Yann LeCun on AI and Identity

It's really a pleasure to talk to you today. I'm really sorry not to be there in person, but I hear it's a wonderful place. If you don't know me, my name is Yann LeCun. I'm the chief AI scientist at Facebook. I'm also a professor at NYU and I'm the co-founder of Element. So today I'm happy to share some thoughts with you about hot topics in AI and the relationship to AF F. Some of you may know me: I'm a research scientist in AI, have been working on machine learning for a very long time, and I was sort of instrumental in bringing what's called deep learning to the fore. Deep learning is really the technology that is under all of the things that we hear about surrounding AI today. That's why we talk about AI in the last few years, because of deep learning. I've been working on a number of different technologies, particularly something called convolutional nets, which is a technology that's almost universally used now for image recognition, also for speech recognition, partially also for language processing. And those are technologies that are really crucial for things like authentication that you find people, but also for helping us in our daily life, or medical applications, for self-driving cars, all kinds of applications. I'm gonna brag here a little bit, but recently I was awarded the Turing Award, which is sort of the Nobel Prize of computing, for that work in particular together with my colleagues Geoff Hinton and Yoshua Bengio. And so I'm really excited about the topic of today's conference, and I will tell you a little bit about how our AI development is relevant to this. So I'm not an economist, but I talk with economists a lot, and they're telling me that they view AI as what they call the GPT, a general purpose technology, which means a piece of technology that is going to disseminate in all corners of the economy all around the world over the next decade or two decades perhaps. So as people learn to use AI or learn to deploy it, it's going to be used in all kinds of different things: healthcare, banking, authentication, services in general, manufacturing, transportation of course with autonomous driving, and just about everything in our lives. And so it's going to transform a lot of society in many ways. Because once we have autonomous driving, we're going to design cities completely differently. Transportation is going to be different. People won't have to necessarily own cars anymore other than for fun. Communicating with each other is a very important topic, of course, because I work at Facebook, which is all about connecting people with each other. But one very important aspect of AI today is that we have language translation systems that work quite well. They're not perfect, very far from it, but they are very useful, and this may allow people to communicate across language barriers. I think it's something very important in Africa. There are so many languages in Africa, that's where basically human language was born in the first place. And so the diversity of languages is so great that it needs to be preserved, and allowing people to speak in their own language, to express themselves in their own language, and using translation to be able to communicate with other people across language barriers, I think it's crucial. Okay, so when we hear about AI today, really we mean one thing: we mean supervised deep learning. What does that mean? It means if I want to train a machine to translate from one language to another, or recognize speech, or say recognize objects in images, I collect millions of images with a cat, a dog, a table, a chair. I show an image of a cat to the machine. If it doesn't say cat, well the training procedure is going to change the parameters inside the machine so that next time I show the same image, the output gets closer to the output I want. And by sort of repeating this process, showing millions of images to the machine, eventually it gets the concept of what's a cat, what's a dog, what's a table, what's a chair. We can train systems to do things like recognize faces, recognize people from the palm of their hand, or babies from the bottom of their feet. We can recognize obscure species of insects by just taking a picture of that. And things like that. So those machines have in certain cases superhuman performance, but they require a lot of data. And so if you want to train, for example, a supervised system to do translation from let's say English to French, we'll need a huge amount of parallel data of text that has been translated from one language to another. And we show the machine a sentence in English and we tell it, here is the corresponding sentence in French. It doesn't work if you want to translate every language into every other language, because there are something like 7,000 languages in the world, and we can't possibly train 49 million machines to translate every language into every other language. We just don't have the data. And it's particularly true for a lot of languages in Africa for which there is hardly any digital data. So what we need to use in that case is what's called self-supervised learning. So it's the ability of machines to learn the nature of language, to learn to represent language for example, without being trained to do a particular task like translation. So you show them a lot of text in a particular language, and the machine learns to represent that text and sort of get to some representation of the meaning. And the magic of it is that you can do this for multiple languages in parallel, so that the representation of the meaning is the same regardless of the language in which the text is written. So you have a neural network, a deep learning system. You show it text in English or in Swahili or in whatever language that you have, and it will produce a representation of the meaning of that sentence regardless of the language. And then you can plug on top of this different neural nets that will produce translations of that text in any language you want. So there is sort of this way of training a system that has a sort of a lingua franca, if you want, some sort of universal representation of language from which we can translate into any other language. There's already some work on this at Facebook and Google and various other companies, and it's making very quick progress. And it's allowing Facebook, for example, to translate languages from one to another for which we don't have parallel text. It's very important also to be able to train machines to learn other world works by observation. Again, this is a form of self-supervised learning. For example, if you can train a machine to predict what's going to happen in a video, so you shoot a few frames in the video and then you ask it what's gonna happen next, and if the machine can train itself to do a good job at this, then it will have understood a lot about how the world works. It will have understood that the world is three-dimensional, that these objects can move, others don't, and that objects follow only certain trajectories, and that inanimate objects not supported will fall, and things like this. It will learn intuitive physics. We basically learn all the basic things we learn about the world as babies in the first few months. When we're babies, we learn all that stuff by just observing the world. We like machines to be able to do this as well. So an important application of AI which is sort of increasingly being developed is using modern AI techniques for identity, for authentication. And this is sort of an enabling technology for people in all countries in the world where identifying themselves is actually already a problem. And being able to authenticate yourself brings the possibility of a lot of new services, particularly financial services, but also health services as well. If you can prove who you are, it makes a lot of services more accessible to a lot of people. And that also includes government services, not just financial or healthcare services, but government services as well. So identity is key. I think it's a way to sort of give more people more access to more services, and if it can be done in a way that preserves privacy, which of course is a very important topic. So I will tell you a story about that. Element: I met Adam several years ago when he had the idea of creating Element, and we started talking about technologies that could be used for people to authenticate themselves, to identify themselves. And my first reaction was, oh, there are a number of techniques we can use for things like fingerprints and things, and they are well known, and I'm sure we can use them. Of course, you know, I'm a big fan of deep learning methods, I invented a bunch of them, but I thought we should try sort of more classical methods because I don't believe deep learning is ready yet for this kind of technology. And so the first systems that we built actually didn't use deep learning. And then the engineering team, they decided to use deep learning anyway, and it worked really, really well. And so that was one case where I was too timid actually with my own inventions, and the engineers at Element actually proved me wrong in a very happy way. I'm very happy that it proved me wrong. Since then, deep learning, convolutional nets, have been used for all kinds of authentication systems: for face recognition, for fingerprints, for palm print, you know, all kinds of different biometrics. And the combination of some of those is conducive to really reliable identity verification. And again, that opens the door to all kinds of new services, particularly in countries where the infrastructure for identity is not fully developed. Okay, so when we talk about AI today, we really mean deep learning. And for image recognition, when we talk about deep learning, we really mean convolutional nets. Convolutional nets is a technology I invented, and it's what is universally used now for image recognition, image analysis. So sometimes we also ask the question, why is AI needed for face recognition or palm print recognition or fingerprint recognition? Fingerprint recognition is relatively simple if you have a specific sensor, but recognizing someone from the face, from just a photo using a smartphone, or of the palm, it's actually much more challenging. To build a system that is capable of doing this using convolutional nets, using deep learning, using AI, you collect lots of data, lots of images of faces of people, and you need to have multiple pictures of the face of a single person so that you can tell how the face is going to change when the person has a different facial expression, when the person changes pose, perhaps when the person takes glasses off, or happens to grow facial hair or something like this. So you need a diverse set of images of the same person, and then you need images of lots of different people. Now you train a convolutional net to recognize the people that you have within the gallery that you have. Supervised training: show it the picture of a person, tell the machine this is person A. If the machine doesn't say person A, adjust the parameters so that next time you show the same picture, the answer gets closer to the one you want. Do this with thousands or millions of people. What you get is a system that's able to recognize those people you've trained it on. So now what you want is a system that's able to recognize anybody with just a picture of them. And so in fact, what you realize is, now you have a neural net. You can show an image to it, and it's going to extract what we call the feature vector, which is a long list of numbers that are kind of a signature of that person. And if you compare that signature to a reference signature obtained from a few pictures of a particular person, you can identify that person, you can at least authenticate that person. And so although the training is only done on a limited number of people, the system in the end is universal. It can be applied to any face and basically authenticate a new face. So that's how, that's why machine learning, AI, deep learning, convolutional nets are the key technology for authentication and for identification. All right, so let me take a very concrete example. Element here in Nigeria is working with Access Bank, which is by far the largest bank in Nigeria, to use AI-based identity verification and authentication to help people create bank accounts and authenticate financial operations, so that financial services can be provided to them without the fear of their identity being stolen or being faked. And so that's a very important concrete application. Element has been kind of doing this kind of work in various parts of the world and with great success. So Element's identification technology has been developed for a lot of applications in very different parts of the world, and therefore Element is really used to dealing with inclusivity, to making sure that systems work for all kinds of different people. This is sort of a crucial aspect of Element. The services span Southeast Asia, the Middle East, Africa, and with a you know enormous diversity of people. So the AI techniques that are used are really sort of crucial to getting the kind of performance you need in those kind of very diverse situations. Access to healthcare and medicine is very important. It's very difficult sometimes to sort of have local access to different types of services, and AI and communication services can sort of democratize medical services like medical image analysis, diagnosis systems, etc. There are wonderful applications of AI in analyzing the retina. There is a disease that's the first cause of blindness in the world, which is a consequence of diabetes, it's prevalent in some parts of Africa. There is wonderful work attempting to detect malaria parasites in microscopic images, which can be taken with a simple microscope and a cell phone, and those images can be shipped to a service that automatically detects the presence or the count of malaria parasites in an image. I mean, there's a huge amount of promise certainly in the area of healthcare. So one big challenge for the next few years is how to get machines to acquire those models of the world by observation, a little bit like babies and animals, and then use this to accelerate the type of learning that we can get them to do. That's in my opinion how we will get to robots that are dexterous and agile, machines that have a little bit of common sense, so intelligent agents that you can interact with, that will answer any questions you have, help you in your daily lives like a human assistant. This is the future. We don't know how to do it yet. It's gonna probably take a few decades. Okay, so AI is really sort of a nascent field in Africa, but it's very exciting. And you know, all the statistics of demographics are there to show us that the African population is very young. So most of the research in the future, in the next several decades, will come out of Africa. It's a numerical necessity. So Facebook has been very active in sort of helping technology get developed in Africa through its sponsorship of the AI Masters in Africa, which is located in Kigali. So that program was created by Mustafa Sisi, who used to be a research scientist at Facebook in Paris. I hired him in Paris, and he really felt very strongly that AI could have a big impact in Africa and really wanted to go back to Africa. So he went back to Africa a little over a year ago and created this master's program located in Kigali. He also runs a research lab in Ghana, which is known to Google, and it's kind of creating a sort of a center of research activity and education in Africa, which I think will be the seed for a lot of activities around AI in Africa. A lot of people from Facebook, from Google, and various other places, several universities, have spent a week or two in Kigali and taught the students there about the latest developments in AI. And it's very, very exciting to see how much interest there is. I see this every day in my Facebook feed. When I make a post, there is a lot of reaction from young people in Africa who are really enthusiastic about the topic. And I think this bodes well for the future. Actually, several years ago I founded a conference called the International Conference on Learning Representations, ICLR. In fact, it just took place last week in New Orleans in the US. Next year, in 2020, ICLR will actually take place in Addis Ababa, Ethiopia. This is the first time a major machine learning conference is taking place in Africa. And we did this on purpose because of the nascent AI research community in Africa, because we thought you know it would send a strong signal that AI in Africa is kind of coming up. And I'm really looking forward to going there next year.