Faculty Friday: Pedro Domingos
Most computer scientists seek to empower machines. Pedro Domingos would rather empower people.
“You may not be interested in technology, but technology is interested in you,” he says, using a witting corruption of Leon Trotsky’s famous quip about politics to underscore why he believes it’s increasingly important for people to take an active interest in how artificially intelligent technology is coming to influence everyday life.
A professor at the University of Washington’s Paul G. Allen School of Computer Science and Engineering, Domingos conducts research in the subfield of computer science known as machine learning, the process by which computers are given the ability to learn through the use of pattern recognition and data analysis—all without being explicitly programmed to do so. After an incubation period of several decades, the field has burst into popular consciousness—and everyday life—over the past several years with predictive algorithms increasingly omnipresent online, on the road, and around the home.
“People say machine learning is really the second phase of the Information Age,” Domingos says. “The first age was when we programmed computers and things happened at a certain speed, but now we’re in a different phase when the computers are going to program themselves and this is going to reach a whole different level.”
One basic way machine learning affects everyday life is that it’s increasing peoples’ range, Domingos says, citing as examples Amazon’s use of algorithms to filter through millions of books to suggest ones you might enjoy or music sites’ ability to surface rare tracks that correspond to obscure individual tastes. One might think understanding how these algorithms function and evolve isn’t child’s play. But, according to Domingos, it’s exactly that.
“Think of how a baby or a child learns: by observing, by learning from its parents, by trying things out and seeing what happens. In a way, machine learning is doing what children do, which is learning by themselves,” he says. “The whole idea of machine learning is that you don’t have to program computers; they program themselves by learning from data.”
Letting a computer know more about yourself—your tastes and desires (and, by certain extension, your hopes and dreams)—helps the machine tailor a more natural experience that facilitates action—and interaction—within a given space. In essence, machine learning makes technology, well, a little more human.
“Technology can feel very cold and impersonal,” Domingos says. “With machine learning, it can feel a little bit warmer.”
Many think it can also feel a little bit intrusive, stirring debates about privacy and power, all while producing a muddle of misconceptions.
“People either think it can do all sorts of things that it can’t or that it can’t do all sorts of things that it can,” Domingos says—quick to quell the notion that machine learning is simply about summarizing data and spitting out an answer. Rather, he says, it’s about continually adjusting to stay one step ahead through generalization and informed inference. “People are overestimating what it can do in the short term, but they underestimate what it can do in the long term.”
Machine misunderstood
In an effort bring some clarity to a complex, often incongruous conversation, Domingos wrote The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World, a best-selling book that surveys how machine learning’s data-driven algorithms are shaping—and will continue to shape—business, science, and society at large. In it, he explores the impact self-teaching, data-driven algorithms are having on how medical diagnoses are made, how hirings happen (or, for that matter, don’t), analyzing and acting on global markets, and in many other fields.
“Making recommendations is really the beginning,” he says. “These things can run a lot of your life for you.”
He cautions, however, that progress is not going to happen all at once. Nor will it happen without adequate understanding of the power on the part of people, who have a responsibility to look out for themselves as consumers and as citizens. It’s not so much that Domingos has come from the future to save us from ourselves, rather, he’s telling us that we have it in our reach to bend the future to our best interests.
“The companies that create these algorithms acquire a lot of power and you want to make sure that all this in your benefit,” he says. “If it’s all under the hood and all in a black box as it tends to be today, you as a user don’t have a lot of control over it. People need to understand what’s going on so they can have control over it.”
While concerns about the steady creep of computer technology should be taken seriously, Domingos believes the best approach isn’t to tune tech out entirely, rather, it’s to engage it head-on with a humanist approach that draws on the power of philosophy and ethics.
“How should we program machines to be ethical?” he asks. “If a self-driving car runs someone over, who’s responsible? The company that made it? The data that it was trained on? The driver?”
It might seem a complicated—not to mention messy—conundrum. But Domingos says that when it comes to the ethics of artificial intelligence, it’s very simple.
“Machines are not independent agents—a machine is an extension of its owner—therefore, whatever ethical rules of behavior I should follow as a human, the machine should do the same. If we keep this firmly in mind,” he says, “a lot of things become simplified and a lot of confusion goes away.”
It’s only simple so far as the ethical spectrum remains incredibly complex, and, as Domingos will be first to admit, everybody doesn’t have the same ethics.
“One of the things that is starting to worry me today is that technologists like me are starting to think it’s their job to be programming ethics into computers, but I don’t think that’s our job, because there isn’t one ethics,” Domingos says. “My job isn’t to program my ethics into your computer; it’s to make it easy for you to program your ethics into your computer without being a programmer.”
He compares his work as a researcher who studies machine learning to that of a mechanic tasked with making sure a car is safe and in driving condition. He wants his machine to get people to where they want to go.
“People don’t need to understand the gory details of how machine learning algorithms work,” he says. “It’s like driving a car—you don’t need to learn how the engine works—the mechanic does—but you need to understand how to use the steering wheel and pedals. These days people don’t even know that machine learning has a steering wheel and pedals. They need the wheel in their hands and this is what hasn’t happened.”
It’s the kind of crash course Domingos determined early on would be worthy of his life’s work.
Future’s past
Pedro Domingos has always been a reader. He almost was a writer.
Once, as a graduate student at UC Irvine, he spent six weeks at the Clarion West Writers Workshop in Seattle writing science fiction stories. It would prove his first visit to the city he would later call home. While there, he had what could be considered the second great epiphany of his career.
“What I realized was science is moving faster than science fiction,” he says. “It’s more interesting to do science.”
His first breakthrough had come several years before when he was still an undergraduate studying computer science and electrical engineering in his native Portugal. One day, he’d walked into a bookstore and spotted a book called Artificial Intelligence.
“It seemed like an oxymoron,” he recalls thinking. “How could something be artificial and intelligent at the same time?”
His second time past the shop, he bought the book and read it. Doing so, it dawned on him that a key element to artificial intelligence was machine learning.
“Back then it sounded crazy, but my belief then was that in the future everyone would do everything using machine learning,” he says. His reasoning was that the ability to learn is the essential ingredient of intelligent behavior. “If you program all sorts of intelligence into a computer, but it doesn’t learn, the following day, it will already be falling behind people.”
Domingos witnessed it first-hand growing up in and around his father’s own computer center, which he’d established as part of his work as a professor and mechanical engineer who used the machines as part of his research on heat transfer and fluid dynamics.
“In a way, I was very lucky, because I was exposed to all the phases of computer development,” Domingos says. “They had the whole IBM mainframe and you put in your punchcards and you got back a stack of print outs—it was that very old paradigm of computing.”
Domingos remembers going into the lab and playing “little video games” on an Apple II, which debuted in 1977 as one of the first personal computers aimed at a consumer market. The center also had one of the first portable computers, which Domingos recalls, “wouldn’t be portable by today’s standard.”
“People at large didn’t know that this wave of information technology was coming,” Domingos says of the late 1970s. From these early experiences, however, it was clear to Domingos that the digital age was about to dawn. “I knew that this was coming and that there was at least a reasonable picture of how this was going to go and that this was all very exciting.”
But, to Domingos’ eyes, even the highest of hopes for the burgeoning computer revolution seemed fatally short-sighted. Machine learning and its applications in the fields of computer science and artificial intelligence, on the other hand, seemed so far out into the future that Domingos determined it was a field worth fixing his gaze on.
“Things seemed to be in such a primitive state and this field was so new, I thought, ‘Alright, I can make a difference.'”
So he did, following a path from the Instituto Superior Técnico in Lisbon to one of the few programs in the world at the time to focus on artificial intelligence at UC Irvine. After receiving his Ph.D., he returned to Lisbon to teach for two years at his alma mater before joining the faculty at the UW in 1999 as the University’s first faculty member with a focus on machine learning.
Even in the throes of the late ’90s tech boom, machine learning was still an exotic field. But for Domingos, that was all part of the fun. Over the next decade and a half, his work in machine learning would prove a dilettante’s delight as a steady stream of people from all kinds of fields started trickling through his door wanting to apply machine learning to a range of projects.
“A little bit of technology today, a little bit of business tomorrow,” Domingos says of any given week. “I have very broad interests and am very omnivorous. Part of why I do machine learning is that it’s an excuse to do a whole bunch of different things besides.”
As a professor, his plate is full supervising thesis projects and advising students, whose work, Domingos adds, constitutes a good portion of the actual programming that goes on in his research. He likens a Ph.D. thesis to an airplane: “The student is the pilot. The role of the advisor is to be the control tower.”
It’s an apt metaphor for Domingos’ own unique vantage point on the progression of computing power into the 21st century.
“We are in this phase where progress is accelerating rapidly, but it’s not going to accelerate forever,” he says, adding that the progress of technology typically occurs in S-curves, in which little happens at first, before a scant progress gives way to rapid acceleration, before slowing down again and settling to a certain level. He says he believes we’re currently at the beginning stages of one such phase of acceleration.
“By the time this is all done, we will be in a different phase and the world will be very different,” he says. “Physicists call this a phase transition where we transition from one phase to another.”
Even so, he cautions that this doesn’t mean we’re fast approaching the technological singularity, a hypothesis that posits the invention of artificial super-intelligence will abruptly trigger a runaway reaction of exponential technological growth, resulting in unfathomable changes to human civilization.
“Computers are already not getting faster at the rate that they used to get faster,” he says. “It’s easy to extrapolate these trends and go too far. Progress is not that predictable.”
For one whose work is carrying the world into the future at a rate that seems faster than ever before, Domingos appears a man at ease in the present age. Hardcover books still line the shelves of an immaculate office.
“The Geographer,” by Johannes Vermeer (1669)
Its one ostentation: a reproduction of a work by the 17th century Dutch master, Johannes Vermeer—what Domingos calls, “the closest thing Vermeer has to a painting about science.”
The painting depicts a geographer looking up from his charts and calculations—the expression on his face something between deep thought and a passing reverie, as though momentarily distracted by the world passing before his window.
“It captures a moment in time,” Domingos says, standing before the work. “He’s measuring distances and he’s pausing for a moment to think. You wonder what he’s thinking about.”
Over the course of our hour-long interview, the late afternoon light has shifted, falling in wide shafts across the floor and along the bookshelves of Domingos’ sixth floor office for an effect that is not dissimilar to the glowing light depicted in the canvas hanging on the back wall. Gazing in tandem at the painting, contemplating the question of what might be on the geographer’s mind, I find my own thoughts wandering to what else might be racing through the mind of the man before me.
Domingos remains pensive.
Whatever the future may hold, one could find some measure of comfort in the fact that the professor who has done so much to shape its course can still pause for a minute to meditate on the mysteries of the past.
Pedro Domingos holds an undergraduate degree and M.S. in Electrical Engineering and Computer Science from Instituto Superior Técnico (IST) and an M.S. and Ph.D. in Information and Computer Science from the University of California, Irvine. In 2014, he received the SIGKDD Innovation Award Award, the highest honor in data science, for his foundational research in data stream analysis, cost-sensitive classification, adversarial learning, and Markov logic networks, as well as applications in viral marketing and information integration. Learn more about Pedro, his book, and other work here.