Synopsis: What would it be like to have a computer examine you and diagnose your illness? How about a robot making a reservation at an exciting new restaurant that is just perfect – without any direct input from you? Sound like science fiction? Well, our guest says that it already is…or soon will be…science fact. We talk to a computer scientist and author about the rise of computers that can learn on their own and then teach other computers to do the same.
Host: Gary Price. Guest: Pedro Domingos, professor of computer science at the University of Washington, author of the book, The Master Algorithm: How the quest for the ultimate learning machine will remake our world.
Link for more info:
The Master Algorithm: Is there a robot in your future?
Gary Price: Robots, super computers and all sorts of amazing technologies are the mainstays of science fiction books, movies and television shows. The stories are usually set in the future, but much of the whizz-bang technology is here right now, and Pedro Domingos is on the cutting edge of much of it. Domingos is a professor of computer science at the University of Washington. He says that engineers are working on creating computers that can teach themselves and other computers — without the aid of humans — using this master algorithm…
Pedro Domingos: An algorithm is really just how we tell computers to do what we want. It’s a sequence of instructions, not unlike a cooking recipe but a lot more detailed and precise. Now, in the old days, human programmers had to write the algorithms line by line. The exciting thing today is that with machine learning the computers can learn to program themselves. So they write their own algorithms by looking at your data and trying to figure out what you want. And the master algorithm is a learning algorithm that can learn anything from data – any knowledge, any skill it can acquire from looking at data.
Price: Domingos goes into the topic in depth in his new book, The Master Algorithm: How the quest for the ultimate learning machine will remake our world. The data – and how much of it — the computer needs to learn a particular skill or set of skills depends on what you are asking it to do. Take an online dating site, for example…
Domingos: Then what the algorithm, for example, might need to look at is things like the profiles of the people and trying to figure out if they’re compatible. Or maybe their pictures and trying to figure out what people find attractive in each other, and so forth. And then how much data you need? Well it depends on how complex the thing is that you’re trying to learn. So if what you’re trying to learn is simple, maybe you don’t need a lot of data. If it is very complex, maybe you do need a lot of data like, for example, if you’re trying to learn to cure cancer, that is a complex problem and you are going to need a lot of data. It’s not just enough to have a lot of data, it has to be good data. But one thing learning algorithms are already very good at is kind of getting rid of the noise, getting rid of the garbage and, basically, only zeroing in on the data that is actually good and useful.
Price: There are learning machines already in use, and you may have seen one of the celebrity ones on Jeopardy! in 2011. IBM’s Watson computer beat the show’s champions Ken Jennings and Brad Rutter, scoring more than three times better than either human. The computer was programmed specifically for the match, and Domingos says that IBM scientists used a variety of data to prepare the computer for the show…
Domingos: It learned in two main ways. One of the ways that it learned was just by having a database of past episodes of Jeopardy! It basically looked at all of that and learned all those questions and answers and then tried to generalize from them to the new questions that it might find. So that was one aspect was the database of past Jeopardy! episodes. But another one was, then it was trained by the computer scientists at IBM when it had these candidate answers, let’s say it had six candidate answers and it picked Answer 1, but the right answer was Answer 2. The people at IBM told it that and you know it got a bunch of learning episodes like this and got better as they did more of it.
Price: Watson was a very public example of machine learning, and a lot of it is going on – some behind closed doors and some right under our noses. Domingos says that medicine is making progress in the diagnosis and treatment of diseases using algorithms, and computers are learning about individual patients by looking through the medical records of millions…
Domingos: You give the algorithm a database of patient records like what their symptoms were, what the test results were, and then, maybe, what the diagnosis was or what the therapeutic was or what surgery they needed. And then what the learning algorithm does, it looks at all those cases and tries to generalize from that and say, “Oh, for example this, let’s say, breast x-ray, there is a tumor in it; this one there isn’t. And then it tries to figure out how do you look at a breast x-ray and decide it has a tumor or not? And it does this on its own, so the learning algorithm actually knows nothing whatsoever about the particular disease or condition going in, it basically does it by looking at that. And the amazing thing is that it is often the case that algorithms that learn like this in, literally, a few minutes, actually are more accurate at diagnosing things than doctors who spend many years training.
Price: One thing the machine does better than a human is to remain objective. Domingos says that there’s no personality involved in an algorithm that’s come up with a diagnosis for a complex disease like cancer…
Domingos: Humans are extremely inconsistent. If they’re in a bad mood, they will give more pessimistic diagnoses. If something else just happened, they might be influenced by it. The algorithms are very consistent, and as a result they can be consistently high quality if they learn well enough, and humans, in some sense, are more flakey, if you will.
Price: But what about the doctor-patient relationship? Doesn’t having a human there to explain and maybe even hold your hand create feelings of comfort and trust?
Domingos: There’s two ways that this could go. One is that computers can learn to understand and react to emotions as well. There’s already a lot of this going on in research labs and some of it will happen. So a computer in the future, if it’s acting as your doctor, it could provide some of that emotional support. The question, however, is whether you will be satisfied with it. I mean will you feel that emotional support means anything? If it’s coming from a machine, maybe not. In which case maybe what you will have is it will still be a human doctor talking with you but they would have made their diagnosis not just by thinking about it themselves, but by largely or even almost entirely relying on the computer.
Price: A place where medical learning machines can excel is in the laboratory. Domingos says that there’s one operating right now that has already created a medical breakthrough…
Domingos: At the University of Manchester a robot scientist called Eve that is actually a machine that knows molecular biology, it formulates hypotheses like a scientist, it carries out experiments with DNA, micro-rays and sequencers to test the hypothesis. Physically, it a real robot that does this completely on its own without any human help. And then it, you know, revises its hypothesis and whatnot. So it’s doing the whole scientific process all by itself. For example, it recently discovered a new malaria drug. Right not there’s only, you know, this one robot but you can imagine having a million of them and now you have a million scientists where before you just, you know, had a few humans.
Price: Robots that teach themselves will come in all shapes and sizes. Domingos says that we already have many of them in our midst…
Domingos: Things like Siri, for example, are a good preview of that. Siri is, you know, has a bunch of learning algorithms that are learning about you. And then you talk to it and it talks to you, so that’s another way in which a learning machine could appear to you. But then over and beyond that there’s going to be a vast number of machines that are learning about you. There already are, and you don’t even know. You don’t notice. It’s like your car is learning about it, it’s learning about your driving style in order to, you know, save you gas. It’s learning about you in order to find, you know, the best route to where you want to go. So every little device from the smallest to the largest will have these learning capabilities and some degree of intelligence. Some of them you be very aware of it and you’ll kind of interact with them as such, and the others you won’t notice and not even necessarily care.
Price: The driverless cars that we’ve been hearing so much about lately are another type of learning machine that we can look forward to in the future. But with robots working in laboratories, hospitals, driving us where we want to go, taking over manufacturing tasks — what will people do? Are we inventing ourselves out of our jobs?
Domingos: The jobs that we do are going to change. So some jobs will disappear but also a lot of new jobs will appear and many jobs, and I think in the short term this is actually the bigger part of it, is that it’s neither that a job will appear or disappear, it will be changed by learning machines. There will be parts of your job that the machine can actually do very well, and what you actually want to do is you want, you know, start using machine learning to do those parts of your job and then you do all of the other things that you actually didn’t have the time to do. Think of it as having a horse, right? If you have a horse, you don’t try to outrun it, you ride it. And you know machine learning is a kind of horse that you can ride and so what people need to do is learn how best to interact with the learning algorithms that they have in order to do their jobs better.
Price: If there are master algorithms developed and machines can program other machines and learn so much about us, isn’t there a danger that they’ll take over? That we’ll end up working for them, or worse, being hurt by them?
Domingos: I think that danger is actually a lot less than people often think when they think of movies like, you know, Terminator and whatnot. The reason is actually quite simple, it’s that when we think of intelligent machines we think of them as being almost human, right? Because intelligence and human, for us, are very closely related. But the truth is that being intelligent and being human are very, very different. In Hollywood movies, the robots, you know and the AIs (Artificial Intelligence) are always humans in disguise. But the real ones, all that they’re really doing is searching for solutions to the problems that we set them. So as long as that’s all that they can do, they could be infinitely intelligent and they are of no danger to us. In fact, the more intelligent they are, the better because they can solve harder problems like curing cancer. I think what we need to do is we set the goals, the AIs solve the problem and then we check the solutions. What we can’t let is the robots start to set their own goals, like, “Oh, my goal is to survive and you know proliferate.” Well, that’s a very dangerous thing to do.
Price: Domingos sees a bigger threat to us from robots and computers inadvertently harming us in some way…
Domingos: By not having understood very well what we ask them to do, by not having common sense and, in fact, some of this is already happening, right? Computers already make a lot of important decisions about who gets credit, who gets what job, who gets flagged as a potential terrorist, and they make mistakes. So, you know, people worry that computers are going to get too smart and take over the world. But the real problem is that they’re too stupid and they’ve already taken over the world.
Price: Pedro Domingos says that his goal in writing his book, The Master Algorithm, is not to scare people with a sci-fi version of the future, but to let everyone know that machine learning is here already and we need to learn how to make the most of it. The book is available in stores, online and on his website, PedroDomingos.org. For more information about all of our guests, visit our site at Viewpointsonline.net. You can find archives of past programs there and on iTunes and Stitcher. I’m Gary Price.