We may not be aware of it, but machine learning is already an integral part of our daily lives, from the product choices that Amazon offers us to the surveillance of our data by the National Security Agency. Few of us understand it or the implications, however.
Enter Pedro Domingos, professor of computer science at the University of Washington in Seattle and author of The Master Algorithm: How The Quest For The Ultimate Learning Machine Will Remake Our World.
Talking from his home near Seattle, he explains how Artificial Intelligence (AI) may one day make going to the doctor a thing of the past; how a Master Algorithm could match Einstein’s theory of relativity in its world-transformative power; and why replacing soldiers on the battlefield with robots might actually make warfare more humane.
The holy grail of computer science is a machine that can teach itself. Explain machine learning in simple language.
In traditional programming, you have to explain to the computer in painstaking detail what you want it to do. That’s what an algorithm is: a set of instructions you give the computer.
With machine learning you program the computer to learn by itself. When you do a web search, machine learning chooses the results you get. Amazon uses it to recommend products; Netflix uses it to recommend movies; Facebook and Twitter use it to choose which posts to show you. Pretty much everything that happens online involves machine learning.
What is the Master Algorithm? And how close are we to developing it?
The Master Algorithm is an algorithm that can learn anything from data. Give it data about planetary motions and inclined planes, and it discovers Newton’s law of gravity. Give it DNA crystallography data and it discovers the Double Helix. Give it a vast database of cancer patient records and it learns to diagnose and cure cancer.
But to do that we need a deeper understanding of how learning works. There are multiple approaches. One of them is to reverse engineer the brain. Another is to mimic evolution.
My own view is that you’re going to need to combine ideas from these paradigms to come up with a Master Algorithm. You need something like a grand unified theory of machine learning, like the grand unified theory of physics.
Somebody could discover it tomorrow or it could take hundreds of years. My gut feeling is that it will happen in our lifetime, and it will probably be someone who is actually not a professional machine-learning researcher.
I would love it if a 20 year-old college student read my book and had the idea that realizes the Master Algorithm.
Huge claims are being made for machine learning, including cures for AIDS and cancer. Might this really be possible in our lifetime?
Machines can attend to vastly more information and more complex processes than human beings, and try out more drugs or vaccines than we can in the lab. Machine learning is used to discover drugs by simulating the process on a computer, which takes a fraction of the time and cost.
The thing that makes the AIDS virus so difficult is that it mutates very quickly. A researcher named David Heckerman has developed the idea that you don’t attack the virus in only one place, like most vaccines do. You attack it in different places at the same time. But to discover all those places takes a scale of data processing and hypothesis testing humans can’t do.
With cancer, the problem is that it’s not just one disease. Everybody’s cancer is different and the cancer mutates as it grows, so today’s cancer in the same patient is not the same as it was six months ago. The metabolism is so complicated and there are so many different possible mutations and combinations of cellular and environmental factors that no human can possibly actually master all that.
As a result, there is no single drug that is going to cure cancer. Machine learning can potentially take in the cancer’s genome, the patient’s genome and medical history, predict which drug or combination of drugs to use or even design a new drug specifically for that cancer.
But we need better machine learning algorithms before we can do that. We also need patients to share their data so that the algorithms can learn from it. David Haussler, who is both a famous molecular biologist and machine-learning researcher, believes if we are able to gather that data from patients we will be able to cure cancer. Otherwise, we won’t.
An Oxford University survey suggested that 47 percent of the world’s jobs could be taken by machine learning in the coming decades. Which jobs will be most at risk?
That is a very pregnant question. Let me answer it the opposite way, by discussing which jobs will be least at risk. One of the surprises of AI in the last 50 years is that people thought we would start by automating the trivial things, like construction work or cleaning toilets and the hardest things would be what doctors and lawyers do.
It actually turns out to be exactly the opposite. Doctors and lawyers are much easier to automate than street sweepers. In fact, one of the big successes of machine learning is that you can take a simple algorithm, give it a database of patient records and it learns to diagnose diabetes or breast cancer better than people who have spent years in Med school.
What makes a job hard to automate? If a job involves routine mental work, and in many ways medical diagnosis is routine work, it is easily automated. If the job involves interacting with the physical world, using your hands and feet, that is very hard to automate.
Jobs that require a lot of common sense are also very hard to automate. Common sense is something we human beings take for granted, but is extremely difficult for machines to acquire. So if your job requires common sense it’s a lot safer than if it doesn’t.
Machine learning per se is not a threat, but it’s definitely one of the tools that an organization like the NSA uses. Let’s assume that the NSA can capture every electronic communication in the world. The problem is that the vast majority of those communications are completely innocent and they don’t have the manpower to look at them.
Again, this is where the machine learning comes in. You can have machine-learning programs that go through all these conversations, even in real time, to pick out the suspicious ones. So if you want to create a surveillance state, machine learning is potentially a really dangerous tool.
At the same time, machine learning can also put a lot of power in our hands. We can have our own machine learning algorithms, which will figure out how to defeat those who want to keep tabs on us. Like any technology, it can be used by both sides.
The important thing is that we, as citizens, have to understand what it is, so that we can make it work for us, as opposed to just allowing companies or states making it work for them.
The Chinese philosopher Laozi claimed, “The best soldier fights without vengeance, without anger and without hate.” Might robots one day replace human soldiers on the battlefield?
There’s a big polemic going on around this question right now. Some people want to ban robots and intelligent weapons from the battlefield. I think that would be a huge mistake, precisely for the reason you mentioned. Robot soldiers have the potential to be much more humane than human soldiers. They don’t get angry or frightened or vengeful; they keep their heads in the heat of battle; and they can put a lot more thought into the decision where or what to shoot.
The United Nations is currently considering banning intelligent weapons, the same way chemical and biological weapons were banned. But that strikes me as backwards. What we want is to develop AI to the point where all wars will be fought by robots. If we ban humans in combat rather than robots, warfare will hopefully just become a competition to see who can build and destroy the most, rather than killing people.
What things in your children’s lives do you think will be improved – or degraded – by machine learning in the future?
In the short term, machine learning, like other technologies, will have winners and losers. In the long term, we will weed out the bad things and mainly have good consequences. One of them, which we’ve touched on already, is health. Our children’s generation is going to regard health problems as something awful that people had in the barbaric past.
Today, when you get sick, you go to the doctor and hopefully the doctor figures out a treatment. In the future is you won’t get sick in the first place because there’s all this work happening in the background that cures it before it happens.
Let me give you a short-term example and a longer one. Your smart phone is full of sensors, and there are going to be more. You might even have sensors in your body one of these days. If your smart phone feels that you are about to have a heart attack, it will call 911 and also warn you, so it might save your life. I think this will happen in the fairly short term.
In the longer term, picture an outbreak of a new disease, like the next Ebola. In the future, the virus will get sequenced very quickly, and then labs will find the vaccine or cure and your immune system will download the instructions for the cure from the Internet into your body, without you even necessarily being consciously aware of it.
More generally, I think machine learning, like other technologies, gives you a kind of superpower. Telephones let you communicate at a distance; airplanes let you fly. Machine learning will be the ultimate superpower.
The human species adapts the world to itself rather than having to adapt to the world. Machine learning is going to take that to a whole new level, which is that the world adapts itself to you. When you go into a car or enter a new environment, whether on or offline, it will automatically configure itself for you. A lot of the battles we have today with information overload, or things going wrong, won’t happen anymore. You’ll just be happier and more productive.