientist why why would you want to go there you say yeah you're right and then two weeks pass and he calls me again and he says um you know i've talked to mark zuckerberg twice and said why you talked to mexico he said yes and you know it's really interesting what they're trying to do i think i'm gonna i'm gonna take the job i'm gonna move from google to facebook and so that was a big of a shot because i knew my career really well and i knew he wouldn't go there unless there was some really definite you know will to kind of get into serious research and so we joined at the end of august i think of 2013 and then a month later i get a call from mark zuckerberg and you know he again a big like trap kind of wants to get the land so try to figure out like you know how can we move forward and then he asked me can you help us and say well you know i'm not going to work for facebook because i don't want to move to california and i want to remain a professor at nyu you know i don't see myself kind of leaving this so you know i can i can help you but i can't like be a facebook employee right i'd say okay um and then you know the company kind of goes into sort of several trials you know several attempts that kind of getting this jump started and then realize after a couple of months that basically the only way they the best way they have to start a real effort is to hire someone senior and kind of build a research lab from the ground up rather than say acquiring a startup or something like that right um and so i had to go to california for the pre-nips uh cfa workshop that you've participated in a few times um and michael radio was going there too so he knew i was going to california so he said like you know why don't you drop drop in at facebook and see what we're doing said sure and then the next thing is i got a call from uh mark's assistant saying why don't you come a day earlier and have dinner with mark sure um so so i get there um i have you know one-on-one dinner with mark at this house and he tells me what he wants to do and i find it really compelling and i realize that he knows a lot about machine learning he's he's read everything there is to read he's read some of my papers which completely floored me um and uh you know he at facebook he he's sitting you know every nobody has offices on facebook right it's all open desk right so there is a little block of six desks and there is you know mark zuckerberg uh mark surfer the cto and just across from mark is uh is marco rayo and right there that's right who he's like the first you know ai researcher hired at facebook and the reason is because mark wants to learn about ai so he keeps coming to to macro radio um and uh and at the end of the day he says okay you know we we just you know we just want to hire you somehow like can you build a ai lab for us wow and and i said well i only have two conditions which is that i don't quit my job at nyu and i don't i don't move from new york yeah and mark says sure say okay what do i sign it's not like that it took 24 hours but you know but like you know i was given a chance to create a research lab from scratch and i had some pretty i've thought about this you know before because i've worked in industry before you know at bell labs and att and yeah so curious what kind of mandate did you did you get because when when you're a professor you're in charge of everything you do effectively yeah it's your lab you do whatever you want but when you run a lab at a company well you're hired by the company and the university hires you and has expectations it's very different your research is completely free most of what you do is pretty free companies have expectations so what was the mandate well so there were demands there were very little demands from mark and shred they were pretty convinced that uh research lab would produce something useful in a reasonable amount of time even if the goals were very long long term and i i you know i really believed in this too because that's kind of the story of my career you know trying to build an intelligent machine and then kind of working your way back and whatever you come up with the first step you come up with turns out to have applications like congressional nets right um or back prop or whatever so um so i i kind of explained that i said like you know we're gonna have very ambitious goals very long term and they wanted that um but at the same time you know things will fall out of it pretty easily and we need to kind of organize this in such a way that this can be picked up and and and useful for the for the company but i think there's two things that are really important if you want to uh this to succeed one uh there has to be a lot of freedom on research scientists on the topics that they pick to work on it has to be essentially bottom-up okay you can do it top-down but top-down is more kind of directed research and this could be much more difficult to attract high-quality people in this kind of environment so if you want to attract the top people um it has to be completely bottom-up um and then the second thing is that the research has to be open okay and i struggled with the openness when i was at bell lab so bell labs you could publish just about anything but everything had to be patented and you could not open source code it was like it was before the time too but it was really difficult you have to go through layers of lawyers it was really difficult and i said you know i told market and shred um if you if you practice open research first of all it's gonna be much easier to attract good scientists because the the currency that measures the the quality of a scientist uh is his or her impact intellectual impact right and that goes through publications you know um that that's one reason for publishing there are other reasons the the you know published research is also more reliable it's also easier to convince people inside the company that the stuff you produce is good if it's published that's right and so the answer was astonishing but wonderful it was oh you don't have to worry about this at all like you know we come from the open source world schweppes previous job was uh you know tech leader at firefox he said like it's on our dna like you know we open source stuff all the time you know we we don't even patent anything except for defensive purpose and even that comes with open license to anyone who wants to use it so we never sue people for using our patents so i said oh that's awesome and i fast forward to just a few months ago and i have a quote from you in the next web that says you take ai out of facebook and basically the services crumble so in 2013 ai got started there now it can't work without it can you say a bit more about that a lot must have happened in those years turning ai from this nerdy research conference kind of activity into real world creation of value and what what happened right uh by the way that's true of google as well okay you take deep learning out of google today and it crumbles it's entirely built maybe yeah wow i mean there's lots of lots of companies that are built on this but these are the most you know the two most prominent ones um so the what happened was uh so the first applications were you know commercial nets essentially uh uh used for image classification and face recognition so face recognition is really useful because on facebook you know you can you know you are in a picture that are you know a friend of yours posts and you get recognized in it and you you get sent a message you know do you want to get you want to be tagged you happen to appear in a picture so you get this kind of automatic tagging and you people get alerted whether they are in a picture or not which is kind of good so that was probably one of the first deployed application and then another one was a kind of internal product called x-ray which uh was a system that would you know take an image extract like run a convolutional net extract the future vector out of it which is basically a long list of numbers and then just store it somewhere and and this could be used to tell if the picture contained uh you know various things like uh you know does it contain people how many people does it contain uh does it have a car in it you know maybe what's the brand of the car does it uh does it have like you know real things like weapons or you know things that are obviously kind of drug related or uh is it terry's propaganda is someone getting their head cut off you know kind of horrible things that you want to do it's like a lot of things that we normally don't see when we're on facebook but that behind the scenes that system is seeing and then maybe removing or something right and it took years before the system was you know progressively deployed and actually worked like to detect you know nudity and violence and uh you know neo-nazi propaganda all kinds of stuff right that uh uh are either and and you know child exploitation i mean there's all kinds of horrible things that people want to use facebook for that facebook has the interaction does the ai see and remove like is the bulk of stuff stuff that just needs to get removed or is it oh right now it's basically all of it like for image uh image stuff it's basically all of it is ai-based yeah all of it gets seen by the ai but then this is like 90 this garbage that needs to be removed or is it yeah it's dominated by things i need to be removed no no no no the i wouldn't say 90 of the material on facebook is is garbage it's not no no i mean it's it's still a relatively small proportion that is has to be taken down uh but most of it is being taken down automatically maybe because it's taken down so effectively it becomes less meaningful to post it because it doesn't go anywhere so right and people play games right so they they okay so now there's also kind of text understanding systems that try to figure out you know is this hate speech is this uh call to violence um you know things like that right that that or is it uh things that are illegal in certain countries right in uh most european countries uh you you know there is uh i mean his future is actually illegal right um it's not it's not just that facebook doesn't want it it's that the government doesn't want it okay i mean it's not just a government it's like you know uh so is it possible that if you took the ai out that effectively people who want to post those things would notice it's gone and start just flooding the place absolutely absolutely and there are social networks where that happens today like you know for 4chan hn you know uh those basically advertise themselves as not regulating content but they end up being a cesspool of uh right yeah yeah so you know and and then sort of became like so only in the last year or so with the progress in natural language processing basically due to uh self-supervised learning you know transformers and birds and things like that right that's you know some of which originated at facebook but you know some of some of those ideas came out of google and we exchanged ideas that's the beauty of open research which is that those ideas actually circulate really quickly um now you know 95 of his feature or so is taken down automatically um without anybody seeing it and this went from you know essentially zero percent two years early two years ago uh you know maybe 20 two years ago too you know it's a big difference as a yeah it gets more precise it becomes a lot easier to put to work right and then there is translation you know translation works really well so i mean facebook is about connecting people one of the best things you can do is erase the the linguistic barriers and you know the progress in recent progress in translation have been completely astonishing uh i mean that's probably one of the coolest works that has happened in ai community in particular at facebook uh over the last few years is the fact that you can you can train multilingual translation systems that can translate from 200 languages to 200 languages from you know any pair any direction and those are trained with very little data and you know they you don't have to go through english or anything the system can develop their internal representation of language that's sort of independent of the particular language you're using i i found that really astonishing there was some really amazing work also by uh guillermo and alexi conor who are actually resident psg students in paris on completely unsupervised translation so you you can represent as i said earlier you can represent say again what is a completely unsupervised translation okay so um here's the idea you you know i explain that there are techniques where you train a system to predict missing words right and with those techniques you can train a system to represent every word by a vector a list of numbers right um so you train that for a language let's say english okay and you do the same for another language uh let's say french these are bad examples because there is lots of parallel text in french and english but let's you know let's let's use that pair for the sake of uh simplicity so an entire language now is a cloud of vectors uh vectors are kind of like points if you want right so the all the words in the language is is a cloud of those points right that has particular shape and so english is a cloudy point and then french is another clutter point and now what you're going to try to do is try to kind of transform uh those two things so that the two cloudy points match the the shape basically match right so you kind of align them on top of each other and what emerges out of this is a mapping between english and french and so now you can build a translation system it's not a you know record-breaking one but it's a decent one and you never had to use any parallel text right of english and french you never had to have a rosetta stone or anything you just have english in one side french on the other side and just the structure of language is such that you know you can build this translator i found that absolutely a start machine that is amazing and as you go through these things it's well clearly a lot of ai is being deployed at facebook and but facebook can can be unique in the sense that it's the only company that is deploying ai many many companies out there are are deploying ai for their own needs and i'm kind of curious if in the last few years if you've seen a shift in terms of really how even because at facebook you're interacting with the ceo of the company which is very different from typically as a researcher you are in your own world you're not necessarily interacting with the ceo of the company right and at facebook are the places where have you have you seen ai play a much more strategic role in how companies just make product decisions or make investment decisions and and so forth and anything you can share about that oh yeah absolutely so um you know this is not just the the sort of default services that facebook offers you know like you know ranking the news feed uh doing ads translating you know filtering objectionable content all that stuff um there's also you know products that facebook physical products that facebook puts out like the portal system which is a video video conferencing system so the portal system has a smart camera uh which is basically a wide-angle high resolution camera and you know by software the camera can you know zoom on the on the speaker or uh you know there are multiple speakers you know kind of widens and it can pan and it's basically like a like a like a you know camera uh operator uh and that uses ai or it uses conversational nets for detecting people figuring out which way they're looking you know all that stuff right um then there is another set of products that are heavy users of making use of ai which is a virtual reality glasses so virtual reality new product lines based on ai that didn't exist that's right before so you know inside that tracking of uh like you know you wear goggles and the goggles has to figure out where you are in the world and track your position accurately so it can render the the world from the right point of view when you move your head uh and traditionally there would be like you know cameras infrared cameras that would be looking at you and try to kind of figure out the geometry of it but now the way it works is the other way around your camera is on the goggles you don't need anything external uh and the goggle looks out and tracks its own position and that's using kind of you know future detection and all that stuff that some of which is based on uh on deep learning it's kind of interesting what you're touching upon here is also a system that would not be purely in the digital world anymore something i'm personally very excited about this getting ai to work in the in the physical world with robots or other ways in the physical world this is physical world observations real real time physical world observations i'm curious if more generally about your thoughts on bringing ai into the physical world i know you've worked on some projects in the past for for not not car driving per se but uh vehicles going across complicated terrains based on vision and so forth so what do you think for example about self-driving cars right i actually have a project on driving car i'm working oh you do yeah it's a project with uh nvidia um i didn't know you're involved in the nvidia self-driving project that's exciting so you might have some very precise thoughts on self-driving cars it's an interesting story actually uh which i can tell you uh the when we were working on robotics uh this was a dapper project uh called lager which you probably heard about i think you were still a student with andrew when that happened and and andrew was involved in this separate project about robotics from from the same project manager uh and my group at nyu got associated with a small company in new jersey called netscale technologies which uh where the the the ceo and founder was the a former colleague of of of mine at bell labs who had worked on commercial nets when he was doing his phd in switzerland or similar so um so they were kind of you know netscale was kind of doing a lot of the uh kind of development for lager and you know my group at nyu was doing a lot of the upstream uh research and it was like it worked really well um so we had a couple projects like this and then you know the deep learning revolution happened things started to pick up and i get a call from elon musk saying you know we need to build our own sort of aisle driving system because the stuff we we you know we have from uh mobileye is not good enough and they don't want to work with us anymore so like how do i go about this they say well you should talk to my friend or similar and then a week later i get a call from jensen wong from nvidia who says exactly the same thing says like we'd like to like sell our hardware to you know for autonomous driving to come manufacturers but you know we can't sell them anything and unless there is a whole stack on it so we need to build our own autonomous driving uh stack uh how should i do about this should i well you should talk to my friend or smailer um and so was you know within a week also interviewed both at nvidia and uh and tesla and nvidia was just faster so to make him an offer so he joined nvidia and then uh you know he's based in new jersey in fact he's in the very same building that we used to work at bell labs which is now kind of another it's even in the same corridor it's kind of okay so so we're talking about self-driving cars and that's that's that's real world uh physical world and but there's actually also a lot of impact on what we do in the digital world on our real lives right and curious how you see the future of social media um as ai becomes ever more capable right how do you envision that will shape social media i mean the role of social media is to connect people with each other right raise cultural barriers uh you know get people to understand each other better i know there is some belief uh among a lot of people that social media is the cause of polarization political polarization for example or the dissemination of conspiracy theories but that's actually not true in the sense that first of all even if it were true it would be sort of you know seeing negative effects and ignoring all the positive effects of social media and there are many but uh but it's probably not true because the polarization is observed only in certain countries like the us and the uk for example but not in other countries that use facebook just just as much or as other social media just as much one could maybe distinguish between causing and which is one thing and another thing which is providing a fabric that if people are so inclined makes it easier for them right to polarize and it seems some of that is is happening right where it's it's a fabric where polarization seems easier to happen okay so it wasn't there there's a huge difference though between traditional media which are one-way and social media which is two-way so when some crazy uh you know we not extremist uh delivers a message on you know some extremist tv network or or radio uh which is one way there is no response there's no one to do there's no feedback there's nothing to to tell you know the the the audience uh this is this is bs or this is wrong this is factually wrong you know blah blah blah whereas in a social network uh you know there's some crazy post by a famous person if you're not you know connected you you'll just see that that thing but if you read the comments there's going to be a lot of discussions about how this is wrong and how this is you know incorrect factually there's going to be a little thing at the bottom that says that fact is actually disputed by our facts checker click on this if you want to know more um so the fact that you can respond the fact that the the the communication is two-way is is you know the essence of democracy in free speech which you don't have in traditional media so if i were to look for cause for polarization i would look in traditional media that you know disseminate conspiracy theories and you know social networks work both ways conspiracy theories can disseminate conspiracy theories but also other people can debunk debunk it just just as well one part that maybe stands out to me about social media is that it's in some sense one of the places where ai is the most active in our worlds right self-driving sure safer cars but most of us are on social media and it's it's driving a lot of our interactions and under the hood these interactions are ai's making decisions right because right i cannot see everything well i don't want to see everything because that's why facebook haven't built the news feed i would i would find it too boring to see everything somewhat people post and then it it helps decide what i'm interested in and it seems like those decisions i mean as an ai becomes better i mean there are all these stories about ai's building paper clips right and it's like okay if you're asking the guy to build paper clips is going to turn the whole world into a paper clip factory and in some sense it seems this is in in some way sometimes a microcosm of the same idea when you ask the ai to do something for example we want people to spend more time on facebook or maybe we want them to click on more ads or maybe some other objective whatever the objective is some objective gets said by somebody yeah and it's not exactly right and the ai is really good these things these things can get out of control right yes and somebody who's about your thoughts on on those things even just a simple thing of should we really be spending maximal time behind our computer right is that really the goal uh i probably spend you know more time on social networks than a lot of people because i use it as a platform essentially for disseminating a lot of uh you know communication about science and and but also about society and politics and all kinds of stuff right but uh but to respond to your question directly so uh there is this myth uh somehow that uh the the ranking algorithms that facebook use for your news feed reinforces the stuff that you like and therefore kind of creates information bubble this may have been true in 2014 it's not true anymore so there were some major like fundamental changes in the way facebook operates or the news feed ranking operates uh around 2017 that uh you know basically completely changes that that that thing and people also remember the old facebook and they say oh this is why um you know facebook is polarizing the world but it's not actually still very hard no and now you must still be setting some objective behind the scenes it's now a better chosen objective in some ways but fully optimizing objective is always risky i would say it's i mean it's not risky it's complicated right i mean uh it's not as complicated as designing a system from scratch designing a system designing an objective function that the system will will maximize is easier than designing the system itself right first of all and it used to be that the newsfeed ranking system was just designed there was no machine learning in it you know back in 2014 or something 13. then there was a little bit of machine learning in it and and now you know there's there's a lot of machine learning in it but there's a lot of things also the criteria that are being optimized have changed so it's not maximizing the time that you spend it's not maximizing uh you know engagement or whatever the the ultimate objective is sort of meaningful interaction so people have to be happy about the sp the time that they spend on facebook and must have the feeling that they're not wasting their time because if you don't try to maximize this objective people eventually are going to get bored and just do something else right they're not going to be on facebook anymore so so that's that's really the that's the ultimate criterion now this is not a criterion that's really easy to measure so you have to use proxy proxies for that and and that's that's where it becomes complicated we're very familiar with this in machine learning right we do this all the time we replace the actual objective function we want to optimize by your proxy a substitute that uh is easier to optimize but and we can be computed right yeah i actually do wonder maybe reinforcement learning could play a role here because oh totally and they could give feedback whether something was meaningful or not whereas they might not be able to construct the meaningfulness themselves so that there might be there might be some opportunities there oh absolutely no i mean there is a lot of you know reinforcement-like learning in there which is you know particularly there's also causal inference so there's a lot of like feedback loops right if you if you show something to someone you know they're more likely to interact with it than so you have exploration exploitation uh dilemmas and and things like this so yeah i mean it's uh so yeah i'm switching gears for a moment but on the same topic this is also the kind of thing where as a scientist we used to be in our own world but now the things you build actually affect things in the world and so to which extent do you essentially think about hey what is even the facebook algorithm what what is it even optimizing and are those things on your mind as an ai researcher at facebook too to really think about how it is being used beyond just building the next technology right so it's it's something that you know we we i am interested in as uh with the person as a personal interest but i'm i'm not actually working on this for the for the company and basically no one at fair is we develop basic machine learning and ai technology and image recognition and natural language understanding translation all that stuff right uh and and so more core machine learning techniques as well as tools and things like this and then it's in the hands of the people actually designing the products to figure out like what are the good objectives what are the you know things like that and it's you know somewhat removed from what you know what what we're doing you know of course we can interact with them we can provide advice we can you know give ideas about like new ways of uh you know assessing the reliability of a post for example you know um to figure out like how to rank it but this is really completely outside of our uh over you um but but of course you know we have a responsibility if we have uh kind of you know good ways of of helping with this to to actually do it but you're not what if you're not sure that the people who set the objectives are setting the right objectives do you want to send them that better more perfect optimizer well if it's something that you know about um so there's stories like that about like you know some kind of weird causal feedback loop in the you know like ad placement or something like this right and you know there might be an expert at fair who knows something about causal influence who realizes that might be a problem so that person might start you know talking to the people doing ad placement and and say like you know you may not realize it but you have an issue here that you know you the the thing you're measuring is not the right thing actually because of this causal loop and so if you were able to do like you know causal inference you would be able to kind of tell apart you know what really matters and what variable you can influence that you know causally influence the stuff you care about so i mean there's all kinds of things like this that uh um you know there are technologies that were developed at fair that are used you know just about everywhere in the company uh for things that are really important so it's a piece of technology which is actually completely open source called feis that means facebook ai similarity search that was developed uh at the at fair in paris uh mostly by uh matisse dus and ere jagu uh and and jeff johnson and it's basically you know it's not kind of the traditional type of ai that you think of like deep learning but it's it's basically a very very very fast uh nearest neighbor uh search system so here is a billion objects vectors uh and here is one object give me the 1000 among this billion that is closest to this to this one in i don't know a few microseconds this can actually be done this is incredibly useful because you have you know some extremist group posting a video or photo of some you know propaganda piece and then you have all kinds of sympathizers you know it's taken down but there's all kinds of sympathizers basically modifying this video and then reposting it and you have to detect approximate copies of it um there's blacklists there's whitelists you know those things and you have to do this for everything that people post this is you know hundreds of billions of things per day and you have to you know you know check against kind of you know blacklisted content and all that stuff this is the kind of stuff that that needs to be done i mean there's incredible technology behind it you know people don't realize like how much technology is behind a service like facebook the amount of ai happening behind the scenes is enormous and there is people in the lube and what i'm curious about is how do you see kind of longer term the synergy between ai and humans and maybe especially also how you might see something like a home robot or something like that emerge right how do humans and robots go together right so you know one of the one of the big kind of dream of both ai researchers and a lot of companies is an intelligent virtual assistant that you know can help you in your daily lives and you can you know ask any question and keeps you out of trouble and uh you know you know organizes your information and your your flow of information your email whatever means of communication you know it displays stuff in your augmented reality glasses you know wherever you go uh tells you the name of the people you're seeing whose name you don't remember you know i mean so all the stuff right i mean many of us have seen the this movie her right you've seen that movie spike jones movie it's actually not a bad depiction of what a truly intelligent virtual assistant you know might look like in the in the future you know not clear about the whole love story and everything but but that uh is is not necessarily about depiction and this is something that a lot of people have have in mind and so there's going to be a future where you know people are going to wear augmented reality glasses and they're going to have uh inter you know intelligent virtual assistants that help them in their daily lives and they're going to be based on ai and so before that happens we need ai that has common sense that can hold a dialogue that knows how the world works that you know kind of understands everything about us and there's gonna be they're gonna be our own personal assistant um just specific to us we're gonna train that assistant to work for us okay it's gonna be our or kind of specific to each of us individually so that's kind of you know in my opinion the future and then there's the physical uh embodiment of this which is um you know either you know domestic robots that you know do everything in our houses uh and and you know our cars that drive themselves and things like this and the question is whether the virtual personality of those virtual assistants is going to kind of migrate from a domestic robot to our you know ar glasses to a car or whether it's going to be a different personality every time uh that's that's kind of uh a question but you know we're very far from having the technology to be able to do those things this is not gonna happen you know it's gonna you know take at least a couple decades as i think about the future and what ai itself will look like in the future the the thing under the hood the thing that would be driving the robot brain effectively or the digital brain in our computer um recently jeff hinton of course you're great friends with and have known for a long time one of the other pioneers of deep learning he said i think deep learning can go all the way we we won't need a fundamentally new paradigm deep learning is not today's deep learning we'll need more breakthroughs within this realm of deep learning but we don't need a completely different thing right um what are your thoughts on that so my version of that statement which i think uh is pretty much what what jeff means as well is that the the basic idea of deep learning is that uh you you construct a machine by assembling parametrized functional block i'm using a technical term here but uh but bear with me for a minute and and then you you you train it for example uh you know end to end using using a what's called a gradient-based technique so a technique that can you know estimate in which direction you have to move all the adjustable parameters to get the answer you want you know whether their learning is supervised reinforcement or unsupervised self-supervised doesn't matter um maybe all of it so yes maybe all of it at the same time so uh this idea that you build a machine by assembling those blocks and you train it using gradient-based methods is here to stay okay so whatever solution ai is going to to be even 20 30 50 years from now is probably going to have this idea as a com as an essential component that doesn't mean that whatever we have today is sufficient like we don't know how to do self-supervised learning that deals with uncertainty for example this is something we need to solve and it's probably going to be part of the solution but uh but this the basic idea of deep learning is here to say i don't see it disappearing so this is this is my opinion but it's also i think my interpretation of what jeff says he says you know you know what whatever we do in the future uh whatever ic term we're going to build this this basic idea of deep learning is going to be part of part of the solution it's pretty wild definitely when i started my phd in 2002 i would not have imagined that ai researchers would get called up by receivers of the world's biggest companies and be asked any questions about anything and there you are being called up by elon musk uh arguably the most visible ceo in in the world these days and uh ceo of uh jensen wang of nvidia asking you the same question within a couple weeks time i think that's just amazing but given that you're also also working on this we know elon said there will be tesla fully self-driving taxis by the end of 2020 well it's 2021 that that right didn't happen um that was uh optimistic let's say right well none of us in the business believe that right well i don't even know if he believed it um but he did set it as a goal at the very least um i i don't know but i had so there was there was a a dinner organized by mark at this place with iran and and shrimp and a few other people uh to basically discuss with you know like what is you know that's at the time that we were saying we should regulate ai because it's going to kill us all right oh he had this period right and mark was kind of concerned about this and really wanted to get the idea of why he was thinking this and also kind of we were kind of hoping to kind of change his mind about it because we didn't think that was that was really uh the case and it was pretty clear to us that he was very optimistic about the speed of progress that he thought you know i think he bought the original um kind of uh you know marketing pitch or a sales pitch from you know deep minor various other companies that said oh we're going to have a gi within five years uh and that didn't happen yeah it didn't happen i mean we're now like six years after that and uh we're still you know nowhere close to human level ai or or even capital ai for that matter i don't think agi exists i i don't believe in the concept of agi because i think human intelligence is actually very specialized so so i don't think general intelligence is something we should be talking about agi is different from human level intelligence those are yes so agi stands for artificial general intelligence and a lot of people associate that with the notion that ai becomes as smart as humans and then because it has more compute power it will quickly exceed the capabilities of humans right brain is limited it's only so much energy you can do compute in your head and the ai can use the entire world of compute in principle right but you're saying agi artificial general intelligence and human level intelligence are very different things human intelligence is very specialized we we don't like to think of ourselves as being so specialized but we are i see so you think of us as very very specialized some people would say we're very flexible though compared to you're flexible we're flexible but we are specialized here is the way how we are specialized there's a lot of things that computers do much much better than humans and so it must be the case that humans are not very good at those things like let's take the example of chess and go playing or poker for that matter where you know machines you know now uh overtake humans in all of those games there was this idea somehow before uh you know good go players that humans were very close to an ideal player maybe you know to to god right uh like two or three stones handicap or something like this it turns out no humans are absolutely terrible that go machines are like so much better than humans and so you know we we thought of ourselves as as being almost kind of ideal and divine we just suck at go we are horrible um and so there are many things like this where humans really are bad but there's also this notion i would say that humans can adjust and i think that's what a lot of people think about when they think about artificial general intelligence this notion that humans can learn new skills like maybe yep learning is probably something maybe there's a board game you've never seen before and somebody explains it to you you now know how to play it you might not be the world's best player but you understand the rules and you you can take joy in it and and become better over time and enjoy that experience and and i think this flexibility of learning something new is quite special right yeah but it but it's it's limited it's limited flexibility so um so let me take an example if uh it's a very well known experiment that was you know carried out by a few people i think at mit where uh you you wear glasses that have a prism in them so that the uh image you receive in your eyes is flipped okay so you see the world upside down and you can adapt to this very quickly i mean in the space of a couple weeks it's now natural for you to wear those glasses and when you remove the glass all of a sudden now the world looks upside down okay you you you re-adapt to it much quicker but um so there is adaptation in that in that sense but let's imagine another experiment where uh through some process fiber optic or electronics um your the glasses you're wearing or the goggles uh do a random permutation of the pixels in your in the image okay a fixed permutation but but random complete scramble if you're looking at this completely scrambled okay that's right will your visual context be able to adapt to this and the answer is probably no i mean it might be able you might be able to recognize a few things you know very low resolution and things like this but it's going to be a complete mess and the organization of your visual cortex is such that there is no way you can catch up so okay so it's an extreme example but it shows that that's an example of specialization that yeah our brain as as you alluded to earlier when we're children we actually learn a lot and maybe a lot of that learning is in our brain consolidating the sometimes geometric structure of the world and so if you now scramble an image it it we can't make sense of it anymore because we've been trained to expect a certain structure this is not training uh you know even if you do this with two children which would be a horrible thing to do um you identical um i don't think children can it might evolutionarily be in our brain that we have certain structures the reason for this is is is the reason why you know i started working with congregational nets right so the reason why so functional net is a particular type of neural network whose architecture is inspired by that of the visual cortex and the characteristic of it is that a neuron in the early layers of a commercial net is influenced by a small area in the in the image this is called local local connections or local receptive fields this works if nearby pixels are correlated so if the pixels organized as in a regular image there is an advantage to having this local processing but if you scramble the pixels in a random permutation you break this correlation and your local kitchen detectors basically are useless so that's you know it's the it's the hardwired architecture of visual context that will prevent you from being able to vision with scrambled pixels when you think about neural networks most people will easily think about the brain because don't we have neurons in our brains and our brain also has the general capability of reasoning about many many things so maybe you think yes neural networks should be should be good can you say a little bit about how the neural networks network that you are building other people are building relate to our brain so they relate to our brain in the way that you know airplanes relate to birds so birds and airplanes use the same basic principle for flight they both have wings they use very very different ways of you know propelling themselves to air um but they use the same principle to generate lift from you know by propping themselves to through air so it's a bit the same thing right the the general kind of idea between neural nets and and and the brain is kind of the same this idea of lots of very simple elements that are connected with each other and you know whatever behavior comes out of it is kind of an emergent behavior that comes from the complexity of the of the network and the way those systems learn is by changing the efficacy of the connections between those those nodes right so that's the general idea now if you dig a little deeper than this the analogy stops so the same way uh airplanes don't fly their wings and don't have muscles uh you know those neural nets uh don't work at all like the neurons in the brain in fact some people say we shouldn't call them neurons at all and the the type of learning that we are able to reproduce in machines is very different from the type of learning that we are observing in uh in biology so the the type of running i described before where you you show an image to the machine and you tell you the answer and then you the machine address itself to get the answer closer to what you want that's called supervised learning and it works really well if you have lots of data so you want to translate one language into another you collect millions of sentences that have been translated from one language to another and you just train the machine to kind of do that and speech recognition is the same you can have tens of thousands of hours of translated speed of transcribed speech image recognition you know you can have millions of images but it only works for things for which it's you know feasible to collect that much data and there's a lot of situations where it's not visible like you know you want to identify uh tumors in certain medical images and you just don't have that many of them because perhaps the disease is rare and it's very different every time it shows up so there are a lot of situations for which is just not feasible but also you refer to this notion of supervised learning what does that mean exactly to be supervised learning so supervised means that you show the system an input let's say an image and you tell it what the correct answer is and it adjusts itself so that it produces the correct answer then there is there are other the two other types of learning one that has become quite popular in the last few years and and you peter has you know have been working on this since you were doing a phd called reinforcement learnings and i'm going to teach you about that but the difference is that you don't tell the machine what the correct answer is you only tell it whether the answer you produce was good or bad and so so if the set of answers is really large the machine has to basically kind of try many things before it figures out the correct answer this is very uh very good for training a machine to to play games for example uh but it's very inefficient in the sense that it takes a very very long time for a machine to learn anything useful like training a robot to you know grab something or to drive a car you know it would take too long if you use the at least the the the current brand of reinforcement running or until it has to stumble upon the solution before it learns about it being the solution has to discover it's on its own so it's a very different kind of learning it has to do with millions and millions of trials and and also the what he's going to see next depends on the decision it just took so that makes things even more complicated and then there is a third type of learning which is the in my opinion the primary type of learning that happens in the in the brain in animals and humans and i call it self-supervised learning people call it lots of different ways but it's basically a type of learning where the the the learning system uh whether it's a machine or a biology uh basically learns to represent the the input without learning a particular task uh it just learns the general organization of the world if you want and and if if the machine can train yourself to do this then um then learning a particular task can build on top of those representations learned by the the system just by observation so uh you know babies human babies and this is true for most animals um you know learn how the world works in the first few months of life by essentially by observation and most of what we learned in our life we learned in the first few months of our lives we we get the impression that all of we you know everything we learn we learn in school and to our parents and you know at university by reading books but no actually the most of the basic things that we learn about the world we we learn in the first you know maybe nine months of life so one more relaxed version of learning in those first few months is my impression if i can remember anything of it it feels learning in school is much harder work any way we can extend this kind of early childhood learning process a bit longer and keep learning at that pace well yeah i mean at that at that time your your brain is uh you know very malleable and you know you're you know a big chunk of your your energy uh is devoted to you know everything is new in the world right when you're a baby everything is surprising uh you can only take a few hours of this a day and you have to sleep on it uh you can reorganize your brain you know every time because it's completely overwhelming uh but but you know perhaps one principle that is uh is used by by biology to to learn how the world works is through prediction so there's a lot of you know internal structure about the world that we learn by uh training ourselves to predict lots of things that we can't directly observe like for you know just now you you can't see the back of my head but you have a pretty good idea what it looks like because you've seen a lot of people and you know uh etc uh you know you can't predict if you know if i'm going to move my my head to the right or to the left like you know in a second but you can predict that my head is not going to jump so suddenly uh 30 centimeters to the to the left um or it's not going to be upside down just just all of a sudden because you know the constraints of the physical world uh you know that if i if i hold an object and i let it go it's going to fall right babies learn this around the age of nine months they learn about the fact that uh unsupported objects fall because of gravity around the age of nine months it takes them all this time to figure out the physics if you want of interest to learn objects fall wow okay right so we're learning fast but maybe maybe not that fast either well but but once once you've done this you know uh you you turn 16 or even earlier than that and you now need to uh learn to ride a bike and you can basically run that in half an hour uh you know with just a few trials without hurting yourself too much uh you want to launch a drive you can drive pretty well with about 15 hours of training for most people without ever crashing uh we can't do this today if we could solve that problem today then we would have level five autonomous driving we would have you know cars driving themselves uh without crashing anywhere but we don't we really don't know how to so so this reminds me actually of something you're you're very well aware of something called the cake of course um and it reminds me of see this was nureps 2016. you gave the opening keynote kicking off the conference this is for context this is the main machine learning conference the biggest one pretty much everybody who works in machine machine learning goes there to share their work and see what other people have been up to and exchange ideas get new ideas and you kicked off the conference with the invited keynote where you introduced this concept of the cake and as as i was listening i was sitting in your audience at the time i was listening to it it occurred to me that in some sense even though my own path which has been more robotics driven and your path has been more pure machine learning driven though we both touched on both sides quite a bit that in some sense what you were presenting was the future of robotics maybe not the future that we're going to get right away tomorrow but the kind of longer term like what the most capable robots of the future might might look like on the inside how their brain gets built so can you maybe say a little bit about this this little cake idea and and how it could play into robotics right so it's a metaphor that i used uh because i was seeing a lot of excitement for deep learning in the community and deep learning at the time was you know 95 supervised learning and there was a lot of mounting excitement and reinforcement learning and uh and and there were some people who said who thought maybe perhaps natively uh or perhaps some good idea that you know by scaling up supervised learning collecting more data having bigger machines or by scaling up reinforcement learning uh we were gonna you know be able to create intelligent machines and i was really convinced that that could not possibly work that the the way forward to really build intelligent machine was to get machines to learn how the wall works and perhaps acquire some sort of common sense and it could not be done through supervised learning and reinforcement running alone because the amount of feedback information you give the machine in supervised learning and reinforcement planning is very very weak it's very small right you know you you when you train supervised supervised machine to recognize images you know you tell it it's a car it's an airplane it's a person it's a table it's a chair you give it just a few bits of information at every sample so the amount of stuff that the system learns is limited by uh you know how much information is in the label which is fed by humans and it's very limited uh in refreshment planning it's even worse right the machine only gets a single number every time you try something you you're doing good right so the so the amount of trials the machine has to do is just enormous uh so if you want you know learning to be efficient the machine has to basically figure out how the world works by itself and so my analogy was if intelligence is a is a cake the reinforcement clowning has so little information you give to the machine that that's going to be just the cherry on the cake and reinforcement clowning is the is the icing on the cake but the bulk of the cake is self-supervisioning or whatever it is that humans and animals are doing the joke i usually say is that you know a house cat has more common sense than the most intelligent of all our ai systems how can we do self-supervised learning and make our robots more naturally smart that's the question you tell me i see it's not solved yet yeah no it's completely unsolved i mean there is there's been very interesting advances in that direction over the last five years and there's been some success in some domain so for example the most successful natural language understanding systems today the most successful translation systems so basically text you know text processing text understanding are very heavily based on self-supervised learning and it's self-representing of the type that i was telling you about earlier where you take a segment of text extract it from the book or whatever or wikipedia article a website you remove some of the words and you train a gigantic neural net to predict the words that are missing in the process of doing so the system learns to represent language and it learns automatically to basically represent a combination of syntax and semantics it it has to understand like if i if i tell you a sentence you know the blank changes the mouse in the kitchen you can you can tell what blank is and that's because you you know how the world works right you know that uh you know cats change mice and and you know cats are in inside you know our houses and stuff like that right um if i say the you know the blank uh uh chases the gazelle in the savannah you can probably also tell what it is at least within you know within some region so so by by training a machine to predict missing information and telling it after a while well this this is the word that you know was supposed to be there the machine basically learns to work with supervised learning then aren't you supervising it by by having the text that has the completion for the blank yes so it's it's basically the the basic algorithm you use to train the system is a supervised learning algorithm because you're telling the system the correct answer but the reason why it's called self-supervised is that this information does not come from someone who has labeled the data manually it comes from the input itself right so you you take a you take an input you corrupt it and then you tell the system recover the original input this by the way applies to not just text but you can apply this to speech and images it doesn't work as well in images but it works astonishingly well in natural language processing to the extent that it completely revolutionized the way people do natural language understanding and translation about two years ago every system now that you use um you know whenever you go to google or facebook the the translation the system that ranks the the the items that you see either search items or search results or or your news feed this is all done by this you know natural language understanding system that basically uh do this you can ask a question to google google will answer right and and that's based on all of this is based on this uh this kind of technique so there's been this enormous success of self-supervised learning in the context of natural language understanding and it has not yet completely translated in other domains like speech recognition or image recognition although why do you think that it is why do you think it works better in language processing so far and not as wider than the other domains okay i'm going to say something that's going to be very controversial but it's because language is simpler language is is you know we we like to think as as humans you know we like to think of of ourselves as humans as the only species that has you know complex language and and it's you know intimately linked with intelligence and and all that stuff right but language is basically an ep phenomenon of of intelligence it's a lot of animals who don't have language that are almost as intelligent as we are like uh orangutans orangutans don't have it don't have language because they are solitary animals they're not social animals they're almost as as smart as we are i mean on the scale of of smartness right they're very close to us um so uh uh you know language in this conflict in particular context of prediction language is simpler because the prediction you're making is you're predicting a missing word and you can never predict the exact word right so the example i said i said you know the the blank chases the mouse in the in the kitchen is basically only one possibility it's going to be a cat uh you know maybe kitten or something uh but if i say the blank chases the blank in the blank there's many many possibilities of what animal where you know which predator and prey right and so you can't so the system cannot make an exact prediction but what it can do is that it can make a probabilistic prediction so it gives you essentially a score for every possible words in the dictionary there's maybe a hundred thousand word in the english dictionary uh you have this giant list of numbers that indicate kind of a score for every possible word what is the chance that this word appears here and so we have an easy way of representing the uncertainty in the prediction by a big list of numbers essentially but what if i do something else if what if i train the system where i show it a little segment of video and i ask you tell me what's happening next like draw the frames in the videos that are going to happen next as a consequence you know as that can follow this there's an infinite number of possibilities and and you know an image is a very complex object by itself but the sequence of of uh or video frames is incredibly complicated and we don't know how to represent the uncertainty in the prediction in that context we cannot we can have a big list of numbers that gives us a score for every possible image that follow a particular frame so the complexity and i'm getting into sort of a technical issue here but this is really what we're facing uh we we do not know how to represent uncertainty in sort of complex high dimensional uh domains like like images video uh audio to some extent so we have to invent other techniques and there are a few that are really promising it also sounds like we're putting out there then is a notion that if somebody can crack this if somebody can actually represent possible futures in the video sense meaning not in their own head but somehow train a neural network to be able to do that that neural network would maybe have some notion of common sense at least some kind of notion of common sense and and be able to learn other things much more quickly absolutely because that system would be able to just watch the spectacle of the world and then you know train itself to predict what's going to happen next and then understand that you know if i if an object is not supported it's going to fall but here is the complexity if i put this this pen on my on my hand and i let it go you can you can predict that the pen is going to fall right but you can't predict in which direction and every time i repeat the experiment it falls in a different direction so so that's the complexity of prediction which is that you know you don't want the machine to predict a particular image you want it to predict you know the pen is going to fall but i can't tell you in which direction how do you represent this uh that's what we don't know how much we have humans represent that though because when i see the pen on your hand and it falls i'm not too surprised even though i can't predict which direction's gonna fall i'm i'm not surprised that it falls i'm not surprised that it is in some direction and yeah that was a plausible direction why not that's right in fact that's exactly how we know that babies uh you know know something about the world you show them a scenario that breaks one of those laws of physics and you observe how surprised they are um you basically you you measure how long they look at the the thing right so you show a little scenario where you know an object kind of appears to be floating in the air because the support of it is not apparent you know we see them somehow and a six-month baby will kind of look at it for a bit and like not really pay attention uh a 10-month baby will go we go like this right and and like stay on it for like a minute because you know her model of the world is is being violated right she she knows that an object is supposed to fall if it's not supported and here is one that flows in the air uh so you know that's that's surprises is how we measure whether our model of the world is being violated [Music] we are dropping new interviews every week so subscribe to the robot brains on whichever platform you listen to your podcasts sayana i want to get back to some of the kind of future of ai research maybe a little later in our chat here what what else maybe hoping we could also touch upon a little bit is if you're willing to share is how how do you how do you start out growing up in france and now you are here in the us i mean you moved a while back like how do you how does somebody just is born and then later they are a ai scientist what was that trajectory for you when did you know that's what you want to do okay it's it's it's it's simultaneously simple and complicated okay so you know i grew up in the suburbs of paris in a you know middle class family my dad was a engineer in the aerospace industry uh my mom was a homemaker uh but my dad was uh like a really an engineer at heart and uh and he taught me he taught me everything basically but you know when i was a kid with him he would you know build moderate airplanes he built his own radio control system actually um you know he taught himself an electronics a long time ago before a long time ago this is in the in the 60s nobody knew how to do these things at the time uh i mean he had built modern airplanes you know before that but radio controlled modern airplanes you know you couldn't just go out and buy a radio control system you basically had to build it yourself in the 60s so he did that he taught himself electronics he was a mechanical engineer told himself electronics and you know he you know and i was like watching this when i was a kid and i was kind of you know amazed i also always had kind of this sort of you know engineering kind of you know virus basically infecting the family my my brother also kind of went through the same process he's actually uh uh he's a research engineer at google in paris um he doesn't work on machine learning he works on a combinatorial optimization optimization guide but um so uh you know i got kind of exposed to science and engineering this way and i was fascinated even when i was a kid by uh you know the appearance of intelligence in humans i was you know interested in paleontology for that and how did it happen that you know humans evolved and became intelligent like i was fascinated by that question and then you know what is intelligence and and things like that right so um i was fascinated by physics you know by astronomy about all kinds of stuff right so then um you know i was okay in high school i was good at very good at physics i was okay at math i wasn't particularly good i was good you know and so the you know there's kind of a strange type of uh educational system in france where if you're really good at math and science you go into one track you go into those like two years of preparatory school where you just do math and physics and then you take a a competitive basically a national competition to get into the best engineering schools from students programs it's amazing right and i didn't go through this because i really wanted to do engineering from day one uh and and and science and and my math teacher in high school convinced me that i wasn't good enough at math to survive that i went to an engineering school that uh you know basically goes right after high school right and it's a five year program you do a lot of math and physics and everything and i did like way more math and physics that you know the average student in that school uh i learned electrical engineering and various other things but while doing this i was interested in this mystery of intelligence you know artificial night and i stumbled on a book which was the transcript of a debate between uh noam chomsky the linguist and jean-pierre who is the swiss developmental psychologist um who studied how children learn so i was you know fascinated by this you know i read this book and it was kind of a debate where piaget and chomsky kind of brought people to argue for their side and on the side of of piaget was this guy seymour pepper who was a math mathematician from mit and his um his talk which was transcribed was about the perceptron which was one of the early learning machines from the 50s and he says you know the perceptron is a very simple learning machine but it can learn surprisingly complex concepts when you train it even though it's so simple and so i it's the first time i read about a machine that could learn i was i was you know transfixed essentially i said like this is what i want to do this is like this is the answer you know intelligence is so complex we're never going to be able to engineer it it's going to have to build itself by learning and in fact i discovered you know decades later that turing had the same argument he said you know if you want to build an italian machine you'd be better off trying to emulate the the mind of a child so that you know this child could learn the way a human learns um so i always thought learning was going to be part of intelligence and then when i learned that people had worked on this you know i decided to basically spend all my spare time kind of studying this so i went to the i spent you know days at a time in the libraries there was no internet back then right we're talking 1980. i had to you know drive to versailles where indriya had this like wonderful library which i gained access to they were nice enough to let me in and i would spend you know the entire wednesday afternoon because i didn't have class then um kind of looking through the old literature and i discovered is that the entire literature on this topic stopped uh in the early in the late 60s around 1969 1970. you couldn't find any paper from the us or europe that talked about uh perceptron you know neural net like machine learning you know after 1969 roughly and i discovered the reason for this was or one reason for this was a book whose title was perceptron by two authors marvin whiskey and seymour pepper so he was simple pepper chanting the singing the praise of the perceptron but he actually contributed to killing it how do you kill neural networks with a book well it's a theoretical book that says you know here are the limits of those models uh you can't you know here is what you can do here so you cannot do with those models without you know them being ridiculously large or or or whatever and it's a very interesting book that i i read multiple times right at that time they were not they were not wrong it was just that the results were over interpreted uh in the sense that some of the problems that that they say perceptrons cannot solve or can only solve with a ridiculous amount of resources first of all our problems that may not be that interesting anyway uh that you may want to do there are problems that are essentially sequential so you might want to use another type of machine to to solve them and second of all the main limitation was the fact that the perceptron basically has a single layer of trainable weights right so you can think of the perceptron basically as a single neuron with you know a bunch of adjustable inputs uh and everything else before that is fixed and and so that's that's a very dire limitation and and what you know the neural nets more the recent brand of neural nets you know from from the late 80s uh and back propagation allowed to do was to train neural nets with multiple layers of adjustable weights that's by the way why we call this deep learning it's called deep learning for the simple fact that those neural nets are composed of multiple layers right to differentiate them from more classical machine learning where you basically only have one layer that's trainable but you didn't achieve that right away right you're sitting in these libraries and in versailles apparently and reading reading these books that that's still a long path to making your own that's actually work and where we are today what what happened next right so i was like uh you know sophomores or maybe junior um i mean it's it's five-year program in france right so i was in third year or something like that and uh the school i was at called e-s-i-e you could do independent studies so i went to see a math professor and they said like you know um i'd like to work on this like you know i want to like do a project and say yeah this you know that sounds like i think it could be fun so i did a couple of those and i was really thankful that the this professor kind of uh helped me i had access to the the school's computers which at the time were kind of rather powerful for for for undergraduate students um and and so i started experimenting with sort of various uh you know various learning algorithms and i you know quickly discovered that really what needed to be discovered was a learning learning method to train those material neural nets this is what people didn't find in the 60s this is what caused the failure if you want of the entire field and that's what needed to be unlocked essentially and i look for all kinds of ideas there the people who are working on this in various places in the eastern europe in japan etc and i kind of stumbled on this idea that perhaps if you were able to sort of propagate signals backwards to tell every neuron like here is what i need from you to get the correct answer that could work and it came up with some sort of really intuitive world to do this that we would now call target prop and then i talked to one of my friends who was doing a phd in optimal control and he was working on planning and he told me about this technique called the i joined state method i don't know if i mean i'm sure you know about this because you are a control guy so this is an algorithm that goes back to the 60s that it was used for example by nasa to compute trajectories for rockets things like that and it's based on the idea that if you want to shoot a rocket and you want it to meet the space station uh you you start by kind of shooting it and it's going to miss the space station so you compute kind of a distance between those two things and you kind of adjust the the controls the nozzle the thrust you know the direction at every time step so that this it gets closer and you do this by basically propagating a signal backward in time this is all done on the computer right um and uh uh this is the this is the back propagation this is this the same idea it's the same mathematics that we're using right yeah and so that's amazing so you're working on this back in france how do you end up in the u.s okay so i was i i finished my engineering degree my specialty was uh site design by the way uh cheap design but as i told you it took a lot of physics and and also actually optimal control stuff and and i stumbled on another book which was talked which which was about uh self-organization and something called automata network and i discovered that there is a small independent lab in paris uh composed of uh academics who had positions in universities but decided to kind of get together in this like little informal organization uh called the laboratory so this means the laboratory for network dynamics and they were kind of studying those emerging properties of large networks of uh of elementary kind of simple objects you know like cellular automata and things like that and i just call them up cold call um say like you know i'm i'm interested in working on neural nets and stuff like you know can i just hang out um and one of them says oh you should sign up for a graduate program uh right now because the deadline is like you know three days from now so or something like that or you're too late but like i can get you in any way or something right so so i registered to this graduate program which was sort of a weird thing because in france you have to do like a like a free phd before you do the phd so this was the pre-phd um and that's how i learned to to do research because you know otherwise i was kind of uh isolated but no one in that lab was working on neural nets they were working on a sort of other type of things uh i had funding from my uh like a scholarship from my engineering school so i didn't need any uh any support but i needed someone to be my official advisor to at least sign the paper so one of the members of this lab mauricion i come to him and said like you know you are the one food professor who can actually take me as a student would you would you take me he says like i have no idea what you're working on i can't help you at all with anything technical but you seem smart enough and i'll sign the papers um you know a bit of uh i was kind of you know out in the out in the cold a little bit kind of how do you got interest more general research world then right so one thing that uh you know from the first month uh i sort of started hanging out with uh people there so there was a a scientist called francoise sulia for german and she you know she she kind of started to get interested in your net pretty early on but she connected me with uh uh the the wider community that was already working on this in particular she gave me a preprint of two things one which was the paper on hubfield networks and the second one was the paper from 1983 by internationally on boston machines and i look at this paper and you know back then you know those papers were like published in obscure conferences i think the also machine paper was you know published at either cvpr or triple ai or one of those things but you know you couldn't get those unless you went to the conference and brought the you know the proceedings back right so um so they got a pre-print somehow you know they they they ask uh intern or whatever and they got they got a photocopy of it there's no email either okay so so it was all by your physical mail um and i look at this paper and they talk encrypted encrypted way they talk about essentially the multi-layer training problem okay so they the possible machine algorithm was kind of the first credible solution to training neural nets that had so-called hidden units with the units that are neither inputs nor outputs but somewhere in the middle and i thought like this is exactly what the kind of stuff you know i'm interested in i absolutely need to talk to these people uh cenosky and hinton like they are my heroes now okay and they are in the u.s and i'm in france and uh how do i do that um so it's all it so happened that the francoise and a few other people organized the workshop in 1985 uh in france where they invited a lot of uh you know physicists who are really interested in neural nets at the time there was a lot of theoretical connections with uh solid state physics or or condensed matter physics like spin glasses and stuff because of fog field and so they invited a lot of different people in particular they invited teresanowski so unfortunately terry wasn't here for my talk i was talking about this uh kind of target prop algorithm in completely broken english i was absolutely terrified when i gave my talk because my english was so bad and also i knew people would not understand what i was saying um you know both for the language because of the language and because of what i was talking about um so he wasn't there for my talk but he he came give a talk about muslim machines and then i cornered him one of the afternoon and tried to explain him what i was working on he was working on net talk at the time which was sort of one of the first big demonstration of what backprop could do but backward was not published yet uh nobody had heard of it except him because he was friend with with jeff hinton uh so he went back to the us and he he told jeff uh who was working on backprop also preparing the paper he told him there's a kid in france who was working on the same stuff we're doing um and uh and then a few months later uh another conference takes place in france in summer 1985 where jeff intern now is a keynote speaker where he talks about possible machines so you know by that time neural nets were starting to get like a little hot and muscle machines were kind of seen as sort of a big thing so uh so he gives his keynote and then he's surrounded by 50 people and i can't get anywhere close to him sounds like you today except 50 would be uh 200 i mean you know jeff was pretty young at the time he was like you know he was not 40 right he was like 37 or something so um so anyway he uh he turned to the organizer i was kind of far away right and he tells the organizer do you know do you know a guy called jan lacoon and i get i go i'm here basically he had read my paper in the proceedings which was in you know french and kind of figured out this was about multi-layer nets and backward like things uh and so we said like you know let's get let's get lunch together tomorrow and uh and we basically you know then he told me he was working on backprop and i i told him well that's what i'm working on too and we were kind of basically completing each other's sentences so this clicked like you know immediately um and and so he he said uh like i'm organizing a summer school next summer 1986 at cmu he was cmu at the time said like i'll invite you there um and that's where i met the entire community right so this was the first connection in summer school you know michael jordan was there and you know terry and j mcclelland and i mean a lot of senior people now that you see in the machine learning community actually were were there at that uh at that summer school because now all these people are your longtime friends but then you're a young researcher and you show up and all these people are names you've seen but you've never met them before uh how was it i mean i can only imagine it's a bit intimidating well it is yeah but um particularly intimidating because my english was so bad and so in fact my english was so bad this is one of the reasons i connected with mike mike jordan uh because he his french is very good so he's french was completely fluent and he's kind of a bit of a francophile so um we connected you know immediately and so became fast friends i think you got a good part of it you got to coming to cmu and meeting jeff intern and everybody there which i think got uh yeah i mean i met jeff into you know earlier earlier in france but uh what happened is i in 1985 in this first this first meeting in uh in the french alps that was organized by the member of this lab that i belong to where all the physicists were and where i met teresanovsky i gave my first talk and there was a guy in the audience a young guy um he looked like a cowboy from the arizona he he had like a uh i mean he was dressed like you know he had like big sideburns and it was very strange and and this this guy was very young but he he was asking very nasty questions to absolutely everyone like it was you know this very famous person giving a giving a talk about their major research and this guy would raise his hand and ask a question that would basically destroy the entire talk or or put it in question uh and so i turn you know to my my friend and say oh this guy these guys yeah i mean they are from bell labs you know that's that's what they do about labs right you know whenever you you talk about something you know someone else someone has done it about last 10 years earlier or it doesn't work um and so i was kind of terrified by this guy and i give my talk in broken english and his complete silence in the room uh because nobody understands what i was saying except this guy who raised his hand and i i literally liquefied in place because i was you know terrified of him and he said something nice about my about my talk he said like you know this is really cool blah blah blah not asking an asking question so i resolidified and two years later they offered me a job so this guy was this guy's name was john danker and he was kind of the person i i worked most with at the labs and the person next to him was larry jackal who was the the the department head who hired me was my boss he ended up being my bus in closing i want to ask you one last question that's a bit different john which is um one thing i noticed about you is that you um you engage quite a bit in in political discussions you know look at facebook twitter and so forth and it's kind of intriguing to me because i think a lot of us scientists clearly not including you but a lot of us kind of love science partially because there is a way to determine the correct answer and there's a way to determine this this is now correct we figured it out right and in politics everything's so well there are definitely clearly wrong things but yes it's much harder to to identify the clearly one right thing that everybody in the world should agree this is the the way forward and that that seems to never happen right so kind of wonder where that comes where do you see that come from in your history and also how important is it uh very important okay so the the thing that gets me riled up in politics uh is not you know necessarily ideology you know classical political ideology but it's more that you know seeing people suffer for no reason for people for reasons that are easily avoidable and and seeing that they suffer because there are decisions that are made that are completely irrational or that are based on false information or biased information or false facts or bad models of the world and or self-interest also sometimes but uh which is clearly the case in a lot of politicians but and so you know i'm uh my philosophy is that you know i'm a i'm an atheist i'm a rationalist and i'm a humanist okay i'm actually more than a humanist because i think it's not just humans who are who deserve uh uh you know happiness but also you know animals and everything but um i'm not a vegetarian by the way but um but but because i'm a rationalist you know as a scientist we have to have at least one corner of our minds that is is rational it's completely rational and when i see political decisions being made for for example reasons that are uh based on you know a religious myth from you know sheep herders that you know lived in the middle east 2000 years ago that end up hurting people in really bad ways i get riled up i i don't find that you know that's completely against my my humanism right i i want to kind of we were talking about objective functions right i want to sort of you know maximize uh the long-term expected value of human welfare and minimize the long-term expected value of human suffering and so you know that's kind of a rational objective function i mean it's an objective function that you may agree or not with and then what is the rational way of optimizing it and it has to be based on your model of the world like you know what is what is going to be the consequence of taking that decision on the long-term uh well well-being of humanity it seems if you had a really good ai system that could simulate the future depending on decisions you make you could actually build a tool that's very powerful to improve the whole human condition how everybody's living that's right so i mean this is you know trying attempting to automate this completely would be like the ultimate technocratic uh uh you know former government which i don't think i i i subscribe to but um but i think there is something in this sort of model of of intelligence you know for an intelligent agent to act whether it's a human or an animal or a machine there has to be three components three main components or four main components one component is the objective that you want to optimize uh all humans have this hard-wired in our brain there's the you know something at the base of our brain called basal ganglia and it basically computes how uncomfortable or comfortable we are this is the thing that tells us you're hungry you're hurt you know uh you know someone is pinching you you're burning or whatever right um the second component is what in our business we call a critic so it's something that tries to predict unless you know prefrontal cortex it's something that tries to predict uh in the future i'm not going to be comfortable or uncomfortable what is the future expected value of this objective that you know my brain computes for me um then a third component is our model of the world so it's something that predicts what is the world going to do uh how is the world going to evolve uh in the future and more importantly how is it going to evolve if i take this action so can i predict the effect on the world of the action i'm taking okay so that's the model of the world and then the the last module uh which you could call the actor is the the module that given the objective and given the model of the world tries to figure out the best sequence of actions that will optimize your objective given your model of the world okay and robotic systems are based on this right you have all those components in exactly robots okay so as a human uh particularly a politician but you know any human any person you can be evil or stupid in in three different ways first of all your objective might be quicker so your objective instead of being the overall welfare of humanity might be just your own self-interest okay that that means you are narcissistic essentially okay or sociopath or whatever um so think about your objective um your the second uh uh component is your model of the world if your model of the world is bad then you know the the prediction of what's going to happen as a consequence of your actions would be wrong right so uh you're gonna say um you know stupid things like uh um you know we want to uh uh you know forbid uh contraception because it encourages teenage sex and that's no factually wrong right i mean there is a lot of studies that just show that it's just not true if you you know uh provide uh uh you know frequent contraception to uh to teenage teenagers you just reduce teenage pregnancy i mean it's just good right and but other countries i realize is the u.s it's kind of not quite yet um and it's because of religious reasons irrational thought right it hurts people and then the third way you can be stupid or evil is you have a good objective you're a good model of the world but somehow you cannot find a good sequence of actions that will optimize your objective given your given your model and then you have some politicians that are stupid or evil in all three different ways all three ways right they have a bad objective because they're narcissistic the mother of the world is completely wrong uh and and even then they cannot actually take a sequence of action that will actually do the stuff they want to do which to some extent make them less evil than if they were able to we're talking about humans um yes here but ai could have the same issues right same components and each of these components need to work correctly otherwise we can't trust it uh well so there is this idea this is question of value alignment right can you design the ultimate objective of the of the machine in such a way that it never uh kind of works against uh the welfare of humanity right and it might be you know people think maybe it's difficult to do this kind of design of this objective function but then this is something that we are extremely familiar with as humanity we do this absolutely all the time designing objective functions to uh entice humans to do the right thing we do this when we teach our children to behave properly we teach them morals we teach them you know to take to distinguish you know good from evil you know to behave in society to uh et cetera right so we we shape their objective function when we teach them all those things uh we do this with laws so laws are nothing more than shaping the objective function of of humans and and and entities like corporations so that whatever they do works for the common good okay so this is not something new this is not something that's new in ai we've been doing this for thousands of years it seems you need something more than the laws and you need some notion that humans are above the law and can tell the ai when it's exploiting a loophole and that is not yeah actually what it's supposed to be doing it's maybe it's not above the law but above the laws that it imposes on the ai so something something complex has to happen it's a little too early to like be too riled up about those questions given that we don't actually have the technology to build machines anywhere close to this yet but yes it's a it's a problem that we are facing that you're facing when you're building kind of you know control systems and robots and you have to design those objectives and and do that yeah we face it all the time you need to tell everybody what to do you need to give it an objective in today's formalism and yep including he needs to know when it doesn't know right and maybe uh defer to human operator because otherwise it's just going to be destructive in whatever it's doing we're back to this problem of representing uncertainty yeah that's definitely one of the big challenges yep yan thank you so much for uh joining us on the podcast it's a real pleasure to have had you such a fun conversation learned a lot thank you so much well thank you so much peter for having me it's uh it was a real pleasure [Music] [Music] you

Channel: The Robot Brains Podcast

Views: 50,403

Keywords: The Robot Brains Podcast, Podcast, AI, Robots, Robotics, Artificial Intelligence, Yann Lecun, pieter abbeel, mark zuckerberg, facebook

Id: pc0FOTQGEfM

Length: 105min 44sec (6344 seconds)

Published: Tue Sep 21 2021