Dr. Gary Marcus, the director of NYU's Center for Language and Music, is a professor of psychology and one of the world's most respected cognitive scientists. Recently, he joined Uber (through the acquisition of his company, Geometric Intelligence) and will lead the transportation company's research arm, Uber AI Labs.
Marcus is a brilliant academic who found himself in the eye of a technological hurricane. It turns out that his research into how children learn can help inform how we might teach computers to learn. The past few years have seen frenzied activity around and investment in "AI," or artificial intelligence (which by the way, happens to comes in a variety of flavors). Technology startups throw the terms "machine learning" and "deep learning" around with a cockiness that makes experts in the field of artifical intelligence snicker. Seemingly overnight, everyone's in AI and dozens of companies are marketing their "smart products," but if this is so, then how difficult can it be? HUMAGNA sat down with Dr. Marcus to get a better idea:
HUMAGNA: How does your background in cognitive science help inform your work in AI?
MARCUS: Natural and artificial intelligence are terrific counterpoints to one another. I actually started with AI, as a teenager, before I went into the cognitive sciences, in college, and I love making the return trip now. The human cognitive sciences are a meditation on how one pretty amazing cognitive creature does some pretty amazing feats, but it is also good to realize that our psychology is only one possibility among many. Thinking about machines, which are a little like alien intelligences, is a great place for getting some perspective. Conversely, AI is still pretty mediocre at a lot of things, like understanding language, and I often to look to humans, especially the tiny ones, known as infants and toddlers, for insight. The biggest open question I think is how small children learn so much from so little data; AI has turned towards brute force. Kids are doing something more subtle and more sophisticated; when we figure out how kids do that we may come a long way in AI.
HUMAGNA: Does it frustrate you when people talk about AI, but they are just slapping those two letters on something that is clearly not AI?
MARCUS: No, I don't care that much; the term definitely does get overused, but what really frustrates me is hype that oversimplifies just how hard AI really is.
HUMAGNA: Is there a general lack of understanding about what AI can do and what it can’t?
MARCUS: Yes, it is very hard for outsiders who haven't actually played with real algorithms and real data to have a sense of what is and isn't feasible with current technology - the details matter a lot -- and the state of the art is also constantly changing. I think there is fair a bit of consensus among experts, but it's hard to convey that expertise to people have never even got their toe wet in AI. I wrote a lot about the scope and limits of AI for the New Yorker [link: New Yorker article] a couple years ago, but launching my startup has taken a lot of my time of late. Perhaps some day I will get to write again! The piece I wrote recently for Backchannel on DeepMind and Go [link: Backchannel article] gives a few pointers.
HUMAGNA: What aspects of AI interest you in particular? Is there an important milestone in particular you expect to be reached in the next year or two?
MARCUS: The single biggest milestone that I see on the horizon is the point at which machines can read arbitrary "unstructured" text, like books and articles and for that matter interviews like these. But I don't think we are close to that. There will be some neat stuff in the next couple years, but major advances may be too much to hope for.
HUMAGNA: Congratulations on your new position at Uber! What will you be doing there?
MARCUS:I will be leading a team called Uber AI Labs; one of the most exciting problems is self-driving cars, which could ultimately save hundreds of thousands of lives around the globe annually, and also greatly help reduce energy consumption, and maybe have some other profound implications we are just starting to thinking about. It's no secret that we are interested autonomous flight, too, which is super exciting. The team will start with the 15 folks that I brought together at Geometric Intelligence, but we plan to grow rapidly, and we are definitely looking to hire smart people in all areas of artificial intelligence and machine learning.
HUMAGNA: Favorite AI movie…?
MARCUS: "Her," if I had to pick one, because it might really capture a reality from 40 - 50 years hence. "The Matrix" has some pretty amazing AI going on the background, and really it's just plain hard to pick. I've also really been enjoying "Black Mirror", which has little bits of AI in the background, much of it very plausible, if not quite ready yet. And it does, as the name suggest, make us reflect on where we're headed.