How AI learns Like the Human Brain

How do you learn new, rich concepts from very little data? To date, machine learning has been primarily about pattern recognition. How can machine learning go beyond merely recognizing patterns, to explaining what is behind those patterns?

Watch Josh Tenenbaum, Guest Expert in the Machine Learning in Business online short course from MIT, discuss how machines can learn like the human brain.

Transcript

Today’s machine learning is basically about pattern recognition, and that means deep learning and other ML technologies. But human intelligence is so much more than pattern recognition.

In my work, I like to understand all the ways in which human intelligence models the world. So that means explaining and understanding what we see, not just finding patterns in data. In the last few years, we in our group have been working on an important, what I’d call, ‘warm-up problem’. It’s not yet the problem of learning, say, an intuitive physics engine, but it’s the problem of learning new concepts, rich concepts from very little data, which has the same form.

So to illustrate what I mean by this, think about the first time you encountered one of these segways, right? That personal transportation device. You’d never seen anything like this before, until you saw one going down the street with somebody on it. And yet you could get that concept from just one example.

You could look at other vehicles, including familiar ones and unfamiliar ones, and figure out which ones were of the same type. Now, in case you forgot what that was like, think about this: unless you are a rock climber, you probably don’t know what this thing is on the top. It’s called a cam and it’s a piece of specialized rock climbing equipment. But even if you’ve just seen this for the first time, you can look at the objects down below and recognize that some of them are probably also cams. Consider, again, they vary in all sorts of ways. They’re different shapes, sizes; they’re in different orientations, different colors, and yet you’ll have no trouble recognizing that those could be the same kind of thing as on the top.

For now, and over the last few years, we’ve been working on the simplest version of this, which is to try to do one-shot learning in the context of handwritten characters. If you take machine learning at MIT or in many more technical classes, one of the first data sets that you meet is what’s called the MNIST data set. It’s a famous data set that was put together a few decades ago of handwritten digits — zero, one, two, three, and so on up to nine. It was used to build some of the first machine pattern recognition systems for reading checks in ATM readers or reading zip codes at the post office for automatic letter routing.

We built a data set that, in some sense, is similar to this. It’s also just a bunch of binary, small binary pixel images. But instead of being the digits zero through nine, it consists of characters in many of the world’s writing systems. It’s called ‘omniglot’. In this data set, what you can see is one instance each of a number of characters in a number of alphabets, and it’s just a very small fraction of the data set. Even though you don’t know any of these languages and alphabets yourself, you can recognize that each character is its own thing and different from the ones next to it. How are you able to do that? You can also do one-shot learning in this domain.

What’s next for machine learning?

What’s going to be the next big idea in machine learning, inspired by something that the human brain and mind does, that could turn into the future of machine learning. If you think about what would it take to capture all the different kinds of programs that a human being comes to know in their life, we have to think about all the different activities that you do when you’re programming and this represents, to me, the vision of what the future of machine learning looks like. It’s the idea of learning as programming.

Looking forward, what I’m most excited about is really being able to take on what is truly, I think AI’s oldest dream and best vision for how we might make machines that really achieve something like human intelligence. It’s the idea of machines that start like a baby and learn like a child; that grow into intelligence the way a human being does. This idea was first proposed by Alan Turing and his famous paper on the Turing Test, and championed by really all of the major AI pioneers since, for good reason, because a human child is the only scaling route in the known universe we know that actually reliably, robustly grows into intelligence. And I think only now are we really in a position to take this seriously as an actual vision for machine learning.