Subtitles section Play video Print subtitles Chris Anderson: Help us understand what machine learning is, because that seems to be the key driver around artificial intelligence. How does machine learning work? Sebastian Thrun: So, artificial intelligence and machine learning is about 60 years old and has not had a great day in its past until recently. And the reason is that today, we have reached a scale of computing and datasets that was necessary to make machines smart. So here's how it works: If you program a computer today, say, your phone, then you hire software engineers that write a very, very long kitchen recipe, like, "If the water is too hot, turn down the temperature. If it's too cold, turn up the temperature." The recipes are not just 10 lines long. They are millions of lines long. A modern cell phone has 12 million lines of code. A browser has five million lines of code. And each bug in this recipe can cause your computer to crash. That's why a software engineer makes so much money. The new thing now is that computers can find their own rules. So instead of an expert deciphering, step by step, a rule for every contingency, what you do now is you give the computer examples and have it infer its own rules. A really good example is AlphaGo, which recently was won by Google. Normally, in game playing, you would really write down all the rules, but in AlphaGo's case, the system looked over a million games and was able to infer its own rules and then beat the world's residing Go champion. That is exciting, because it relieves the software engineer of the need of being super smart, and pushes the burden towards the data. As I said, the inflection point where this has become really possible -- very embarrassing, my thesis was about machine learning. It was completely insignificant, don't read it, because it was 20 years ago and back then, the computers were as big as a cockroach brain. Now they are powerful enough to really emulate kind of specialized human thinking. And then the computers take advantage of the fact that they can look at much more data than people can. So I'd say AlphaGo looked at more than a million games. No human expert can ever study a million games. Google has looked at over a hundred billion web pages. No person can ever study a hundred billion web pages. So as a result, the computer can find rules that even people can't find. CA: So instead of looking ahead to, "If he does that, I will do that," it's more saying, "Here is what looks like a winning pattern, here is what looks like a winning pattern." ST: Yeah. I mean, think about how you raise children. You don't spend the first 18 years giving kids a rule for every contingency and set them free and they have this big program. They stumble, fall, get up, they get slapped or spanked, and they have a positive experience, a good grade in school, and they figure it out on their own. That's happening with computers now, which makes computer programming so much easier all of a sudden. Now we don't have to think anymore. We just give them lots of data. CA: And so, this has been key to the spectacular improvement in power of self-driving cars. I think you gave me an example. Can you explain what's happening here? ST: This is a drive of a self-driving car that we happened to have at Udacity and recently made into a spin-off called Voyage. We have used this thing called deep learning to train a car to drive itself, and this is driving from Mountain View, California, to San Francisco on El Camino Real on a rainy day, with bicyclists and pedestrians and 133 traffic lights. And the novel thing here is, many, many moons ago, I started the Google self-driving car team. And back in the day, I hired the world's best software engineers to find the world's best rules. This is just trained. We drive this road 20 times, we put all this data into the computer brain, and after a few hours of processing, it comes up with behavior that often surpasses human agility. So it's become really easy to program it. This is 100 percent autonomous, about 33 miles, an hour and a half. CA: So, explain it -- on the big part of this program on the left, you're seeing basically what the computer sees as trucks and cars and those dots overtaking it and so forth. ST: On the right side, you see the camera image, which is the main input here, and it's used to find lanes, other cars, traffic lights. The vehicle has a radar to do distance estimation. This is very commonly used in these kind of systems. On the left side you see a laser diagram, where you see obstacles like trees and so on depicted by the laser. But almost all the interesting work is centering on the camera image now. We're really shifting over from precision sensors like radars and lasers into very cheap, commoditized sensors. A camera costs less than eight dollars. CA: And that green dot on the left thing, what is that? Is that anything meaningful? ST: This is a look-ahead point for your adaptive cruise control, so it helps us understand how to regulate velocity based on how far the cars in front of you are. CA: And so, you've also got an example, I think, of how the actual learning part takes place. Maybe we can see that. Talk about this. ST: This is an example where we posed a challenge to Udacity students to take what we call a self-driving car Nanodegree. We gave them this dataset and said "Hey, can you guys figure out how to steer this car?" And if you look at the images, it's, even for humans, quite impossible to get the steering right. And we ran a competition and said, "It's a deep learning competition, AI competition," and we gave the students 48 hours. So if you are a software house like Google or Facebook, something like this costs you at least six months of work. So we figured 48 hours is great. And within 48 hours, we got about 100 submissions from students, and the top four got it perfectly right. It drives better than I could drive on this imagery, using deep learning. And again, it's the same methodology. It's this magical thing. When you give enough data to a computer now, and give enough time to comprehend the data, it finds its own rules. CA: And so that has led to the development of powerful applications in all sorts of areas. You were talking to me the other day about cancer. Can I show this video? ST: Yeah, absolutely, please. CA: This is cool. ST: This is kind of an insight into what's happening in a completely different domain. This is augmenting, or competing -- it's in the eye of the beholder -- with people who are being paid 400,000 dollars a year, dermatologists, highly trained specialists. It takes more than a decade of training to be a good dermatologist. What you see here is the machine learning version of it. It's called a neural network. "Neural networks" is the technical term for these machine learning algorithms. They've been around since the 1980s. This one was invented in 1988 by a Facebook Fellow called Yann LeCun, and it propagates data stages through what you could think of as the human brain. It's not quite the same thing, but it emulates the same thing. It goes stage after stage. In the very first stage, it takes the visual input and extracts edges and rods and dots. And the next one becomes more complicated edges and shapes like little half-moons. And eventually, it's able to build really complicated concepts. Andrew Ng has been able to show that it's able to find cat faces and dog faces in vast amounts of images. What my student team at Stanford has shown is that if you train it on 129,000 images of skin conditions, including melanoma and carcinomas, you can do as good a job as the best human dermatologists. And to convince ourselves that this is the case, we captured an independent dataset that we presented to our network and to 25 board-certified Stanford-level dermatologists, and compared those. And in most cases, they were either on par or above the performance classification accuracy of human dermatologists. CA: You were telling me an anecdote. I think about this image right here. What happened here? ST: This was last Thursday. That's a moving piece. What we've shown before and we published in "Nature" earlier this year was this idea that we show dermatologists images and our computer program images, and count how often they're right. But all these images are past images. They've all been biopsied to make sure we had the correct classification. This one wasn't. This one was actually done at Stanford by one of our collaborators. The story goes that our collaborator, who is a world-famous dermatologist, one of the three best, apparently, looked at this mole and said, "This is not skin cancer." And then he had a second moment, where he said, "Well, let me just check with the app." So he took out his iPhone and ran our piece of software, our "pocket dermatologist," so to speak, and the iPhone said: cancer. It said melanoma. And then he was confused. And he decided, "OK, maybe I trust the iPhone a little bit more than myself," and he sent it out to the lab to get it biopsied. And it came up as an aggressive melanoma. So I think this might be the first time that we actually found, in the practice of using deep learning, an actual person whose melanoma would have gone unclassified, had it not been for deep learning. CA: I mean, that's incredible. It feels like there'd be an instant demand for an app like this right now, that you might freak out a lot of people. Are you thinking of doing this, making an app that allows self-checking? ST: So my in-box is flooded about cancer apps, with heartbreaking stories of people. I mean, some people have had 10, 15, 20 melanomas removed, and are scared that one might be overlooked, like this one, and also, about, I don't know, flying cars and speaker inquiries these days, I guess. My take is, we need more testing. I want to be very careful. It's very easy to give a flashy result and impress a TED audience. It's much harder to put something out that's ethical. And if people were to use the app and choose not to consult the assistance of a doctor because we get it wrong, I would feel really bad about it. So we're currently doing clinical tests, and if these clinical tests commence and our data holds up, we might be able at some point to take this kind of technology and take it out of the Stanford clinic and bring it to the entire world, places where Stanford doctors never, ever set foot. CA: And do I hear this right, that it seemed like what you were saying, because you are working with this army of Udacity students, that in a way, you're applying a different form of machine learning than might take place in a company, which is you're combining machine learning with a form of crowd wisdom. Are you saying that sometimes you think that could actually outperform what a company can do, even a vast company? ST: I believe there's now instances that blow my mind, and I'm still trying to understand. What Chris is referring to is these competitions that we run. We turn them around in 48 hours, and we've been able to build a self-driving car that can drive from Mountain View to San Francisco on surface streets. It's not quite on par with Google after seven years of Google work, but it's getting there. And it took us only two engineers and three months to do this. And the reason is, we have an army of students who participate in competitions. We're not the only ones who use crowdsourcing. Uber and Didi use crowdsource for driving. Airbnb uses crowdsourcing for hotels.