What excites you most about artificial intelligence?

For many of us, machine learning is still a futuristic vision. Recently, however, it has also become more and more common in our lives - whether in the form of a Google software program that plays an impressive game of Go, or through the automatic replies from Inbox by Gmail. While this is all exciting, some of us still wonder what exactly machine learning is. Or why it matters. Or why automatically recognizing a dog in a photo isn't as easy as it sounds. That's why we asked Maya Gupta, a researcher at Google.

Let's start with the basics. What exactly is machine learning?

Machine learning analyzes numerous examples, finds patterns that explain those examples, and then uses those patterns to make predictions about new examples.

An example would be movie recommendations. Let's say a billion people tell us their ten favorite films. That's quite a few examples that the computer can use to find out what these films have in common. The computer then creates templates to explain these examples, such as, "People who like horror movies usually don't like romance movies, but people like movies that feature the same actors." Then, when you tell the computer that you enjoyed The Shining, starring Jack Nicholson, the computer can guess you like the romantic comedy "What Your Heart Desires" starring Jack Nicholson and recommend videos to you on YouTube.

I understand. Something like that, at least. But how does it work in practice?

In practice, the patterns the computer learns are very complicated and difficult to explain. Think Google Photos. It can be used to search your photos to find pictures with dogs. How does Google do it? Well, first of all we get - thanks to the internet - countless photos with the name "dog". We also receive a number of photos with the name "cat" and photos with countless other names, but I won't list them all here :).

Then the computer looks for patterns of pixels and colors to find out if it's a cat or a dog or something else. At first, only a vague guess is made as to which patterns might be good for recognizing dogs. Then the computer looks at an example of a dog's picture and judges whether its current patterns apply to it. If he mistakenly mistook a cat for a dog, he tweaked the patterns used a little. Then he looks at a picture of a cat and optimizes his patterns again to come up with the correct answer. And he repeats that about a billion times: look at an example, and if he's wrong, tweak the patterns used to get that one example better.

In the end, the patterns form a machine-learned model, similar to a complex neural network that dogs, cats and many, many other things can (mostly) correctly recognize.

That sounds pretty futuristic. What other Google products are already leveraging machine learning?

There are a whole host of new things that Google is using machine learning to do. This is how Google Translate can take a photo of a street sign or menu in one language, recognize what words and language are in the photo, and magically translate them into your language in real time.

You can also say just about anything to Google Translate and machine-learned speech recognition will kick off. Speech recognition is also used in many other products, for example to process your voice search in the Google app and to make YouTube videos easier to find.

{% include "www.google.com/about/views/main/machine-learning-qa/translate-demo/index.html"%}

Is machine learning the same as artificial intelligence?

While these terms aren't used consistently, in essence, artificial intelligence (AI) is a general term for computer programs that try to solve the kind of problems that are easy for humans, such as telling a story about what's inside Picture is going on. One thing people can easily do is learn from examples. And that's what you try to achieve with machine learning programs: you want to teach computers to learn from examples.

It's really great when we actually develop such computer programs and manage to expand them to the point where they can process large amounts of data very quickly. We can then use it to solve really difficult problems such as playing Go, guiding all road users through the traffic at the same time, reducing energy consumption across the country and - which is of course my favorite - finding the best search results on Google.

Why is machine learning such a central topic for Google right now?

Machine learning is not entirely new and has its roots in the 18th century. But you are right that the topic has been gaining interest recently. There are three reasons for this.

First, we need a large number of examples to teach computers how to make good predictions - even about things that would be easy for you or me, such as recognizing a dog in a photo. With all of the activity on the Internet, we now have a great source of examples for computers to learn from. For example, there are millions of dog photos with the label "dog" in every language on websites around the world.

But it is not enough to have numerous examples. You can't just show a few photos of dogs to a webcam and expect them to learn something. The computer needs a tutorial. And recently the industry - and Google - made some impressive breakthroughs in how complex and powerful these machine learning programs can be.

However, our programs are not yet perfect, and computers are still pretty simple-minded. Therefore, we have to look at a large number of examples multiple times in order to optimize numerous digital controllers and achieve a correct result. All of this requires enormous computing power and complicated parallel processing. But new advances in software and hardware have also made this possible.

What can computers not do today, but will soon be able to do it thanks to machine learning?

Practically yesterday, voice recognition was struggling to only recognize ten different digits when you read your credit card number over the phone. Speech recognition has made incredible strides over the past five years through sophisticated machine learning techniques. Today you can use it to do Google searches. And all of this is getting better and better, and at a rapid pace.

I think machine learning will even help us look better. I don't know about you, but I hate trying on clothes! If I find a brand of jeans that fits me, I'll buy five of them. But machine learning can use the brands that suit us as examples to make recommendations about what else might suit us. This topic is not entirely on Google's line, but I hope someone is working on such an application!

What will machine learning look like in ten years?

The whole industry is working on how computers can learn faster from fewer examples. One approach that Google is taking particularly hard is to make our machines more common sense. In the industry this is called "regularization".

What does common sense look like in a machine? It means, for example, that if an example changes only slightly, the machine shouldn't completely change its mind. For example, a photo of a dog in a cowboy hat is still a dog.

We implement this kind of common sense in the tutorial by making the machine insensitive to small, insignificant changes, like a cowboy hat. That sounds easy. But if that doesn't work, you don't make the machine sensitive enough to important changes. So this is a balancing act that we are still working on.

What do you find most exciting 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 like Lewis and Clark. The machine learning work is based on the same spirit of research - we see things for the first time and try to find a way into a better future.

If you could give machine learning a slogan on Google, what would it be?

If you can't do it the first time, keep trying.