Breaking it down: A Q&A on machine learning
For many of us, machine learning seems pretty futuristic. Recently, though, it’s been appearing more and more in our lives – whether it’s a Google computer playing an amazing game of Go, or Inbox by Gmail creating auto replies. And while that’s all exciting, some of us are still wondering what exactly machine learning is. Or why it matters. Or why identifying a dog in a photo isn’t as easy as it sounds. So we sat down with Maya Gupta, research scientist for machine learning at Google, to break it down.
Let’s start with the basics. What exactly is machine learning?
Machine learning takes a bunch of examples, works out patterns that explain the examples, then uses those patterns to make predictions about new examples.
Take movie recommendations, for example. Say a billion people each tell us their ten favourite movies. That’s a bunch of examples the computer can use to learn what movies that people like have in common. Then the computer comes up with patterns to explain those examples such as, for example, 'People who like horror movies don’t usually like romances, but people do like movies with the same actors in them.' Then if you tell the computer that you liked The Shining with Jack Nicholson, it can make a good guess about whether you’d like the romantic comedy Something’s Gotta Give with Jack Nicholson, and which other videos to recommend to you on YouTube.
Got it. Sort of. How does that work in practice, though?
In practice, the patterns that the machine learns can be very complicated and hard to explain in words. Consider Google Photos, which lets you search your photos to find pictures with dogs. How does Google do that? Well, first we get a bunch of examples of photos labelled 'dog' (thanks Internet!). We also get a bunch of photos labelled 'cat', and photos with about a million other labels, but I won’t list them all here :).
Then the computer looks for patterns of pixels and patterns of colours that help it guess if it’s a cat or dog (or…). First, it just makes a random guess at what good patterns might be to identify dogs. Then it looks at an example dog image, and sees if its current patterns get it right. If it’s mistakenly calling a cat a dog, then it makes some tiny adjustments to the patterns it’s using. Then it looks at a cat image, and again tweaks its patterns to try to get that one right. And it repeats this about a billion times: look at an example, and if it’s not getting it right, tweak the patterns it’s using to do a better job on that one example.
In the end, the patterns form a machine-learned model, such as a deep neural network, which can (mostly) correctly identify dogs and cats and fire fighters and many, many other things.
That sounds pretty futuristic. What are some of the other Google products that use machine learning today?
There’s a whole host of new things Google is doing with machine learning; for example, Google Translate can take a photo of a street sign or menu in one language, work out the words and language that are in the photo, and magically translate it in real-time into your language.
You can also say just about anything to Translate and machine-learned speech recognition will kick in. Speech recognition is used in a bunch of other products as well, such as working out your voice queries for the Google app, and making YouTube videos more searchable.
Is machine learning the same thing as artificial intelligence?
Well, actually, these words can mean different things to different people, but essentially, artificial intelligence (AI) is a loose term for computer programs that try to solve the kind of problems that humans find easy, like telling a story about what’s happening in a picture. One of the cool things that humans also do easily is learn from examples. And that’s what machine learning programs try to do, too: teach computers to learn from examples.
The cool thing is when we work out how to make these computer programs, we can sometimes scale them up to handle a lot of data really fast, and then we can solve really hard problems like mastering Go, routing everyone through traffic simultaneously, optimising energy usage nationwide and, of course, my favourite, finding you the best Search results on Google.
So, why is Google making such a big deal about machine learning now?
Machine learning is not brand new, and has its roots in 18th century statistics. But you’re right it has really heated up lately for three reasons.
First off, we need a huge number of examples to teach computers how to make good predictions, even about stuff you or I would find easy (like finding a dog in a photo). With all the activity on the Internet, we’ve now got a rich source of examples that computers can learn from. For example, there are now millions of dog photos labelled as 'dog' on websites around the world, in every language.
But it’s not enough to have a lot of examples. You can’t just show a bunch of photos of dogs to a webcam and expect it to learn anything – the computer needs a learning program. And lately the field (and Google) has made some exciting breakthroughs in how complicated and powerful those machine learning programs can be.
However, our programs are still not perfect, and computers are still pretty dumb, so we have to see a lot of examples a bunch of times to tweak a lot of digital knobs to get it right. That all takes a huge amount of computing power, and fancy parallelised processing. But new software and hardware advancements have made that possible, too.
What’s one thing computers can’t do today, but will be able to do soon, thanks to machine learning?
Practically yesterday, speech recognition struggled to recognise just ten different digits when you read your credit card number over the phone. Speech recognition has made incredible advances in the last five years using sophisticated machine learning, and now you can use it to issue Google searches. And it’s getting even better, fast.
I think machine learning is even going to make us all look better. I don’t know about you, but I hate trying on clothes! I find a brand of jeans that works, and I buy five of them. But machine learning can turn examples of which brands fit us into recommendations for what else should fit us. That problem’s a little outside of Google’s scope, but I hope someone’s working on it!
What does machine learning look like in 10 years?
One thing the whole field is working on is how to learn faster from fewer examples. One approach to that (that Google is working particularly hard on) is giving our machines more common sense, which in the field we call 'regularisation.'
What does common sense look like to a machine? Well one thing it means is that, in general, if an example only changes a little, the machine shouldn’t totally change its mind. For example, a photo of a dog with a cowboy hat is still a dog.
We enforce this kind of common sense in the learning program by making the machine learning insensitive to small, unimportant changes, like a cowboy hat. While that’s easy to say, if you do it wrong, you make the machine not sensitive enough to important changes! So this is a balancing act we’re still trying to work out.
What excites you most about machine learning? What motivates you to work on it?
I grew up in Seattle, where we learned a lot about early explorers of the American West such as Lewis and Clark. Machine learning research has that same spirit of exploration – we’re seeing things for the first time, and trying to map out a path to a great future.
If you could give machine learning at Google a bumper sticker slogan, what would it be?
If at first you don’t succeed, try another billion times.