Yann LeCun is director of artificial intelligence research at Facebook, and founding director of New York University’s Center for Data Science.
He is the author of more than 180 academic papers, and has created character recognition technology widely used by banks to verify checks.
He is also one of the preeminent developers of so-called deep learning, a dramatic advance in computer-based understanding.
In the long run, he said in this condensed and edited conversation, advanced computing techniques will create digital partners that will accompany us throughout life.
Below, a conversation with Mr. LeCun:
What is the significance of artificial intelligence?
A.I. is how we will make sense of all of the information that will be out there in the digital world. A lot of interaction with each other and with the digital world will come from what you could call “digital companions,” that will help us work through things.
What does that mean for Facebook?
Facebook is in the business of connecting people, giving them the information that is informative, entertaining, necessary even if painful, to help them reach their goals. Based on the amount of posts, pictures and news items someone typically gets, we could show you 2,000 things a day. But people’s time is precious, and we can only show about 100 to 150 things a day. They should be the most useful ones.
To do that efficiently we have to know what is in the content. We do that by labeling images, recognizing faces or classifying text. And we have to know your interests, what you want to do, who your friends are in different situations.
What will that look like in the future?
I’m not saying this is a future product, but a way to think about this if there is an intelligent digital companion that allows you to think about things in a new way, the way you interact with friends, expand your thinking. There will be a single point of control that knows and respects your private information.
This kind of A.I. will come progressively. In some ways, parts of this are already there in the Facebook News Feed, in Apple’s Siri, or Microsoft Cortana. They are shallow now, in the kind of interactions you are having. They are somewhat scripted.
How does deep learning, or the ability for a machine to figure out things on its own, work?
Deep learning enables complex interactions. The machine has to go back through different levels of problem solving and think of the step it took when things became different. When you are playing chess, at some point you make a mistake, you may go back several steps that were “correct” to find the one that was wrong. When you fall off a bicycle, you think of when you lost your balance. Deep learning does that. The credit assignment in a deep learning exercise can be tens, even hundreds, of levels deep.
You make it sound so easy.
We have a way of representing objects by what we call vectors, which are long strings of numbers. The vector for “cat” is similar to the vector for “dog,” so that is a close relationship. It would be further apart for a symbolic object. The system seeks associations to determine meaning.
How do you create a system that has good vectors, that is, can chose to represent a word differently, depending on the context? How do you make it learn rules of language, for example?
Tomas Mikolov, who did graduate work and was at Google before he came here, has done a lot of work on vector-based language relations, what we call “word to vec.” You show a sequence of words, one leading logically to another over, say, 11 words. You can ask it what the word in the middle is, and it will predict it. In doing so the system is learning to represent individual words to overall meanings.
We are working on a vector representation of a language that you can use for another language. That involves figuring out how to make a text meaningful for another person.
You could work anywhere. Why Facebook?
Solving A.I. will require contributions from the tech industry, academics and the government. And it has to be done in the open. There are very few companies that can do this work.
Apple is completely secretive, so it’s not a good place to try and do this. Google is partly secretive. They have a culture of hybrid research, alongside engineers and don’t say much. Google X is a secretive research modell you can’t expect a breakthrough from there. Deep Mind, Google’s A.I. work in the U.K., is more open, so I’m more optimistic about the model there.
Facebook has a culture of openness in its DNA. A lot of our software and hardware are open. It sees itself as a fundamentally open company, and in the business of connecting people. We release a lot of code on open source, publish a lot on what we do.
How we will deploy A.I. commercially, we won’t talk about that. It’s not important to our research, and we have to keep a competitive advantage.
What is the long-term goal?
You can’t have intelligence without motivations or emotions.
You can’t render emotions in software.
O.K., lets talk about emotions. People have lots of mental representations of the world that they learn. We are prediction machines, and we change the world to be in a state that we like. What are emotions, but registers of things we like or don’t like? You could assign values to these.
Right now, you eat, reproduce, avoid pain, to have or avoid outcomes. You think, “If I don’t go to school, then life will be painful,” so you go. We make these predictions, and yes, a lot of the time we are conflicted. There is no reason to think we can’t encode this in a machine.
What is the big challenge to proving this is possible?
There are major conceptual advances that we still don’t know how to make. One of the biggest is how we do unsupervised learning.
Supervised learning is like when you train a computer to recognize images of dogs or cars. Reinforcement learning is where you don’t tell the machine the correct answer. Instead, you just score performance. The machine figures out the rules by figuring out where it made a mistake.
Unsupervised learning is what humans do a lot: learn about how the world works by having an interest in things. A baby learns that when you put a toy behind a box, the toy is still in the world. Humans and animals have that capacity. Contrast that with machines, where most learning is still supervised. We don’t have a good grand model yet.
“Vec to word” is perhaps a kind of unsupervised learning. We are still missing a basic principle of what unsupervised learning should be built upon.