Yoshua Bengio’s AI Masterclass: The Future of Machine Learning & Deep Learning

Yes my name is yosa Benjo and I'm my professor here at University of Montreal I also lead uh an Institute called the Montreal Institute for learning algorithms that is specializing in my area of science which is machine learning how computers learn from examples and um uh what is the difference between you say machine learning yes but there's also this new thing called Deep learning Right once what's the easy way to to uh yes so deep learning is inside machine learning is one of the approaches to to machine learning uh machine learning is is very general

it's about learning from examples and and scientists over the last few decades have proposed many approaches for allowing computers to learn from examples um deep learning is um introducing a particular notion That the computer learns to represent information and to do so at multiple levels of abstraction what I'm seeing is a bit abstract but to make it easier I could say that deep learning is also uh heavily inspired by what we know of the brain of how neurons compute and um it's a follow up on Decades of earlier work on what's called neural networks or

artificial neural networks so um uh what what is your background that you you got to this I Got interested in new networks uh and machine learning right at the beginning of my graduate study so when I was doing my masters I was looking for a subject and I started reading some of these papers on new networks and this this was the early days of the so-called connectionists movement and I got really really excited and I started reading more and I told U the professor who was going to supervise me that this is what I want

to do and uh And that's what I did and I continued doing it and I'm still doing it and and do you think with your uh research that you are on a on a um on a on a route or on a main line main thinking line which will get you somewhere so so it's funny that you ask this question cuz it depends it's like some days I feel very clearly that I know where I'm going and um and I can see very far I have the impression that I'm seeing far in the Future um

and I see also where I've been and and it it's there's there's a very clear path and sometimes maybe I get more discouraged and I feel uh where am I going it's all exploration I don't know where the future where the future holds of course um so I go between these two states which you need um I'm uh right now I'm uh pretty positive about a particular Direction um I've uh moved to some fundamental questions that I find really exciting and that's kind of driving a lot of my thinking looking forward can you tell me

I'm not a scientist and most of our viewers uh known as well but can you can you describe for me where where you think your path leads to because you sometimes you have a clear goal and you know where you're going right where are you Going so my main question is to understand the principles that underly intelligence and I believe that this happens through learning that intelligent Behavior arises in nature and in the computers that we're building through learning the machine the the the animal the human becomes intelligent because um it learns and um understanding

the underlying principles is like understanding the uh laws of Aerodynamics for building uh airplanes right so I and others in in my field are trying to figure out where is the equivalent of the laws of a aerodynamics but for intelligence so that's that's the quests and uh we're taking inspiration from um brains we're taking inspiration from uh a lot of our experiments that we're doing with computers trying to learn from data um we're taking inspiration from from other disciplines from physics from um uh Psychology Neuroscience um and and and and other fields uh even uh

you know uh electrical engineering and of course statistics I mean it's it's a very multidisciplinary area we must have a clue yes um I do um so one of the well it may not be so easy to explain but uh one of the big mysteries about how brains manage to do what they do is What scientists have called for many decades the the question of credit assignment that is how do neurons uh in the middle of your brain hidden somewhere uh get to know how they should change what they should be doing that would be

useful for the the whole uh Collective that is the brain and um we don't know how brain do it we now have algorithms that do a pretty good job at it uh they have their limitations but one of the things I'm Trying to do is to better understand this uh this credit assignment question and it's crucial for deep planning Because deep planning is about having um many levels of of neurons talking to each other so that's why we call them deep they're they're many layers of neurons and that's what gives them their power but the

the challenge is uh how do we train them how do they learn and it gets hotter uh the more layers you have so uh in the 80s people found how to Train networks with a single hin layer uh so just not very deep uh but they were already able to do interesting things and about 10 years ago we started discovering ways to train much deeper networks and that's what led to this current Revolution called Deep learning ah and this Revolution and uh I didn't read it in the papers so it's not font page news but

but for the Science World it's it's a breakthrough yes so in the world of uh artificial intelligence There's been a a big shift um brought by Deep planning um so there's been some scientific advances but then it turned into um advances in applications so uh very quickly these techniques turned out to be very useful for improving how computers understand speech for example uh that's speech recognition and then later uh a much bigger I would say in terms of uh impact uh effect happened when uh we discovered that these algorithms could be very good for object

Recognition from images and now many other tasks in computer vision are being done using these kinds of networks these deep networks or uh some specialized version of Deep networks called convolutional networks that work really well for images and uh and then it it moves on so now people are doing a lot of work on natural language trying to have the computer understand English sentences uh what you mean uh being able to answer some questions and so on so These are applications but uh they have a huge uh economic impact and even more in the future

that has attracted a lot of attention from from uh other scientists uh from the media and and then from of course business people who are investing billions of dollars in this right now yeah but is it exciting for you to to to to be in the middle of this this this new development it is it is very exciting uh and it's not something I had Really expected because uh 10 years ago when we started working on this there were very few people in the world maybe a handful of people interested in these questions and initially

it started started very slowly we uh it was difficult to to to get money for these kinds of things it was difficult to convince students to work on these kinds of things oh maybe you can you can explain to me the 10 years AG or whatever 12 years ago you were with Three people and then you would you because it was not popular right that's right that's right yes that's right so um there has been uh a decade before the last decade where this kind of research essentially went out of fashion uh people moved on

to other interests they lost the ambition to actually get AI to get machines to be as intelligent as us um and and also the connection between neuroscience and machine learning it got a bit divorced But a few people uh including myself and Jeff Hinton and Yan loar uh continued doing this and we started to have good results um uh and other people uh you know in the world we're also doing this uh and more people joined us and in a matter of about five years it started to be a more accepted uh area and then

and then the uh applications the success in applications started to to happen and now it's it's crazy I mean uh uh we get hundreds of applicants for example for Doing you gr studies here and uh companies are hiring like crazy and and and and and buying scientists for their research Labs do they appro you as well yeah yeah yeah be come please yes yes so I could be much richer but I chose to stay in Academia that's but so you've made some some right did some some good thing good thinking and now has become popular

yes But it has become valuable as well yes very valuable yes why uh basically it's at the heart of what companies like Google Microsoft IBM Facebook Samsung Amazon Twitter all of these companies they they see this as the key a key technology for their future products and some of the existing products already already and [Music] um are they right yeah they are of course I don't have a crystal ball so there are a lot of research questions which remain uh unsolved and it might take just a couple of years or decades to solve them we

don't know but even if say scientific research on topic stop right now and you took the current state of the art part in terms of the science and you just applied it right uh collecting Lots of data sets and to to to because these organs need a lot of data uh just applying the current science would already have a huge impact on society so so they're not I don't think they're making a very risky bet but it could be even better because we could we could actually approach human level intelligence you know that or you

think we could I think that we will have other Um other challenges to deal with uh that and some of them we currently know are in front of us others we probably will discover when we get there so now you're in the middle of the field of the of the exciting research yeah you know your right and you have a goal and you see sometimes you see it clearly and it become popular a lot of people want to study here and and a lot of companies want to invest in you so you must feel a

lot of attention or a Lot of lot of true that's true so it feel to be in the middle of this develop um so initially it's exhilarating to have all this attention and it's great to great to have all this recognition and uh and also uh it's great to attract really the you know best mind that are coming here for for doing phds and things like that it's it's it's absolutely great uh but sometimes I feel that uh it's a bit too much and that I don't deserve that much attention and That um all this

all these interactions with uh media and and and so on are taking time away from my research uh and so it's you know I have to find the right balance here uh I I think it is really important to uh continue to explain what we're doing so that more people can learn about it and take advantage of it or become researchers themselves in this area uh but I need to also focus on my main strength which is not speaking to Journalists my main strength is to come up with new ideas crazy schemes um and and

and you know interacting with students to to build new things have you thought of the possibility that you're wrong um well of course um science is an exploration um and I'm I'm often wrong I you know propos 10 things nine of which end up not working but um but we make progress so uh I I Get frequent positive feedback that tells me that we're moving in the right direction if you're right enough to go yes yes yes and and in these days um because the number of people working on this has grown really fast the

rate at which advances come is is is incredible right the the the speed of progress in this field has greatly accelerated and mostly because there are more people doing it and this is also reflected in what Companies with it yes so companies are investing a lot in basic research in this field which is unusual uh typically companies would invest in uh applied research where you take they take existing algorithms and and try to make them use them for products but but right now there's a big war between these big it companies to attract talent and

also they they understand that there's the the the potential impact the potential benefit of uh future research is Probably even greater than what we have already achieved so for these two reasons they have invested a lot in in in basic research and they are basically making offers to professors and students in the field to come work with them uh in an environment that looks a little bit like what you have in universities where they have a lot of freedom they can publish they can go to conferences and talk to their peers so so it's a

it's a good time for For the progress of science because companies are working in the same direction as universities towards really fundamental questions what then they own it gets difference so yeah that's uh something that's one of the reasons why staying in Academia uh I want to make sure that what I do is going to be uh not owned by a particular person but available for anyone um but is that the risk is is that is that really a risk that that because the knowledge is Is owned by conf me uh that uh why why

would that be a risk maybe I don't I don't think it's I don't think it's a big deal right now um so the the major research Industrial Research centers they publish a lot of what they do and they have uh they they they do have patents but they say that these patterns are protective so in case somebody would sue them but they won't prevent other people other companies from using their technology at case That's what they say um so right now there is a lot of openness in in the in the business environment for for

for this field we'll see how things are in the future there's always a danger of companies uh coming to a point where they they become protective and and but then what I think is that companies who pull themselves out of the community and not participate to the scientific progress and and and exchange with the Others they will not progress as fast and I think that's the reason why they're they understand that if they want to receive the most benefits from this progress they have to be part of the public gain of you know exchanging information

and not keeping information secret part of the mind of the universe yes exactly part of the collective that we're building of all our ideas and our understanding of the world and uh there is something about uh uh doing it part Speaking to it that enables us to be more powerful in in this understanding if we're just trying to be consumers of ideas we're not mastering those ideas as well as if we're actually trying to improve them improve them so when we do research we we get on top of things much more than if we're simply

trying to understand some existing paper and trying to use it for some product so there's something that is strongly enabling for companies to do that kind Of the but that's new it's one decade ago for example um many companies were shutting down their research labs and so on so it was it was a different Spirit but right now the spirit is openness sharing and participating in the sort of common uh development of ideas through science and and publication and so on it's funny that that the uh you say basic research it's same thing as parliamentary

yes yes yes becomes popular in it some Way well I think first of all it's appealing I mean as a person uh if I'm a a researcher a phds candidate or Professor or something uh it's much more appealing to me to know that what I do will be a contribution to humanity right rather than something secret that only I and a few people would know about and maybe some people will make a lot of money out of it but uh I don't think it's as satisfying um and as I said I think There are circumstances

right now that even from purely an economic point of view it's more interesting for companies to share right now and be part of the of the research um so I think for us to understand what you're really into I I would like to know from you some some some basic um definitions yes uh for example what what in your way of Thinking is how what you describe thinking yes is thinking right well obviously we don't know because the brain is still we don't know how the brain works yeah have a lot of information about it

uh too much maybe uh but not enough of the the kind that allows us to figure out the basic principles of how you know how we think and what does it mean at a very abstract level but of course I have my own understanding so I can share that And uh uh with the kinds of equations I drew on the board there uh and um other people in my field uh there's the this notion that what thinking is about is um adjusting your uh mental configuration um to be more coherent more consistent with everything you've

observed right and more typically uh the things you're thinking about or what you're currently observing so if I observe a picture my neurons change Their state to be in agreement with that picture and in agreement given everything that the brain already knows means that they are looking for an interpretation for that image uh which may be related to things I could do that are related like maybe oh I see this I need to go there because it tells me a message that's matters to me so everything we know is somehow built in this internal model

of the world that our brain has and um and we Get all these pieces of evidence each time we we hear something something we we listen to something and our brain is accumulating all that stuff and then what it does is try to make sense of it uh reconcile the pieces like the pieces of a puzzle and so sometimes you know it happens to you something clicks right suddenly you see a connection that explains different things um your brain does that all the time not always that You get this conscious uh impression and uh and

the thinking is is this according to me it's um it's it's finding uh structure and meaning in the things that we are observing and that we've seen and that's also what science does right science is about finding explanations for for what is around us but thinking is happening in our head where science is a is a is a social Thing and is it's it's outside go ahead s science has a part inside yeah science as a part inside of course because we are thinking when we we do science but uh science has a social aspect

science is a community uh of Minds working together in a history of of Minds having you know discovered uh Concepts that explain the world around us and sharing that in ways that that are efficient Second one thing I could talk about too is learning right you probably ask me about thinking but I think a very important Concept in my area is learning I I can explain what that how that can happen in in in in those models or brains we'll get you get the yeah yeah will be yeah I think never from [Music] okay the

yeah um so you explain what thinking is uh now I would like to know what what is Intelligence Ah that's a good question I don't think that there's a consensus on that either on what on what is intelligence if you reframe my question then I can can okay okay so what is intelligence that's that's a good question and I don't think that there's a consensus but um in my area of research uh people uh generally understand intelligence is the ability to take good decisions and what good decisions like Good for me right oh okay good

uh in a sense that they allow me to achieve my goals to uh if I was an animal to survive my predators to find food to find mates and for humans good might be achieving social status or being happy or whatever you know it's hidden in your mind what is your what is it that's good for you but but somehow we are striving to take decisions that are good for us and um in order to do that it's very clear that we need some form of Knowledge so H even a a mouse that's choosing to

go left or right in in a maze is using knowledge and and um that kind of knowledge is not necessarily the kind of knowledge you find in the book right and the mouse cannot read a book cannot write a book but in the mouse's brain there is knowledge about how to how to control the mouse's body in order to survive to find food and so on so so the mouse is actually very intelligent in the context of of the life of a mouse If you were suddenly teleported in the in the mouse you would probably

find it difficult to do the right things um so intelligence about taking right decision and it requires knowledge and now the question is to build intelligent machines or to understand how humans and animals are intelligent where are we getting the knowledge uh where can we get the knowledge and some of it is hardwired in your brain from from birth right and Some of it is going to be learned through experience and that's the the thing that we're studying in my field how do we learn or rather what are the mathematical principles for learning that could

be applied to computers and not just trying to figure out what animals how animals learn ah and there we get to to the point the learning right so uh can you explain to me because for everybody Else you think of learning You' learn at school yeah you read books and and there's someone telling you uh how the world works so what in your concept is the definition of learning yes my definition of learning is not the kind of learning that people think about when they are in school and listening to a teacher learning is something

we do all the time uh our brain is changing all the time in Response to what we are seeing experiencing and it's an adaptation and uh we are um not just um storing in our brain our experiences it's not not learning by heart that's easy know a file in a computer isar is like learning by heart you can store facts but that's trivial that's not what learning really is about learning is about integrating the information we're we're getting through experience uh into some more abstract Form that allows us to take good decisions that allow us

to predict what will happen next um that allow us to uh understand the connections between things we've seen so that's what learning is really about uh in my field we talk about the notion of generalization so that the machine can generalize from things it has seen and learned from to new situations that's the kind of learning we talk about in my field and um the way We typically do it in machines and how we think it's happening in the brain is that it's a it's a slow gradual process each time you live an experience one

second of your life uh there's going to be some changes in your brain small changes so it's like your your whole system is gradually shifting towards uh what would make it take better decisions so that's how you get to be intelligent right because you learn meaning you change the way you Perceive and act so so that next time you would see something some you would have some experience similar to what happened before you would act better or you would predict better what would have happened so it's it's very experience based yes learning learning is completely

experience-based of course um in school we think of learning as teaching Knowledge from a book or some Blackboard um but that's not really the main kind of learning Uh there is some learning happening when we when the student integrates all that information and tries to make sense of it but just storing those facts is is kind of useless is the difference that you have to have an interest in it well motivation for humans is very important because we are wired like this and the reason we're wired like this is there are so many things happening

around us uh that uh emotions help us to filter and and focus on some aspects more than Others those that matter to us right so motivation might be fear as well sometimes um but for computers uh basically they will learn what we ask them to learn we don't need to introduce uh motivation or emotions at least up to now we haven't needed to do that but uh when you uh explained this deep learning yes yes uh maybe from the perspective of of uh uh a machine and and a human um you can learn an computer

experience I think MH but no interest or well you Can you can emotions are something we're born with we're we're born with born with u uh circuits that uh make us experience emotions uh because some situations matter more to us um um so in the case of the computer we we also in a sense hardwire these things by telling the computer well uh this matters more than that and you have to learn to predict well here and here it matters less so we don't call that emotions but it could play a similar Role it looks

like emotions right but then it's still programmed absolutely so AI is completely programmed yeah but but as I understand it well you are uh reaching searching in it in it in this area where where this program where it's beyond programming that they start to think for themselves okay so there's there's an interesting connection between learning and and programming so the traditional way of putting knowledge into computers is to write a program That essentially contains all our our knowledge and uh step by step you tell the computer if this happens you do this and then you

do that and then you do that and then this happens you do that and so on and so on that's what a program is but uh when we allow the computer to learn uh we also program it but the program that is there is different it's not a program that contains the knowledge we want a computer to have we Don't program the computer with the knowledge of doors and cars and images and sounds we program the computer with the ability to learn and then the computer experiences uh uh you know uh images or videos or

sound sounds or texts and learns the knowledge from those experiences so you can think of the learning program as a meta program and we have something like that in our brain so if one part of your cortex dies you Have an accident that part used to be doing some job like maybe dealing uh interpreting music or some some types of songs or something um well if you continue uh listening to music then some other part will take over and that function may have been sort of impaired for some time but then it will be um

taken on by some other part of your cortex what what does that mean it means that the same program that does the learning was there in those two Regions of your cortex the one that used to be the doing the job and the one that does it now and that means that your your brain has this general purpose learning uh recipe uh that it can apply to different problems and that this different parts of your brain will be specialized on different tasks depending on what you do and which how the brain is connected if uh

if if we remove that part of your brain then some other parts will start Doing the job if the job is needed because you you do those experiences right so if if I had a part of my brain that was essentially dealing with playing tennis and I you know that part dies um I'm I'm not going to be able to play tennis anymore but if I continue practicing uh it's going to come back and that means that the same learning general purpose learning recipe is used everywhere at least in cortex um and this is important

not just for for Understanding brains but for for companies Building Products because we have this general purpose recipe um or family of recipes that can be applied for many tasks the only thing that really differs between those different tasks is the data the the examples that the computer sees so that's why companies are so excited about this because they can use this for many problems that they want to solve so long as they can teach the Machine by Showing it examples is it always um just learning always positive learning is positive by construction uh in

the sense that it's moving the learner towards a state of understanding of its experiences so in general yes uh because learning is about improving something now if the something you're improving is not the thing you should be improving then you could be in trouble like you we could learn uh people could Be trained into a wrong understanding of the world and then they start doing bad things right um so that's why education is so important for for humans uh and for machines right now the things we're asking the machines to do is are very simple

like understanding the content of images and and text and videos and things like that so learning is not per se positive because also you can learn wrong things right but if you're just Observing things around you taken randomly then it's just what the world is right and that's the state of the of the some kind of primitive learning of computers right now or now yeah the the the learning that computers do is very primitive it's it's it's mostly about perception and the case of language uh some kind of semantic understanding but it's still a pretty

low level understanding yeah is it is it possible for you to explain in a simple way um How is it possible for a computer to so the the way that the computer is learning is by small uh iterative changes right so so let's let's go back to my um artificial neuron Network which is a bunch of neurons connected to each other and they're connected through these these synaptic connections and at each of these connections there is uh the strength of the connection which controls how a neuron influences a ob Neuron so you can think of

that strength as a knob and what happens during learning is those knobs change we don't know how they change in the brain but in our algorithms we know how they change and we understand mathematically why it makes sense to do that and they changed a little bit each time you see an example so I show the image of a cat but the computer says it's a dog so I'm going to change those knobs so that uh It's going to be more likely that the computer is going to say cat maybe the computer outputs a score

for dog and a score for cat and so what we want to do is decrease the score for for dog and increase the score for cat so that the computer uh eventually after seeing many millions of of images starts uh seeing the right class more often and and eventually gets it as well as humans but that still sounds like putting just enough data or in the end less data uh To for good to recognize something but how do you know that the computer is learning how do you know that your well you can test it

on new images right so if the computer was only learning by heart copying the examples that it has seen it wouldn't be able to recognize a new image of say new breed of dog or a new angle uh new lighting um at the level of pixels those images could be very very different um but if the computer really Figured uh catness at least from the point of view of images it will be able to recognize new images of new cats taken in new postures and so on and that's what we call generalization so we do

that all the time we test the computer to see if it can generalize to new uh examples new images new sentences can you show that to us buil it a down but yeah you can show that that proof of of a learning skill uh yeah yeah uh I'll try to show You some some examples of that um yeah oh great uh so is there something I'm I'm missing that right now for for understanding deep burning yes okay tell me Oh I thought this was uh a a statement not a question um well but yes of

course many things that you're missing um so there are many many interesting questions in B planning but uh one of the um interesting challenges has to do with the question of uh supervised learning Versus unovis learning uh right now the way we teach the machine to um do things or to recognize things is we use what's called supervised learning where we tell the computer exactly what it should do or what output it should have uh for a given input so let's say I I'm showing at the image of a cat again uh I tell the

computer this is a cat and I have to show it millions of such images Uh that's not the way humans learn to see and understand the world or even understand language for the most part we just um make sense of what we observe without having a teacher that is sitting uh by us and telling us every second of our life this is account this is darkg supervisor that's right there's no supervisor I we do get some feedback but it's pretty rare um and sometimes it's only implicit so Uh you you do something and you know

you you get uh you get a reward but you don't know exactly what it was you did that gave you that reward or you get a um you know you talk to somebody the person is unhappy and you're not sure exactly what you did that was wrong and the person is not going to tell you general what you should have done so this is called um reinforcement learning when you get some feedback but it's a very weak like you know you did well or You didn't do well like you have an exam and you you

you you achieved uh you know 65% um well you don't know if you don't know what the errors were and what the right answers are it's it's very difficult to learn from that but we we are able to learn from that from very weak signals or no re no reinforcement role no no feedback just by observation and trying to make sense of uh all of these pieces of information that's called unsupervised learning and uh we We're not yet uh we're much more advanced with supervisor learning than with an supervised learning so all of the products

that these companies are building right now it's mostly based on supervised learning so the next step is unsupervised yes yes does it mean that unsupervised learning that a computer cann't think for themselves that means the computer will be more autonomous in some sense Right that um we don't need more autonomous right well more autonomous in its learning it's we're not talking about robots here right we're just talking about computers gradually making sense of the world around us by observation and um we probably will still need to give them some guidance but the question is how

much guidance right now we have to give them a lot of guidance basically we have to spel Everything uh very precisely for them so we're trying to move away from that so that they can uh essentially become more intelligent because they can take advantage of all of the information out there which isn't doesn't come with a a human that explains every bit and P bits and pieces but when a computer starts to learn yes is it possible to stop it stop the computer from learn sure oh it it's it sounds like if it if it

starts to learn Then it look it's just a program running it's stored in files there's nothing like there's no robot there's no I mean at least in the work we do uh it's just a program that contains files that are like the the contains the those synaptic weights for example and uh as we see more examples we change those files so that they they will correspond to taking the right decisions but there's no uh those computers don't have um A Consciousness uh there's no such thing right now at least for a while is it right

when I say uh well uh a deep learning or a self self learning computer yeah becoming more autonomous autonomous in in its learning right yes yes free well again it's probably going to be a gradual thing where the computer requires less and less of our guidance but we probably so if you think about Humans we still need guidance if you take a human uh and a baby if you nobody wants to do that experiment but you could imagine a a baby being isolated from society um that child probably would not grow to be very intelligent

would not understand the world around us as well as we do that's because we've had Parents teachers and so on guide us and we've been immersed in culture so all that matters and it it's possible that it will also be required for Computers to reach our level of intelligence the same kind of attention we're giving to humans we might need to give to computers but right now the amount of attention we have to give to computers for them to learn about very simple things is much larger than what we need to give to humans humans

are much more autonomous in their learning than machines are right now so we have a lot of progress to do in that direction is is the difference also just a simple Fact that we have biology well biology is is not magical uh biology is uh can be understood it's uh it's what biologists are trying to do and we understand a lot but there as far as the brain is concerned U there's still big holes in our understanding but so I mean I mean a baby grows yes my computer doesn't sure it can I mean we

can we can give it more uh memory and and so on right so that you can you can grow the size of the Model uh that's not a big obstacle I mean computing power is an obstacle but uh I'm pretty confident that over the next few years we're going to see um more and more computing power available as it has been in the past uh that will make it more possible to train models to do more complex tasks right so how do you tackle all the people who get who think uh this is a horror

Scenario course people start to think it and it's not it's not about that um so I I think you have you have to have a standpoint that's right I do I do um um so first of all I think there's been a bit of um excessive expression of fear uh about AI maybe because the progress has been so fast that has made us some people worried uh but if you ask people like me who are into it every day They're not worried because they can see how stupid the machines are right now and how much

guidance they need to to to move forward so um it it to us it looks like we're very far from Human level intelligence and and even you know have no idea whether one day computers will be smarter than us now that may be a short-term view what will happen in the future is hard to say uh but we can we can think about It and I think it's good that some people are thinking about the potential dangers um I think it's difficult right now to have a grasp on how what could go wrong but with

the kind of intelligence that we're building in machines right now I'm not very worried the it's not the kind of intelligence that I could for see um um exploding becoming one more intelligent by itself uh I don't think that's plausible for the kinds of like deep learning methods and so on it Even if they were much more powerful and so on it's not something I can Envision um that being said um it's good that there are people who are thinking about these long-term issues one thing I'm more worried about is the use of Technology uh

now or in the next couple of years or five or 10 years where the technology could be developed then used in a way that could either be very good for many people or not so good for many people so for example military use and Um other uses which I think are are I would consider not appropriate uh are things we need to worry about can you examples of that yeah so so so uh there's been a fuss and a letter signed by a number of scientists who try to tell the world uh we should have

a ban on um uh the use of AI for autonomous weapons that could essentially take the decision to kill by themselves so that's something that's Not very farfetched in terms of technology and the given science basically the the science is there it's a matter of building these things um but it's not something we would like to see uh in in and there could be an arms race of these things so we need to prevent it the same way that collectively uh the Nations decide Ed to uh have bans on biological weapons and chemical weapons and

to some extent on on uh um nuclear weapons the same thing Should be done for that and then there are other uses of the this technology especially as as it matures uh which I think are questionable from an ethical point of view so I think that the use of these Technologies to convince you to do things like with publicity and um uh trying to influence uh uh maybe think about influencing your vote right um as the technology becomes really stronger you could imagine people essentially using this technology to manipulate you In ways you don't realize

but is good for them but it's not good for you uh and uh I think we have to start being aware of that and uh all the issues of privacy are connected to that as well but but but uh in general because we're training currently companies are using these systems for uh advertisements where they trying to predict uh what they should show you so that you will be more likely to buy some product right so it Seems you know uh not so bad but if you push it you know he they might bring you into

doing things that are not so good for you I don't know like smoking or whatever right well we well we just stopped at a point where I was going to ask you about is that why you wrote the Manifest about diversity into in thinking because I okay if you okay I'm door because computers uh you can learn a lot of Things but uh it's almost imaginable unimaginable that you learn in diversity am I correct that that that that's that has a connection if you want I will elaborate no so you're asking me about diversity and

um I can say several things first of all people were not aware of um the kinds of things do in in AI with machine learning deep learning and so on may not realize that um the Algorithms the methods we're using already include a lot of um what may look like diversity creativity so for the same input the computer could produce different answers and so there's a bit of Randomness just like for us place twice in the same situation we don't always take the same decision and there are good reasons for that both for us and

for Compu Compu so that's that's the first part of it but uh there's another aspect of diversity which I have Studied in a paper a few years ago which is um maybe uh even more interesting um diversity uh is very important for example for evolution to succeed because Evolution uh performs a kind of search in the space of uh genomes of of the the the the the blueprint of each individuals and up to now machine learning has considered what happens in a single individual how we learn how our machine can learn but has not really

Investigated much the role of having uh a group of individuals learning together so a kind of society and um in this paper a few years ago I postulated that learning in uh an individual could get stuck that um if we were alone learning by observing the world around us we might get stuck with poor model of the world and we get unstuck by talking to other people and by learning from other people In the sense of uh they can communicate some of the ideas they have how they interpret the world and that's what culture is

about uh culture has many meanings but that's the the meaning that I have that it's uh the the not just the accumulation of knowledge but but how knowledge gets created through communication and sharing and um what I postulated in that paper is that uh there is a it's called an optimization problem that uh can get The learning of an individual um uh to not progress anymore in a sense that as I said before learning is uh a lot of small changes but sometimes there's no small change that really makes you progress um so you need

some kind of uh external kick that brings a new light to things um and another uh connection to Evolution the connection to Evolution actually is that uh this uh this small kick we get from others is like uh we Are building on top of existing solutions that others have uh uh come up with and of course the process of science is very much like this we're building other scientists ideas but it's true for culture in general and uh this actually makes the whole process of um building more intelligent beings uh much more efficient in fact

we know that since humans um have made progress thanks to Evolution and not just uh thanks to Culture and not just to Evolution um we've been making our intelligence has been increasing much faster so Evolution uh is slow whereas uh you can think of culture the evolution of culture as a process that's much more efficient because we are manipulating the right objects so what does this mean in practice it means that um just like Evolution needs diversity to succeed because they many different uh uh variants of the same type of genes That are uh randomly

Chosen and tried and and the best ones combined together to to create new Solutions just like this in in cultural Evolution which is really uh an important uh important for our intelligence I was saying we need diversity we need uh not just one school of thought we need to allow all kinds of exploration most of which may fail um so in science we need to be open to new ideas even it's very likely it's not going to work it's good that people Explore right um otherwise we're going to get stuck in some in the space

of possible interpretations of the world it may take forever before we we escape is it like doing basic research that you don't have yes that's right so basic research is exploratory it's not trying to build a product it's just trying to understand and it's going in all possible directions according to our intuitions Of know what may be more interesting but but without a strong constraint so yeah basic research uh is like this but but there's a danger because humans they like you know fashionable things and Trends and compare each other and so on that uh

um that we are we're not giving enough freedom for exploration and it's not just science it's in general right in society we should allow a lot more freedom uh we should allow marginal ways Of being and and and and and and doing things to to coexist but if you if you uh allow this Freedom um of course a lot of people think well let's don't go that way because then you have autonomous self-thinking computers uh uh creating their own diversity and so there are a lot of scenarios which people think of because they don't know

um which yeah so this this did well it's a gamble um and I'm more on the positive Side I think that the the rewards we can get by having more intelligence in our machines is is immense and the way I I think about it is it's not a competition between machines and humans uh technology is is expanding what we are thanks to technology we are now already much stronger and more intelligent than we were the same way that uh um the industrial revolution has kind of increased our strength and our ability to do things physically

the the Sort of computer Revolution and now the AI Revolution is going to increase continue to increase our cognitive abilities it sounds very logical but but I can imagine you not get tired of all those people who don't uh who fear fear this development right but I think we should be conscious that a lot of that fear is due to um a projection into things we are familiar with so we are thinking of AI like we think like we see them in movies we're Thinking of AI like we see some kind of alien from another

planet like we see animals when we think about another being we think that other being is like us and and so we are greedy we want to dominate the rests and uh if our survival is at stake we're ready to kill right so we project that some machine is going to be just like us and if that machine is more powerful than we are then we're in deep trouble right so it's just because we're making that Projection but actually the machines are not some being that has an ego and a survival Instinct it's actually something

we decide to put together it's a program and so we should be smart enough and wise enough to program these machines to uh be useful to us rather than go towards their own needs they will cater to our needs because we will Design them that way I understand that but then there's also this theory of suppose you you can Develop machines or or robots or uh that can uh the self learn uh so if that grows with this uh uh Power of of the yes there's there is some acceleration in their in their intelligence or

maybe maybe not I I don't that's not the way I what you're saying is appealing if I was to read a science fiction book but it doesn't correspond to how I see AI uh and the kinds of AI we're doing um I don't see such Acceleration in fact what I see is the opposite what I foresee is more like barriers than acceleration so our slowing you down yes so uh our experience in research is that we make progress and then we encounter a barrier a challenge a difficulty that the algorithm goes so far and then

can't make progress even if we have more compute power that's not really the issue the issue are more are like basically computer science issues that Things get harder as you try to solve exponentially harder really much much harder as you try to solve more complex problems um so it's actually the opposite I think that happens that and I think that would also explain maybe to some extent why we're not super intelligent ourselves I mean the sense that uh our intelligence is kind of limited there are many things for which we do the wrong take the

wrong decision and and it's true also of animals like Why is it that animals some animals have much larger brains than we do and they're not that smart um and you know you could come up with a bunch of reasons but it's not they they they have a bigger brain right so and their brain like mammal's brain is very very close to ours so um there's it's it's hard to say now I think it's fair to consider the worst scenarios and to study it and have you know people we seriously considering What could happen and

how we could prevent any dangerous thing I think it's actually important that some people do that but right now I I see this as a very long-term potential and and the most plausible scenario is not that according to my vision does have to do with fact that you try to develop this deep deep learning um that if you know how it works then you also know how to deal with it Is that why you are confident in not seeing any problem you're right that I think we are more afraid of things we don't understand and

scientists who are working with de planning every day don't feel that they have anything to fear because they can they understand what's going on and they can see clearly that there is no danger that's foreseeable um so you're right that's part of it there's the psychology of uh seeing the machine as some other being There's the lack of knowledge there is the influence of Science Fiction so all these factors come together and also the fact that the technology has been making a lot of priv recently so all of that I think creates a kind of

exaggerated fear I'm not saying we shouldn't have any fear I'm just saying it's exaggerated right now um [Music] well it is your your main part of life Or your how you uh fill the day um is that thinking is your work thinking or do you physically physically do I'm thinking all the time yes and whether I'm thinking on the things that matter to me the most uh maybe not enough uh managing a big Institute with a lot of students and so on means my time is dispersed but but when I can you know um focus

or when I'm uh in in a scientific discussion with people and so on of course there's a lot of thinking And it's really important that's how we move forward um yeah yeah what what does it mean but the first question I asked you was about what is thinking yes and now we are uh we are back to that question yeah yeah so you are a a thinker so so what what what happens okay during the day yeah with you so when I listen to Somebody explaining something maybe one of my students talking about an experiment

or another researcher talking about their Idea um something builds up in my mind to try to understand what is going on and um and that's already thinking but then things happen so uh other pieces of information and understanding connect to this and I see some flaw or some uh some some connection and um and that's that's where the creativity creativity comes in and I it I have the impulse of talking about it and that's just one turn in in a Discussion uh and we go like this and uh and and new ideas spring like this

um and it's very very rewarding uh is it possible for you not to think well I uh yes yes it is possible not to think it's hard but if you really you know if you really relax or you are experiencing something very intensely then you're you're not you're not into your thoughts you're you're into just the the some some Present time experience yes like it's more emotional rather than the rational for example yes and but thinking isn't just rational a lot of it is I don't mean it's irrational but it's a lot of the thinking

is something that happens uh somehow behind the scenes it has to do with intuition that has to do with um analogies and um it's not necessarily A causes B causes C it's not that kind of logical thinking that's going on in my mind most of the time It's much softer and uh that's why we need the math in order to filter and and fine tune the ideas but the raw thinking is very um um fuzzy and uh but it's very rich because it's connecting a lot of things together and it's discovering the the inconsistencies that

allow us to move to the next stage and and and solve Problems are you aware of that you are in that situation when you are thinking it happens to me uh I used to spend some time uh meditating and there you learn to pay attention to your own thoughts um so it does happen to me it happens to me also that I get so immerse in my thoughts in ordinary daily activities that people thought think that I'm very distracted and not present and they can be Offended um but it's not always like this sometimes I'm

actually very very present I can be very very present to some somebody talking to me um and really important for what I my job right because uh if I listen to somebody in a way that's not complete uh I can't really understand fully and participate in in in the in a rich uh exchange I kind I can imagine that when you are focused on on a thought yeah uh you are having this problem and you Thinking about it thinking yeah and then you are in a situation that other people demand something else of you right

like attention for your children or whatever that's yes then there's something in you which decides uh to to keep on FAL or how does it work with you right you don't want to lose the thought of course that's right so I write I have some notebooks I write my ideas um often when I wake up or sometimes an idea comes and I want to Write it down like if I was afraid of losing it but actually the good ideas they don't they don't go it turns out very often I write them but I don't even

go back reading them it's just that uh it makes me feel better and it anchors Al the fact of writing an idea kind of makes it uh take more room in my mind um and um and there's also something to be said about concentration so so my work now because I'm immersed with so many people um can be very Distractive but um to really make big progress in science um I also need times when I can be very focused and um and where the the ideas about a problem and a different points of view and

and and all the elements sort of fill my mind I'm completely filled with this that's when you can be really productive and it might take a long time before you reach that state sometimes it could take years for a student to really Uh go deep into a subject so that he can be fully immersed in it and that's when you can really start seeing through things and getting things to stand together and solidly and now you can you can extend science right now when things are solid in your mind you can move forward like a

face of understanding yeah yeah there it is when when you need enough concentration on something to really really to really get these these moves there there's the Other mode of thinking which is the brainstorming mode where out of the blue I started discussion five minutes later something comes up so that's more like random and it's also very um it could be very productive as well uh it it depends on the stimulation from someone else so someone introduces a problem and and immediately I get a something comes up and we have a maybe an exchange um

so that more superficial but A lot of good things come out of that exchange because of the the the the brainstorming uh whereas the other there's the other mode of thinking which is I'm alone uh nobody bothers me nobody's asking for my attention I'm walking uh I'm I'm half asleep and there I can fully concentrate eyes closed or not really paying attention to what's going on in front of me because I'm completely in my thoughts when do you think when yeah Um your day let's start this so the the time when uh the two times

when I spend more um on this concentrated thinking is uh usually when I wake up and uh when I'm walking back and forth between home University just enlarge this moments and what happens well so I emerge to Consciousness like everybody does every morning and uh eyes close and so on and um some some thought related to a Research question or maybe non-research question uh comes up and um and if I'm interested in it I start like going deeper into it and and still your eyes closed still my eyes closed and um and then it's like

if you you see uh a thread dangling and you pull on it and then more stuff comes down and now you see more things and you pull more and like there's an avalanche of things coming right so the more you you you pull on Those strings and the more new things come or information comes together um and sometimes it goes nowhere and sometimes that's how new ideas come about and in what stage in this pulling the thread you open your eyes uh I could stay like this for an hour eyes closed yeah pulling a thread

seeing What's happen me yeah uh often what happens is uh I see something that I haven't seen before and I get too excited so that Waks me up and I want to write it down so I have my notebook not far and I write it down um or I want to send an email to somebody saying Oh I thought about this and it's like 6: in the morning and and they wonder if I'm working all the time yeah so and then what happens then then you then you you woke up oh yeah open your eyes

or you yeah you wrot it down Well so once I've I'm runting it down my eyes are open and uh and it's like I'm I feel relieved right it's like okay now now I go and maybe have breakfast or take a shower or something so having written it down um and it might take some time to write it down um um also sometimes I write an email and then it's longer and now it it the the act of writing it is a different thing so there's Initial sort of spark of vision which is still very

fuzzy but then when you have to communicate the idea to someone else say in an email you have to really make a different kind of effort you realize some flaws in your initial ideas and you have to clean it up and make sure it's understandable and now it takes a different form um and uh and sometimes you realize when you do it that it Was n nothing really be good yeah it was just half dream you know um and what does your partner think oh he has they I use game with some IDE did I

I didn't understand the question but when and what does your partner think of this that you wake up and then or you have to write something down she she's fine with that um I think she she's she's glad to see this this kind of thing happen she she's happy for me that I I I live these very uh rewarding Moments um but she understands what happens yeah I mean I I I tell her often oh I just had an idea I want to say oh I just want to that she understands uh what do you

mean the science yes uh no no but she understands that uh it's really important for me and this is how uh I move forward in in my my work and um and also how emotionally uh fulfilling it is uh okay and then then you uh at a certain moment of that you have To go to work yes let's talk about the the walk you do every day yes so what's what does it mean so that walk is you can really think of it as a kind of meditation so you know you walk what you doing

if you want to so every day I walk for yeah yeah so every day uh I walk up the hill from my home to the university and it's about half an hour and uh it's more or less always the same path um and I and because I know this path so well I don't have to really pay Much attention to what's going on and I can just relax and let thoughts go by and eventually focus on something uh or not sometimes it's just um uh maybe more in the evening where I'm tired maybe just a

way to relax and uh let go let go quality thinking time yes absolutely because uh I'm uh I'm not bombarded by the outside world uh I can just noral people are bombarded by every signs and Cars and sounds yeah the weather yeah I I I kind of ignore that so you are you are whether there are FS around you so when I was young I used to uh hit my head on on [Laughter] poles because you were thinking yeah so yeah or weeding while walking so and now it doesn't happen anymore no well actually it

does now Because I sometimes I check my my phone I see lots of people do that not being paying attention to what's going on yeah yeah so well we will film your walk maybe something happens uh but uh are they doing this walk if you do it for such a long time yeah walking uphill yeah uh that's kind of a nice metaphor walking up the hill yeah um uh are there uh all this ruled uh situations or Or positions or places when you had some really good ideas that that you that you can remember well

I was I was waiting at the traffic light or or there was yeah I have some memories of specific moments uh going up um thinking about some of the ideas i' that I have been uh going through my mind over the last year in particular I guess these are more recent Memories um so can you enlarge one of those H like you did with waking up right right so as I said earlier it's like if the rest of the world is in a haze right it's like there's automatic control of the walking and watching for

other people and cars potentially uh but but I I'm it's like it's like if I had a 3D projection of my thoughts in front of me that are taking most of the Room and uh and I I my thinking works a lot by visualization and I think a lot of people like this it's a very nice tool that we have using our uh kind of visual analogies to uh understand things even if it's not an a faithful portrait of what's going on on the visual analogies are really helping me at least to make sense of

things so it's like if I had pictures in my mind to illustrate what's going on and it's like I see little you know Uh uh what do I see I see uh information flow in neural networks uh it's like if I was running a simulation in my mind of what would happen if some rule of conduct was followed by by you know in this algorithm in this process and that's when you walk up the that's what you see yeah yeah so so it's it's like if I was running a a a computer simulation in my

mind to try to Figure out uh what would happen if some if I made such choices or if we consider such equation and what would it entail what would happen imagine different situations and then um of course it's not as detailed if we as if we did a real computer simulation but uh it provides a lot of insight for what's going on but then you walk up the hill every day yeah and describe the the most Defining moment during one of those walks um where you were where you stood which corner which um well so

I remember a particular moment uh I was walking on the North sidewalk of the Queen Mary Street and uh I was seeing the big uh church we have there which is called the aatar it's beautiful and uh and then I I got this insight about Perturbations propagating in in brains uh maybe you want do that um from the beginning or just the last sentence the last last one and so then I got this Insight uh as visually of uh these perturbations happening on neurons that propagate to other neurons that propagate to other neurons and it

was like some like I'm doing with my hands but it was like something Visual um and and and suddenly I had the thought that this this could this could work that this could explain things that I was trying to understand how did it feel great I think uh of all the good feelings that we can have in life uh though the feeling we get when something clicks the Eureka is probably maybe the the strongest the most powerful the one that we can seek again and again and only Brings positive things uh maybe you know stronger

than food and sex and these usual um good things we get from from our experience those what you mean this moment this this kind of these kinds of moments uh provide pleasure it's a different kind of pleasure just like you know different Pleasures a different sensory pleasure and so on but it's it's really like I Think your when your brain realizes something understand something it's like you send yourself some some molecules to reward you say oh great do it again if you can right do it again yeah yeah that's that's my job so uh this

is one moment at the church was it a coincidence that it was at a church no there's nothing to do with it I don't believe in God but the when when yeah of I don't Believe in God either but but if you think of God as uh someone who created us as is and he is our example yes um uh trying to understand what's happening in your head or your brass isn't that what other people call God or looking for I'm not sure I understand your question [Music] um how can I rephrase that one uh

um When you understand what how brain works yes maybe then you understand who go is when we understand how brains work we understand who we are to some extent I mean a very important part of us that's one of my motivations um and the process of doing it is something um that defines us individually but also as a as a collective as a group as a society so uh there maybe some Connections to religion which are about connecting us to some extent yeah it's one of those layers you were talking about religion is one of

them of course yep so but during this will you this half an hour then you are almost here so um sometimes I think it's too short but then and know I have things to do so but let's continue this metaphor it's uphill when you are uphill yeah what do you See uh I feel so when going uphill my body is working hard I mean I'm not running but I'm walking and I can feel the muscles warming up and uh my whole body becoming more full with energy and I think that helps the brain as well

that's how it feels anyway but I mean when you Moses went up to the m right and he saw the the promis L and you go uphill yes what do you see Uh when I go uphill I see I see the university but uh but there's something that's related to your question which is yeah each time I have these insights these Eureka moments it it's like seeing the promised land it's very much like that it's like you have a glimpse of something you had never seen before and it looks great and you feel like you

now see a path to go there so I think it's very very close to this Idea of seeing the promised land but of course it's not just one promised land it's one step to you know the next Val in the next Valley and that's how we climb really you know big mountains so is there anything you want to add to this yourself because I think we are ready now to go uphill no fine maybe just a few questions about Friday so okay what you're going to do uh what what are you Going to do on

Friday so Friday I'm going to um make a presentation to the rest of the researchers in the lab in The Institute about one of the topics I'm most excited about these days which uh is trying to bridge the gap between what we do in machine learning has which has to do with AI and building intelligent machines and uh and the brain I'm not really a brain expert I'm more a machine learning person but I Talk to neuroscientists and so on and I I try I really care about um the big question of how is the

brain doing the really complex things that it does and so the work I'm going to tell you tell about Friday is one of small step in that direction that we've achieved in the last few months uh on your path to the promised land uh yes exactly that's right and I've been making those small steps you know on this particular topic for for about a Year and a half so it's not like just something happens and you're there right uh it's a lot of in insights that make you move and and and and and get understanding

and um um science makes progress by steps most of those steps are small some are slightly bigger uh seen from the outside sometimes people have the impression that oh there's this big breakthrough breakthrough and journalists like to talk about breakthrough breakthrough Breakr break but actually science is very very Progressive because we gradually understand better the world