A human can learn to drive a car in 15 hours of training without crashing into anything. If you want to use the current reinforcement learning methods to train a car to drive itself, the machine will have to drive off cliffs 10,000 times before it figures out how not to do that.
VP & CHIEF AI SCIENTIST, FACEBOOK PROFESSOR OF COMPUTER SCIENCE, NYU
Yann LeCun has been involved in the academic and industry side of AI and Machine Learning for over 30 years. Prior to joining Facebook, Yann worked at AT&T’s Bell Labs, where he is credited with developing convolutional neural networks—a machine learning architecture inspired by the brain’s visual cortex. Along with Geoff Hinton and Yoshua Bengio, Yann is part of a small group of researchers whose effort and persistence led directly to the current revolution in deep learning neural networks.
MARTIN FORD: Let’s jump right in and talk about the deep learning revolution that’s been unfolding over the past decade or so. How did that get started? Am I right that it was the confluence of some refinements to neural network technology, together with much faster computers and an explosion in the amount of training data available?
YANN LECUN: Yes, but it was more deliberate than that. With the emergence of the backpropagation algorithm in 1986-87, people were able to train neural nets with multiple layers, which was something that the old models didn’t do. This resulted in a wave of interest that lasted right through to around 1995 before petering out.
Then in 2003, Geoffrey Hinton, Yoshua Bengio, and I got together and said, we know these techniques are eventually going to win out, and we need to get together and hash out a plan to renew the community interest in these methods. That’s what became deep learning. It was a deliberate conspiracy, if you will.
MARTIN FORD: Looking back, did you imagine the extent to which you would be successful? Today, people think artificial intelligence and deep learning are synonymous.
YANN LECUN: Yes and no. Yes, in the sense that we knew eventually those techniques would come to the fore for computer vision, speech recognition, and maybe a couple of other things—but no, we didn’t realize it would become synonymous with deep learning.
We didn’t realize that there would be so much of an interest from the wider industry that it would create a new industry altogether. We didn’t realize that there would be so much interest from the public, and that it would not just revolutionize computer vision and speech recognition, but also natural language understanding, robotics, medical imaging analysis, and that it would enable self-driving cars that actually work. That took us by surprise, that’s for sure.
Back in the early ‘90s, I would have thought that that this kind of progress would have happened slightly earlier but more progressively, rather than the big revolution that occurred around 2013.
MARTIN FORD: How did you first become interested in AI and machine learning?
YANN LECUN: As a kid, I was interested in science and engineering and the big scientific questions—life, intelligence, the origin of humanity. Artificial intelligence was something that fascinated me, even though it didn’t really exist as a field in France during the 1960s and 1970s. Even with a fascination for those questions, when I finished high school I believed that I would eventually become an engineer rather than a scientist, so I began my studies in the field of engineering.
Early on in my studies, around 1980, I stumbled on a philosophy book which was a transcription of a debate between Jean Piaget, the developmental psychologist, and Noam Chomsky, the linguist, called, Language and Learning: The Debate Between Jean Piaget and Noam Chomsky. The book contained a really interesting debate between the concepts of nature and nurture and the emergence of language and intelligence.
On the side of Piaget in the debate was Seymour Papert, who was a professor at MIT in computer science and who was involved with early machine learning and arguably actually killed the field off in the first wave of neural nets in the late 1960s. Here he was, 10 years later, singing the praise of a very simple machine learning model called the perceptron that had been invented in the 1950s, and that he had been working on in the 1960s. That was the first time I read about the concept of a learning machine, and I was absolutely fascinated by the idea that a machine could learn. I thought learning was an integral part of intelligence.
As an undergrad, I dug up all the literature I could find about machine learning and did a couple of projects on it. I discovered that nobody in the West was working on neural nets. A few Japanese researchers were working on what became known as neural networks, but no one in the West was, because the field had been killed in the late ‘60s in part by Seymour Papert and Marvin Minsky, the famous American AI researcher.
I carried on working on neural nets on my own, and I did a PhD in 1987 titled, Modeles connexionnistes de l’apprentissage (Connectionist learning models). My advisor, Maurice Milgram, was not actually working on this topic, and he told me outright, “I can be your official advisor, but I can’t help you technically.”
I discovered through my work that in the early 1980s, there was a community of people around the world who were working on neural nets, and I connected with them and ended up discovering things like backpropagation in parallel with people like David Rumelhart and Geoffrey Hinton.
MARTIN FORD: So, in the early 1980s there was a lot of research in this area going on in Canada?
YANN LECUN: No, this was the United States. Canada was not on the map for this type of research yet. In the early 1980s, Geoffrey Hinton was a postdoc at the University of California, San Diego where he was working with cognitive scientists like David Rumelhart and Jay McClelland. Eventually they published a book explaining psychology by simple neural nets and models of computation. Geoffrey then became Associate Professor at Carnegie Mellon University, and only moved to Toronto in 1987. That’s when I also moved to Toronto, where I was a postdoc in his lab for one year.
MARTIN FORD: I was an undergraduate studying computer engineering in the early 1980s, and I don’t recall much exposure to neural networks at all. It was a concept that was out there, but it was definitely very much marginalized. Now, in 2018, that has changed dramatically.
YANN LECUN: It was worse than marginalized. In the ‘70s and early ‘80s it was anathema within the community. You couldn’t publish a paper that even mentioned the phrase neural networks because it would immediately be rejected by your peers.
In fact, Geoffrey Hinton and Terry Sejnowski published a very famous paper in 1983 called, Optimal Perceptual Inference, which described an early deep learning or neural network model. Hinton and Sejnowski had to use code words to avoid mentioning that it was a neural network. Even the title of their paper was cryptic; it was all very strange!
MARTIN FORD: One of the main innovations you’re known for is the convolutional neural network. Could you explain what that is and how it’s different from other approaches in deep learning?
YANN LECUN: The motivation for convolutional neural networks was building a neural network that was appropriate for recognizing images. It turned out to be useful for a wide-range of tasks, such as speech recognition and language translation. It’s somewhat inspired by the architecture of the visual cortex in animals or humans.
David Hubel and Torsten Wiesel did some Nobel prize-winning work in neuroscience in the 1950s and 1960s about the type of functions that the neurons in the visual cortex perform and how they’re connected with each other.
A convolutional network is a particular way of connecting the neurons with each other in such a way that the processing that takes place is appropriate for things like images. I should add that we don’t normally call them neurons because they’re not really an accurate reflection of biological neurons.
The basic principle of how the neurons are connected is that they’re organized in multiple layers and each neuron in the first layer is connected with a small patch of pixels in the input image. Each neuron computes a weighted sum of its inputs. The weights are the quantities that are modified by learning. The neurons only see a tiny window of pixels of the input, and there’s a whole bunch of neurons that look at the same little window. Then, there’s a whole bunch of neurons that look at another slightly shifted window, but this bunch performs the same operation as the other bunch. If you have a neuron that detects a particular motif in one window, you’re going to have another neuron that detects exactly the same motif in the next window and other neurons for all windows across the image.
Once you put all those neurons together and you realize what kind of mathematical operation they do, that operation is called a discrete convolution, which is why this is called a convolutional net.
That’s the first layer, and then there’s a second layer, which is a non-linearity layer—basically a threshold where each neuron turns on or turns off if the weighted sum computed by the convolution layer is above or below the threshold.
Finally, there’s a third layer that performs what’s called a pooling operation. I’m not going to cover it in detail, but it basically plays a role in making sure that when the input image is slightly shifted or deformed, the output responses don’t change that much. That’s a way of building a bit of invariance to distortion shifts or deformation of the object in the input image.
The convolutional net is basically a stack of layers of this type—convolution, non-linearity, pooling. You stack multiple layers of those, and by the time you get to the top, you have neurons that are supposed to detect individual objects.
You might have a neuron that turns on if you put an image of a horse in the image, and then you have one for cars, people, chairs, and all other categories you might want to recognize.
The trick is that the function that this neural network is doing is determined by the strength of the connections between the neurons, the weights, and those are not programmed; they’re trained.
This is what is learned when you train the neural net. You show it the image of a horse, and if it doesn’t say “horse,” you tell it that it’s wrong and here is the answer that it should have said. Then by using the backpropagation algorithm, it adjusts all the weights of all the connections in the network so that next time you show the same image of a horse, the output would be closer to the one you want, and you keep doing this for thousands of images.
MARTIN FORD: That process of training a network by giving it images of cats or horses, and so on, is what’s called supervised learning, correct? Is it true to say that supervised learning is the dominant approach today, and that it takes huge amounts of data?
YANN LECUN: Exactly. Almost all of the applications of deep learning today use supervised learning.
Supervised learning is when you give the correct answer to the machine when you’re training it, and then it corrects itself to give the correct answer. The magic of it is that after it’s been trained, it produces a correct answer most of the time in categories that it’s been trained on, even for images it’s never seen before. You’re correct, that does typically require a lot of samples, at least the first time you train the network.
MARTIN FORD: How do you see the field moving forward in the future? Supervised learning is very different from the way a human child learns. You could point at a cat once and say, “there’s a cat,” and that one sample might be enough for a child to learn. That’s dramatically different from where AI is today.
YANN LECUN: Well, yes and no. As I said, the first time you train a convolutional network you train it with thousands, possibly even millions of images of various categories. If you then want to add a new category, for example if the machine has never seen a cat and you want to train it to recognize cats, then it only requires a few samples of cats. That is because it has already been trained to recognize images of any type and it knows how to represent images; it knows what an object is, and it knows a lot of things about various objects. So, to train it to recognize a new object, you just show it a few samples, and you just need to train a couple of the top layers.
MARTIN FORD: So, if you trained a network to recognize other kinds of animals like dogs and bears, then would it only take a small amount of data to get to a cat? That seems not so different from what a child is probably doing.
YANN LECUN: But it is different, and that’s the unfortunate thing. The way a child learns (and animals, for that matter) is that most of the learning they do is before you can tell them, “this is a cat.” In the first few months of life, babies learn a huge amount by observation without having any notion of language. They learn an enormous amount of knowledge about how the world works just by observation and with a little interaction with the world.
This sort of accumulation of enormous amounts of background knowledge about the world is what we don’t know how to do with machines. We don’t know what to call this, some people call this unsupervised learning, but it’s a loaded term. It’s sometimes called predictive learning, or imputative learning. I call it self-supervised learning. It’s the kind of learning where you don’t train for a task, you just observe the world and figure out how it works, essentially.
MARTIN FORD: Would reinforcement learning, or learning by practice with a reward for succeeding, be in the category of unsupervised learning?
YANN LECUN: No, that’s a different category altogether. There are three categories essentially; it’s more of a continuum, but there is reinforcement learning, supervised learning, and self-supervised learning.
Reinforcement learning is learning by trial and error, getting rewards when you succeed and not getting rewards when you don’t succeed. That form of learning in its purest form is incredibly inefficient in terms of samples, and as a consequence works well for games, where you can try things as many times as you want, but doesn’t work in many real-world scenarios.
You can use reinforcement learning to train a machine to play Go or chess. That works really well, as we’ve seen with AlphaGo, for example, but it requires a ridiculous number of samples or trials. A machine has to basically play more games than all of humanity in the last 3,000 years to reach good performance, and it works really well if you can do that, but it is often impractical in the real world.
If you want to use reinforcement learning to train a robot to grab objects, it will take a ridiculous amount of time to achieve that. A human can learn to drive a car in 15 hours of training without crashing into anything. If you want to use the current reinforcement learning methods to train a car to drive itself, the machine will have to drive off cliffs 10,000 times before it figures out how not to do that.
MARTIN FORD: I guess that’s the argument for simulation.
YANN LECUN: I don’t agree. It might be an argument for simulation, but it’s also an argument for the fact that the kind of learning that we can do as humans is very, very different from pure reinforcement learning.
It’s more akin to what people call model-based reinforcement learning. This is where you have your internal model of the world that allows you to predict that when you turn the wheel in a particular direction then the car is going to go in a particular direction, and if another car comes in front you’re going to hit it, or if there is a cliff you are going to fall off that cliff. You have this predictive model that allows you to predict in advance the consequence of your actions. As a result, you can plan ahead and not take the actions that result in bad outcomes.
Learning to drive in this context is called model-based reinforcement learning, and that’s one of the things we don’t really know how to do. There is a name for it, but there’s no real way to make it work reliably! Most of the learning is not in the reinforcement, it’s in learning the predictive models in a self-supervised manner, and that’s the main problem we don’t know how to solve today.
MARTIN FORD: Is this an area that you’re focused on with your work at Facebook?
YANN LECUN: Yes, it is one of the things that we’re working on at Facebook. We’re working on a lot of different things, including getting machines to learn by observation from different data sources—learning how the world works. We’re building a model of the world so that perhaps some form of common sense will emerge and perhaps that model could be used as kind of a predictive model that would allow a machine to learn the way people do without having to try and fail 10,000 times before they’ve succeeded.
MARTIN FORD: Some people argue that deep learning alone is not going to be enough, or that there needs to be more structure in the networks, some kind of intelligent design from the onset. You seem to be a strong believer in the idea that intelligence will emerge organically from relatively generic neural networks.
YANN LECUN: I think that would be an exaggeration. Everybody agrees that there is a need for some structure, the question is how much, and what kind of structure is needed. I guess when you say that some people believe that there should be structures such as logic and reasoning, you’re probably referring to Gary Marcus and maybe Oren Etzioni.
I actually had a debate with Gary Marcus on this earlier today. Gary’s view isn’t particularly well accepted in the community because he’s been writing critically about deep learning, but he’s not been contributing to it. That’s not the case for Oren Etzioni because he’s been in the field for a while, but his view is considerably milder than Gary’s. The one thing all of us agree on, though, is that there is a need for some structure.
In fact, the very idea of convolutional networks is to put a structure in neural networks. Convolutional networks are not a blank slate, they do have a little bit of structure. The question is, if we want AI to emerge, and we’re talking general intelligence or human-level AI, how much structure do we need? That’s where people’s views may differ, like whether we need explicit structures that will allow a machine to manipulate symbols, or if we need explicit structures for representing hierarchical structures in language.
A lot of my colleagues, like Geoffrey Hinton and Yoshua Bengio, agree that in the long run we won’t need precise specific structures for this. It might be useful in the short term because we may not have figured out a general learning method for self-supervised learning. So, one way to cut corners is to hardwire the architecture; that is a perfectly fine thing to do. In the long run, though, it’s not clear how much of that we need. The microstructure of the cortex seems to be very, very uniform all over, whether you’re looking at the visual or prefrontal cortex.
MARTIN FORD: Does the brain use something like backpropagation?
YANN LECUN: We don’t really know. There are more fundamental questions than that, though. Most of the learning algorithms that people have come up with essentially consist of minimizing some objective function.
We don’t even know if the brain minimizes an objective function. If the brain does minimize an objective function, does it do it through a gradient-based method? Does the brain have some way of estimating in which direction to modify all of its synaptic connections in such a way as to improve this objective function? We don’t know that. If it estimates that gradient, does it do it by some form of backpropagation?
It’s probably not backpropagation as we know it, but it could be a form of approximation of gradient estimation that is very similar to backpropagation. Yoshua Bengio has been working on biologically plausible forms of gradient estimation, so it’s not entirely impossible that the brain does some sort of gradient estimation of some objective function, we just simply don’t know.
MARTIN FORD: What other important topics are you working on at Facebook?
YANN LECUN: We’re working on a lot of fundamental research and questions on machine learning, so things that have more to do with applied mathematics and optimization. We are working on reinforcement learning, and we are also working on something called generative models, which are a form of self-supervised or predictive learning.
MARTIN FORD: Is Facebook working on building systems that can actually carry out a conversation?
YANN LECUN: What I’ve mentioned so far are the fundamental topics of research, but there are a whole bunch of application areas.
Facebook is very active in computer vision, and I think we can claim to have the best computer vision research group in the world. It’s a mature group and there are a lot of really cool activities there. We’re putting quite a lot of work into natural language processing, and that includes translation, summarization, text categorization—figuring out what topic a text talks about, as well as dialog systems. Actually, dialog systems are a very important area of research for virtual assistants, question and answering systems, and so on.
MARTIN FORD: Do you anticipate the creation of an AI that someday could pass the Turing test?
YANN LECUN: It’s going to happen at some point, but the Turing test is not actually an interesting test. In fact, I don’t think a lot of people in the AI field at the moment consider the Turing test to be a good test. It’s too easy to trick it, and to some extent, the Turing test has already been and gone.
We give a lot of importance to language as humans because we are used to discussing intelligent topics with other humans through language. However, language is sort of an epiphenomenon of intelligence, and when I say this, my colleagues who work on natural language processing disagree vehemently!
Look at orangutans, who are essentially almost as smart as we are. They have a huge amount of common sense and very good models of the world, and they can build tools, just like humans. However, they don’t have language, they’re not social animals, and they barely interact with other members of the species outside the non-linguistic mother-and-child interaction. There is a whole component of intelligence that has nothing to do with language, and we are ignoring this if we reduce AI to just satisfying the Turing test.
MARTIN FORD: What is the path to artificial general intelligence and what do we need to overcome to get there?
YANN LECUN: There are probably other problems that we do not see at the moment that we’re going to eventually encounter, but one thing I think we’ll need to figure out is the ability that babies and animals have to learn how the world works by observation in the first few days, weeks, and months of life.
In that time, you learn that the world is three-dimensional. You learn that there are objects that move in front of others in different ways when you move your head. You learn object permanence, so you learn that when an object is hidden behind another one, it’s still there. As time goes on, you learn about gravity, inertia, and rigidity—very basic concepts that are learnt essentially by observation.
Babies don’t have a huge amount of means to act on the world, but they observe a lot, and they learn a huge amount by observing. Baby animals also do this. They probably have more hardwired stuff, but it’s very similar.
Until we figure out how to do this unsupervised/self-supervised/predictive learning, we’re not going to make significant progress because I think that’s the key to learning enough background knowledge about the world so that common sense will emerge. That’s the main hurdle. There are more technical subproblems of this that I can’t get into, like prediction under uncertainty, but that’s the main thing.
How long is it going to take before we figure out a way to train machines so that they learn how the world works by watching YouTube videos? That’s not entirely clear. We could have a breakthrough in two years that might take another 10 years to actually make it work, or it might take 10 or 20 years. I have no idea when it will happen, but I do know it has to happen.
That’s just the first mountain we have to climb, and we don’t know how many mountains are behind it. There might be other huge issues and major questions that we do not see yet because we haven’t been there yet and it’s unexplored territory.
It will probably take 10 years before we find this kind of breakthrough and before it has some consequence in the real world, and that has to happen way before we reach human-level artificial general intelligence. The question is, once we clear this hurdle, what other problems are going to pop up?
How much prior structure do we need to build into those systems for them to actually work appropriately and be stable, and for them to have intrinsic motivations so that they behave properly around humans? There’s a whole lot of problems that will absolutely pop up, so AGI might take 50 years, it might take 100 years, I’m not too sure.
MARTIN FORD: But you think it’s achievable?
YANN LECUN: Oh, definitely.
MARTIN FORD: Do you think it’s inevitable?
YANN LECUN: Yes, there’s no question about that.
MARTIN FORD: When you think of an AGI, would it be conscious, or could it be a zombie with no conscious experience at all?
YANN LECUN: We don’t know what that means. We have no idea what consciousness is. I think it’s a non-problem. It’s one of those questions that in the end, when you realize how things actually work, you realize that question was immaterial.
Back in the 17th century when people figured out that the image in the back of the eye on the retina forms upside down, they were puzzled by the fact that we see right-side up. When you understand what kind of processing is required after this, and that it doesn’t really matter in which order the pixels come, you realize it’s kind of a funny question because it doesn’t make any sense. It’s the same thing here. I think consciousness is a subjective experience and it could be a very simple epiphenomenon of being smart.
There are several hypotheses for what causes this illusion of consciousness—because I think it is an illusion. One possibility is that we have essentially a single engine in our prefrontal cortex that allows us to model the world, and a conscious decision to pay attention to a particular situation configures that model of the world for the situation at hand.
The conscious state is sort of an important form of attention, if you will. We may not have the same conscious experience if our brain were ten times the size and we didn’t have a single engine to model the world, but a whole bunch of them.
MARTIN FORD: Let’s talk about some of the risks associated with AI. Do you believe that we’re on the cusp of a big economic disruption with the potential for wide spread job losses?
YANN LECUN: I’m not an economist, but I’m obviously interested in those questions, too. I’ve talked to a bunch of economists, and I’ve attended a number of conferences with a whole bunch of very famous economists who were discussing those very questions. First of all, what they say is that AI is what they call a general-purpose technology or GPT for short. What that means is that it’s a piece of technology that will diffuse into all corners of the economy and transform pretty much how we do everything. I’m not saying this; they are saying this. If I was saying this, I would sound self-serving or arrogant, and I would not repeat it unless I had heard it from other people who know what they’re talking about. So, they’re saying this, and I didn’t really realize that this was the case before I heard them say it. They say this is something on the scale of electricity, the steam engine, or the electric motor.
One thing I’m worried about, and this was before talking to the economists, is the problem of technological unemployment. The idea that technology progresses rapidly and the skills that are required by the new economy are not matched by the skills of the population. A whole proportion of the population suddenly doesn’t have the right skills, and it’s left behind.
You would think that as technological progress accelerates, there’d be more and more people left behind, but what the economists say is that the speed at which a piece of technology disseminates in the economy is actually limited by the proportion of people who are not trained to use it. In other words, the more people are left behind, the less quickly the technology can diffuse in the economy. It’s interesting because it means that the evil has kind of a self-regulating mechanism in it. We’re not going to have widely disseminated AI technology unless a significant proportion of the population is trained to actually take advantage of it, and the example they use to demonstrate this is computer technology.
Computer technology popped up in the 1960s and 1970s but did not have an impact on productivity on the economy until the 1990s because it took that long for people to get familiar with keyboards, mice, etc., and for software and computers to become cheap enough for them to have mass appeal.
MARTIN FORD: I think there is a question of whether this time is different relative to those historical cases, because machines are taking on cognitive capability now.
You now have machines that can learn to do a lot of routine, predictable things, and a significant percentage of our workforce is engaged in things that are predictable. So, I think the disruption could turn out to be bigger this time than what we’ve seen in the past.
YANN LECUN: I don’t actually think that’s the case. I don’t think that we’re going to face mass unemployment because of the appearance of this technology. I think certainly the economic landscape is going to be vastly different in the same way that 100 years ago most of the population were working in the fields, and now it’s 2% of the population.
Certainly, over the next several decades, you’re going to see this kind of shift and people are going to have to retrain for it. We’ll need some form of continuous learning, and it’s not going to be easy for everyone. I don’t believe, though, that we’re going to run out of jobs. I heard an economist say, “We’re not going to run out of jobs because we’re not going to run out of problems.”
The upcoming AI systems are going to be an amplification of human intelligence in the way that mechanical machines have been an amplification of physical strength. They’re not going to be a replacement. It’s not like just because AI systems that analyze MRI images would be better at detecting tumors, then radiologists are out of a job. It’s going to be a very different job, and it’s going to be a much more interesting job. They’re going to spend their time doing more interesting things like talking to patients instead of staring at screens for 8 hours a day.
MARTIN FORD: Not everyone’s a doctor, though. A lot of people are taxi drivers or truck drivers or fast food workers and they may have a harder time transitioning.
YANN LECUN: What’s going to happen is the value of things and services is going to change. Everything that’s by done by machine is going to get a lot cheaper, and anything that’s done by humans is going to get more expensive. We’re going to pay more for authentic human experience, and the stuff that can be done by machine is going to get cheap.
As an example, you can buy a Blu-ray player for $46. If you think about how much incredibly sophisticated technology goes into a Blu-ray player, it’s insane that it costs $46. It’s got technology in the form of blue lasers that didn’t exist 20 years ago. It’s got an incredibly precise servo mechanism to drive the laser to microns of precision. It’s also got, H.264 video compression and superfast processors. It has a ridiculous amount of technology that goes in there, and it’s $46 because it’s essentially mass-produced by machines. Now, go on the web and search for a handmade ceramic salad bowl, and the first couple of hits you’re going to get are going to propose handmade ceramic bowl, a 10,000-year-old technology, for something in the region of $500. Why $500? Because it’s handmade and you’re paying for the human experience and the human connection. You can download a piece of music for a buck, but then if you want to go to a show where that music is being played live, it’s going to be $200. That’s for human experience.
The value of things is going to change, with more value placed on human experience and less to things that are automated. A taxi ride is going to be cheap because it can be driven by the AI system, but a restaurant where an actual person serves you or an actual human cook creates something, is going to be more expensive.
MARTIN FORD: That does presume that everyone’s got a skill or talent that’s marketable, which I’m not sure is true. What do you think of the idea of a universal basic income as a way to adapt to these changes?
YANN LECUN: I’m not an economist, so I don’t have an informed opinion on this, but every economist I talked to seemed against the idea of a universal basic income. They all agree with the fact that as a result of increased inequality brought about by technological progress, some measures have to be taken by governments to compensate. All of them believe this has to do with fiscal policy in the form of taxing, and wealth and income redistribution.
This income inequality is something that is particularly apparent in the US, but also to a smaller scale in Western Europe. The Gini index—a measure of income inequality—of France or Scandinavia is around 25 or 30. In the US, it’s 45, and that’s the same level as third-world countries. In the US, Erik Brynjolfsson, an economist at MIT, wrote a couple of books with his colleague from MIT, Andrew McAfee, studying the impact of technology on the economy. They say that the median income of a household in America has been flat since the 1980s where we had Reaganomics and the lowering of taxes for higher incomes, whereas productivity has gone up more or less continuously. None of that occurred in Western Europe. So, it’s purely down to fiscal policy. It’s maybe fueled by technological progress, but there are easy things that governments can do to compensate for the disruption, and they’re just not doing it in the US.
MARTIN FORD: What other risks are there, beyond the impact on the job market and economy, that come coupled with AI?
YANN LECUN: Let me start with one thing we should not worry about, the Terminator scenario. This idea that somehow we’ll come up with the secret to artificial general intelligence, and that we’ll create a human-level intelligence that will escape our control and all of a sudden robots will want to take over the world. The desire to take over the world is not correlated with intelligence, it’s correlated with testosterone.
We have a lot of examples today in American politics, clearly illustrating that the desire for power is not correlated with intelligence.
MARTIN FORD: There is a pretty reasoned argument, though, that Nick Bostrom, in particular, has raised. The problem is not an innate need to take over the world, but rather that an AI could be given a goal and then it might decide to pursue that goal in a way that turns out to be harmful to us.
YANN LECUN: So, somehow we’re smart enough to build artificial general intelligence machines, then the first thing we do is tell them to build as many paper clips as they can and they turn the entire universe into paper clips? That sounds unrealistic to me.
MARTIN FORD: I think Nick intends that as kind of a cartoonish example. Those kinds of scenarios all seem far-fetched, but if you are truly talking about superintelligence, then you would have a machine that might act in ways that would be incomprehensible to us.
YANN LECUN: Well, there is the issue of objective function design. All of those scenarios assume that somehow, you’re going to design the objective function—the intrinsic motivations—of those machines in advance, and that if you get it wrong, they’re going to do crazy things. That’s not the way humans are built. Our intrinsic objective functions are not hardwired. A piece of it is hardwired in a sense that we have the instinct to eat, breathe, and reproduce, but a lot of our behavior and value system is learned.
We can very much do the same with machines, where their value system is going to be trained and we’re going to train them to essentially behave in society and be beneficial to humanity. It’s not just a problem of designing those functions but also training them, and it’s much easier to train an entity to behave. We do it with our kids to educate them in what’s right and wrong, and if we know how to do it with kids why wouldn’t we be able to do this with robots or AI systems?
Clearly, there are issues there, but it’s a bit like we haven’t invented the internal combustion engine yet and we are already worrying that we’re not going to be able to invent the brake and the safety belt. The problem of inventing the internal combustion engine is considerably more complicated than inventing brakes and safety belts.
MARTIN FORD: What do you think of the fast takeoff scenario, where you have recursive improvement that happens at an extraordinary rate, and before you know it, we’ve got something that makes us look like a mouse or an insect in comparison?
YANN LECUN: I absolutely do not believe in that. Clearly there’s going to be continuous improvement, and certainly, the more intelligent machines become, the more they’re going to help us design the next generation. It’s already the case, and it’s going to accelerate.
There is some sort of differential equation that governs the progress of technology, the economy, consumption of resources, communication, the sophistication of technology, and all that stuff. There’s a whole bunch of friction terms in this equation that is completely ignored by the proponent of singularity or fast takeoff. Every physical process at some point has to saturate, by exhausting resources if nothing else. So, I don’t believe in a fast takeoff. It’s a fallacy that someone will figure out the secret to AGI, then all of a sudden, we’re going to go from machines that are as intelligent as a rat to some that are as intelligent as an orangutan, and then a week later they are more intelligent than us, and then a month later, way more intelligent.
There’s also no reason necessarily to believe that being way more intelligent than a single human will allow a machine to be completely superior to a single human. Humans can get killed by viruses that are extremely stupid, but they are specialized to kill us.
If we can build an artificial intelligence system that has general intelligence in that sense, then we can probably also build a more specialized intelligence designed to destroy the first one. It would be much more efficient at killing the AGI because more specialized machines are more efficient than general ones. I just think that every issue has its own solution built in.
MARTIN FORD: So, what should we legitimately be worried about in the next decade or two?
YANN LECUN: Economic disruption is clearly an issue. It’s not an issue without a solution, but it’s an issue with considerable political obstacles, particularly in cultures like the US where income and wealth redistribution are not something that’s culturally accepted. There is an issue of disseminating the technology so that it doesn’t only profit the developed world, but it’s shared across the world.
There is a concentration of power. Currently, AI research is very public and open, but it’s widely deployed by a relatively small number of companies at the moment. It’s going to take a while before it’s used by a wider swath of the economy and that’s a redistribution of the cards of power. That will affect the world in some ways, it may be positive but it may also be negative, and we need to ensure that it’s positive.
I think the acceleration of technological progress and the emergence of AI is going to prompt governments to invest more massively into education, particularly continuous education because people are going to have to learn new jobs. That’s a real aspect of the disruption that needs to be dealt with. It’s not something that doesn’t have a solution, it’s just a problem that people have to realize exists in order for them to solve it.
If you have a government that doesn’t even believe in established scientific facts like global warming, how can they believe in this kind of stuff? There are a lot of issues of this type, including ones in the area of bias and equity. If we use supervised learning to train our systems, they’re going to reflect the biases that are in the data, so how can you make sure they don’t prolong the status quo in terms of biases?
MARTIN FORD: The problem there is that the biases are encapsulated in the data so that a machine learning algorithm would naturally acquire them. One would hope that it might be much easier to fix bias in an algorithm than in a human.
YANN LECUN: Absolutely. I’m actually quite optimistic in that dimension because I think it would indeed be a lot easier to reduce bias in a machine than it currently is with people. People are biased in ways that are extremely difficult to fix.
MARTIN FORD: Do you worry about military applications, like autonomous weapons?
YANN LECUN: Yes and no. Yes, because of course AI technology can be used for building weapons, but some people, like Stuart Russell, have characterized a potential new generation of AI-powered weapons as weapons of mass destruction and I completely disagree with that.
I think the way that militaries are going to use AI technology is exactly the opposite. It’s for what the military calls, surgical actions. You don’t drop a bomb that destroys an entire building, you send in your drone that just puts the person you are interested in capturing to sleep; it could be non-lethal.
When it gets to that point, it makes the military look more like police. Is that good in the long term? I don’t think anyone can guess. It’s less destructive than nukes—it can’t be more destructive than nukes!
MARTIN FORD: Do you worry about a race with China in terms of advancing artificial intelligence? They have over a billion people, so they have got more data and along with that, fewer constraints on privacy. Is that going to give them an advantage in moving forward?
YANN LECUN: I don’t think so. I think currently progress in the science is not conditioned on the wide availability of data. There may be more than 1 billion people in China, but the proportion of people who are actually involved in technology and research is actually relatively small.
There’s no question that it will grow, China is really progressing in that direction. I think the style of government and the type of education they have may be stifling for creativity after a while. There is good work coming out of China, though, with some very smart people there, and they’re going to make contributions to this field.
There was the same kind of fear of the West being overrun by Japanese technology in the 1980s, and it happened for a while and then it kind of saturated. Then it was the Koreans, and now it’s the Chinese. There are going to be big mutations in Chinese society that will have to happen over the next few decades that will probably change the situation completely.
MARTIN FORD: Do you think that AI needs to be regulated at some level? Is there a place for government regulation for the kind of research you’re doing and the systems that you’re building?
YANN LECUN: While I don’t think there is any point in regulating AI research at the moment, I do think there is certainly a need for regulating applications. Not because they use AI, but because of the domain of applications that they are.
Take the use of AI in the context of drug design; you always want to regulate how drugs are being tested, how they are deployed, and how they are used. It’s already the case. Take self-driving cars: cars are regulated, and there are strict road safety regulations. Certainly, those are application areas where existing regulations might need to be tweaked because AI is going to become preponderant.
However, I don’t see any need for the regulation of AI at the moment.
MARTIN FORD: So, I assume you disagree quite strongly with the kind of rhetoric Elon Musk has been using?
YANN LECUN: Oh, I completely and absolutely disagree with him. I’ve talked to him several times, but I don’t know where his views are coming from. He’s a very smart guy and I’m in awe of some of his projects, but I’m not sure what his motivation is. He wants to save humanity, so maybe he needs another existential threat for it. I think he is genuinely worried, but none of us have been able to convince him that Bostrom-style, hard take-off scenarios are not going to happen.
MARTIN FORD: Are you an optimist overall? Do you believe that the benefits of AI are going to outweigh the downsides?
YANN LECUN: Yes, I would agree with that.
MARTIN FORD: In what areas do you think it will bring the most benefits?
YANN LECUN: Well, I really hope that we figure out the way to get machines to learn like baby humans and animals. That’s my scientific program for the next few years. I also hope we’re going to make some convincing breakthrough before the people funding all this research get tired, because that’s what happened in previous decades.
MARTIN FORD: You’ve warned that AI is being overhyped and that this might even lead to another “AI Winter.” Do you really think there’s a risk of that? Deep learning has become so central to the business models of Google, Facebook, Amazon, Tencent, and all these other incredibly wealthy corporations. So, it seems hard to imagine that investment in the technology would fall off dramatically.
YANN LECUN: I don’t think we’re going to see an AI winter in the way we saw before because there is a big industry around it and there are real applications that are bringing real revenue to these companies.
There’s still a huge amount of investment, with the hope that, for example, self-driving cars are going to be working in the next five years and that medical imaging is going to be radically revolutionized. Those are probably going to be the most visible effects over the next few years, medicine and health care, transportation, and information access.
Virtual assistants are another case. They are only mildly useful today because they’re kind of scripted by hand. They don’t have any common sense, and they don’t really understand what you tell them at a deep level. The question is whether we need to solve the AGI problem before we get virtual assistants that are not frustrating, or whether we can make more continuous progress before that. Right now, I don’t know.
When that becomes available, though, that’s going to change a lot of how people interact with each other and how people interact with the digital world. If everyone has a personal assistant that has human-level intelligence, that’s going to make a huge difference.
I don’t know if you’ve seen the movie Her? It’s not a bad depiction in some ways of what might happen. Among all the sci-fi movies on AI, it’s probably one of the least ridiculous.
I think a lot of AI-related technology is going to be widely available in the hands of people because of hardware progress. There’s a lot of effort now to develop low-power and cheap hardware that can fit in your smartphone or your vacuum cleaner that can run a convolutional network on 100 milliwatts of power, and the chip can be bought for 3 bucks. That’s going to change a lot of how the world around us works.
Instead of going randomly around your room, your vacuum cleaner is now going to be able to see where it needs to go, and your lawnmower is going to be able to mow your lawn without running over your flowerbeds. It’s not just your car that will drive itself.
It might also have interesting environmental consequences, like wildlife monitoring. AI is going to be in the hands of everyone because of progress in hardware technology that is specialized for deep learning, and that’s coming in the next 2 or 3 years.
YANN LECUN is a Vice President and Chief AI Scientist at Facebook, as well as a professor of computer science at New York University. Along with Geoff Hinton and Yoshua Bengio, Yann is part of the so-called “Canadian Mafia”—the trio of researchers whose effort and persistence led directly to the current revolution in deep learning neural networks.
Prior to joining Facebook, Yann worked at AT&T’s Bell Labs, where he is credited with developing convolutional neural networks—a machine learning architecture inspired by the brain’s visual cortex. Yann used convolutional neural nets to develop a handwriting recognition system that became widely used in ATMs and at banks to read the information on checks. In recent years, deep convolutional nets, powered by ever faster computer hardware, have revolutionized computer image recognition and analysis.
Yann received an Electrical Engineer Diploma from Ecole Superieure d’Ingenieurs en Electrotechnique et Electronique (ESIEE) in Paris, and a PhD in Computer Science from Universite Pierre et Marie Curie in 1987. He later worked as a post-doctoral researcher in Geoff Hinton’s lab at the University of Toronto. He joined Facebook in 2013 to establish and run the Facebook AI Research (FAIR) organization, headquartered in New York City.