Yann LeCun, Courant Institute of Mathematical Sciences Silver Professor of Computer Science, Neural Science and Electrical and Computer Engineering and founding director of NYU’s Center for Data Science, was recently named director of Facebook’s laboratory devoted to artificial intelligence and machine learning research. This new laboratory will support AI research teams in New York City, London and Facebook’s Menlo Park, CA headquarters.
A pioneer in the exploding field of artificial intelligence, LeCun co-developed an early version of the back-propagation algorithm, the most popular method for training artificial neural networks. While at AT&T Bell Laboratories, he developed the convolutional network model, whose architecture is somewhat inspired by the visual cortex of animals and humans to create a pattern recognition system for machines. In the area of deep learning, an emerging branch of machine learning in which researchers aim to reproduce the perception process, LeCun has been called “perhaps the world’s best-known deep learning scientist.”[MIT Technology Review]
LeCun will remain at the NYU Center for Data Science and at the Courant Institute, continuing his research and teaching, on a part-time basis. He will step down from his position as director of the Center for Data Science; Raghu Varadhan has been appointed interim director. Yann recently sat down with reporter ML Ball to discuss Facebook, artificial intelligence and the evolutionary usefulness of love.
How did Facebook approach you for this position?
Facebook wanted some advice for how to get into deep learning, so they called me and Yoshua Bengio. I was then put in touch with their CTO, Mike Schroepfer, a very smart guy. They asked me if I was interested in joining Facebook, and I said, “I’m very happy at NYU and I see myself as an academic in the long-term. There are a lot of things I’m working on which I can’t do in industry, like computational neuroscience and robotics, and I am not ready to quit my job. And I don’t want to move from New York.” So they asked me for names of other people who were good, and I gave them a bunch of people whom they could talk to.
Then I had a chat with Mark Zuckerberg over the phone about the fact that he wanted to create a research lab. Again I told him, “I’m not leaving NYU and I’m not leaving New York.” Eventually, on a trip to California, I visited Facebook, and had another chat with Mark Zuckerberg, in person. It was pretty clear that what he wanted to do was start an advanced research lab with very, very ambitious goals. And they were okay with me staying at NYU part-time and not moving from New York.
It made perfect sense, particularly because it would be a partnership between NYU and Facebook, and they would be right next door ― 770 Broadway, just up the street. I’ll be able to do academic-style research here at Courant, and at the same time, lead research projects that are better done in an industry environment. I was a department head at AT&T Labs-Research, so I have quite a bit of experience with managing research organizations in industry. It’s not every day you are given the opportunity of creating a research lab from scratch, which is basically what I’m in charge of doing.
Why did Facebook create a New York lab?
There is a huge pool of talent around New York to work on the problems Facebook is interested in. We have the Center for Data Science here at NYU, the Center for Urban Science and Progress (CUSP), the Institute for Data Sciences and Engineering at Columbia, the Simons Center for Data Analysis, Cornell’s new tech campus…there is a lot of things happening around machine learning and data science in and around New York.
More importantly, there is a pool of people here who have experience with industry research from which we can recruit. It’s the same reason Google has a branch of Google Research in New York. Plus, there a lot of people coming from Europe who will move to New York but won’t move to California. So it’s a combination of things.
Will this change your relationship with NYU, Courant and the Center for Data Science?
I am going to step down from being the director of the Center for Data Science, now that Raghu Varadhan has been appointed interim director. I very much enjoyed creating it and getting it on track. I will still teach, I will keep my lab with my students and postdocs. I will keep a research activity here. I’m probably going to change the focus of the research in my lab to focus more on research and principles, and less on applications.
How will NYU benefit from your new endeavor?
This is creating a lot of interest for CDS. There are people ― students, faculty, postdocs ― who before were kind of hesitant to consider NYU, for various reasons. There is now a bit of a snowball effect in New York with all the initiatives in data science: CDS, the Simons Center and Columbia, as well as Princeton and other universities in the area starting similar programs to CDS. And there is a very active startup scene: Silicon Alley, big industry and telecom industries in the area, pushing data science and AI. Plus, all the big web companies have research labs here: Facebook, Google, Microsoft and Yahoo.
What are your research goals?
There are two problems I want to solve, or have an impact on. I don’t know if we will be able to solve them, but we will try. One is creating intelligent machines. And the other is, understanding human intelligence. Those two things cannot be separated, in the sense that we have to understand the principle underlying intelligence for both. Solving one will help solve the other, in the same way our understanding of aerodynamics was driven by building airplanes, but was inspired by bird flight. This is basically the same idea ― an engineering approach to neuroscience. You don’t understand a complex system until you build one.
Do you feel the solutions are close or far away?
Far away. Building machines as intelligent as humans is far away, there’s no question of this. There has been wave after wave of optimism in AI, where people have discovered what they thought was a solution, only to realize they were driving in the fog and there was a big brick wall further down the road. Right now, we are in a period of elation where we think we have a new angle on image recognition and speech recognition, because of deep learning. If nothing else, there will be a lot of useful short-term and medium-term applications of deep learning that are going to come out. And that is to a large extent what the web companies like Facebook and Google are interested in. But beyond that, they want to build really intelligent machines for the purpose of helping their users.
A big part of what Facebook and Google are trying to do is figure out what is interesting to people. This is where AI comes in. To know what is likely to interest you, Facebook needs to give you a “digital best friend” that can understand what your aspirations, interests and desires are…an intelligent entity that can be your main point of contact to the digital world, in the sense that you can hide yourself behind it, and it can protect you. This is not going to happen in one day. This will take a long time and a lot of effort.
Do you think artificial intelligence will ever equal human intelligence, including emotions?
In science fiction, you see intelligent machines that don’t have emotions. I think that emotions are an integral part of intelligence. I don’t think you can build intelligent machines without emotions.
There are theories about where emotions come from. We have in our brains hardwired circuits to tell us when to feel pleasure and pain and hunger and things like this. It drives us to do things that keep us alive as individuals and as a species, things that our conscious mind has little control over. Also, there are other things built in that give us a sense of relationship with other people―empathy, for instance.
Things like love are related to chemical compounds, like oxytocin, that govern how you attach to people. Those are mechanisms that are established by evolution to essentially preserve our species so that we don’t kill each other all the time, hopefully.
In deep learning, are you trying to build machines that have emotions?
Most of the machines we build now are not intelligent in the sense that they have no notion of pain or reward, but can recognize, say, a car or a boat in a picture. It becomes more complicated when the machine needs to train itself―for instance, whether bumping into a chair is a good thing or not. You have to have some built-in mechanism to tell a robot that when it bumps into something, or when it hurts people, it’s not good. You have to give it goals. And then you have to give it notions that it’s succeeding in the goal or not. Then the system can actually train itself to achieve the goal.
Human learning works this way, too. For example, learning to walk or ride a bike. So if we want robots to learn by themselves, we need to give them this ability. And that will come with emotions, because they will have to predict somehow whether something they might do is going to help them achieve their goal or not. That’s what emotions are: the predictions of future rewards. There is a branch of machine learning called “reinforcement learning” that deals with these issues.
Using reinforcement learning, computers have been able to train themselves to play games such as backgammon and simple video games as well as humans. Using similar techniques, they can also be trained to choose which story to show you on your Facebook newsfeed. The reward signal is a click on the “Like” button.
Are you excited about this opportunity with Facebook?
It’s very exciting because it’s not just a research lab for the purpose of having a research lab. It’s a research lab with a mission. And the mission is to make progress towards AI and produce more intelligent machines that can actually help people.
By ML Ball