The View From a Visionary
Yann LeCun, a pioneer in the field, traces AI's evolution
LeCun was not even 30 when he joined the Adaptive Systems Research
Department at AT&T Bell Laboratories in New Jersey. There, his
enthusiasm for artificial intelligence couldn’t be contained.
At Bell Labs, LeCun developed a number of new machine learning methods,
including the convolutional neural network—modeled after the visual
cortex in animals. LeCun’s work also contributed to the advancement of
image and video recognition, as well as natural language processing.
A few decades later, LeCun is now Chief AI Scientist at Facebook.
Yann LeCun: An AI Groundbreaker Takes Stock
Yet another obstacle dogged any dreams of AI from taking form. In 1984, the American Association of Artificial Intelligence held a fateful meeting where field pioneer Marvin Minsky, of all people, warned the business community that investor enthusiasm for artificial intelligence would eventually lead to disappointment. Sure enough, AI investment began to collapse.
It’s a good thing, then, that visionaries such as Yann LeCun chose not to pay the pessimism much mind. The native of France was not even 30 when he joined the Adaptive Systems Research Department at AT&T Bell Laboratories in New Jersey. There, his enthusiasm for artificial intelligence couldn’t be contained.
At Bell Labs, LeCun developed a number of new machine learning methods, including the convolutional neural network—modeled after the visual cortex in animals. LeCun’s work also contributed to the advancement of image and video recognition, as well as natural language processing.
“The whole idea of statistical learning in the context of AI kind of died in the late 1960s,” LeCun recalls. “People more or less abandoned it. Then it came back to the fore in the late ’80s with interest in neural nets. So when learning algorithms to train multilayer neural nets popped up in the mid-’80s, it created a wave of interest.”
In capturing this revolution, LeCun is modest to a fault. He’s made history for his discoveries, but he barely mentions his own name or accomplishments. He refuses to take himself seriously; in fact, a whole section of his personal website is devoted to puns, with this self-admonition: “The Geneva convention against torture, and the U.S. constitutional protection against cruel and unusual punishments, forbid me to write more than three atrocious puns in a row.”
LeCun also refuses to rest on any of his well-earned laurels in computer science; today, he serves as Facebook’s chief AI scientist, where he works tirelessly towards new breakthroughs. Here, he takes us on a privileged tour—better than a front-row seat, because he’s a star of the show—through the growth, recent changes and potential of artificial intelligence.
AI Begins—Perceptrons To The Precipice Of Learning
As a student of AI’s past, LeCun can cite the milestones as well as anyone, starting with the summer 1956 brainstorming session at Dartmouth where the term “artificial intelligence” was coined. Just a year later, Frank Rosenblatt invented the perceptron at the Cornell Aeronautical Laboratory. One of its first implementations was the Mark 1 Perceptron, a mammoth rectangular machine that contained 400 photocells randomly connected to simple motif detectors fed to a trainable classifier.“It was the first neural network that could learn to recognize simple patterns in a kind of non-trivial way,” LeCun says. “You could use them to do simple image recognition but not to recognize objects in photos and not for any kind of reasoning or planning.”
Until the last decade, pattern recognition systems required a lot of human grunt work to recognize objects in natural images. “You’d have to work a lot on building an engineered module that would turn the images into a representation—generally a long list of numbers that can be processed by those simple learning algorithms. So you basically had to do the work by hand.” Ditto, he adds, for early speech recognition and computer-driven translation: Hand-engineering meant maximal sweat with minimal results.
So what changed insofar as the computer science? “In all of those applications, deep learning and neural nets have brought significant improvements in performance—and also considerable reduction in sort of the manual labor that is necessary,” LeCun says. “And that allows people to expand the applications of these to a lot of different domains.”
This raises the question of how computers can “learn” in the first place. Neural nets function as software simulations of the brain; they process information such as a visual image and attempt to arrive at a correct answer. But what if that answer isn’t quite right? Enter “backpropagation,” an algorithm for feedback flow that enables neural networks to learn.
LeCun And The Backpropagation Proposition
The breakthrough discovery of backpropagation came in 1986, when professor Geoffrey Hinton became one of the first researchers to describe a way computers could learn by performing a task over and over, each time with the computer’s neural network “then adjusted in the direction that decreases the error.”LeCun not only made good on Hinton’s groundwork—he helped lay the foundation. Hinton had first floated the idea of “backprop” in the early 1980s but abandoned it because he didn’t think it could work.
But in 1985 LeCun wrote a paper that described a form of backpropagation for, as he puts it, “an obscure conference. It was in French and basically wasn’t read by many people—but at least by one important person.” That would be Hinton. LeCun then served as a postdoctoral research associate under Hinton at the University of Toronto before he began his work at AT&T Bell Labs (also the birthplace of the transistor).
“All of machine learning is about error correction,” LeCun explains. Imagine showing a computer “thousands of images of cars and airplanes, and every time the parameters adjust themselves a little bit, the outputs get closer to the correct one—and eventually settle on a configuration where every car and every airplane you turn the machine on are correctly recognized, if you’re lucky enough.”
Awe colors LeCun’s voice when he describes the end result: “The magic of learning is that even images the system has never seen will be correctly classified.”
Still, he can’t avoid being just a bit puckish. “There are all kinds of tricks to get backprop to work, and it’s still a bit of a black art—but now we have a recipe. If you follow the recipe, it’s going to work every time.”
"All of machine learning is about error correction."
Data, AI And Business: The Skies, And The Limits
Data in the age of AI has been described in any number of ways: the new gold, the new oil, the new currency and even the new bacon. By now, everyone gets it: Data is worth a lot to businesses, from auditing to e-commerce. But it helps to understand what it can and cannot do, a distinction many in the business world still must come to grips with.“Data is important for making a business out of machine learning,” LeCun acknowledges. “You need data to train your system, and the more data you have, the more accurate your system will be. So, from a technology goal and business point of view, having more data is better.”
But there also comes a time when, if you will, data becomes the greasy bacon: That is, it can’t make the machines that use artificial intelligence more intelligent. “With the research aspect of AI—the stuff that we work on here at Facebook and that a lot of people at Deep Mind and Google Brain and other places work on—we don’t use internal data to test them,” LeCun says. “We use public data because we like to compare our methods with other people’s in the academic research community. And having more data is not critical to developing better methods. In fact, a lot of effort is devoted to reducing the amount of data required to reach a given level of performance.”
That’s especially true in the case of academia, where the critical role isn’t to crunch terabytes of terabytes, but rather to serve as what LeCun calls “the vanguard of new ideas.”
"Data is important for making a business out of machine learning. You need data to train your system, and the more data you have, the more accurate your system will be."Meanwhile, businesses building an AI strategy need to self-assess before they look for solutions. “It depends how critical AI is to your operation,” LeCun points out. “If you just want to apply existing AI methods, you can use cloud services that a number of companies are offering. It’s relatively easy.” Some businesses and for-hire technologies can help with AI deployment; LeCun mentions Element AI in Montreal as an example.
The biggest challenge is for companies looking to build their own engineering teams. “Basically, AI engineers and scientists are in high demand these days, and so you have to pay them. They’re not cheap, and it’s because they’re rare.”
Two Types Of Learning, One Luminous Future
LeCun outlines two different types of learning that form the basis of artificial intelligence today: supervised and unsupervised. With supervised learning—applicable to more than 95% of all machine learning applications—human operators train machines to better recognize images or other forms of input over time. By way of analogy, think of it as knobs you can adjust automatically so that the output of the machine gets closer to the one you want.Unsupervised or “self-supervised” learning, though it represents a much smaller percentage of today’s machine learning, holds very large potential. “It’s based on essentially predicting everything about what we perceive in the world from everything else,” LeCun says. He uses “video prediction” as an example: “You show a small snippet of video to a machine, and you ask it to predict what’s going to happen next.”
Right now, that’s a little like predicting what will happen next to yield that particular breakthrough. But to be certain, there’s more than enough enticement for scientists, academics and high-tech giants to pursue unsupervised learning. “The prize for this is being able to perform all the applications that we currently can’t do,” LeCun says. “So, we’d like to have intelligent virtual assistants that you can talk to that can understand everything you say. They’ll have enough background about how it all works to really help you in your daily lives.”
He pauses. “It’s kind of like the movie 'Her.' You’ve seen that movie?” A quick recap: The 2013 Spike Jonze film stars Joaquin Phoenix as a lonely writer who falls in love with his virtual assistant, voiced by Scarlett Johansson. It turns out LeCun likes it.
“It’s not a bad depiction of a possible interaction between people and their virtual assistants once they become intelligent,” he says. “We’re very far from having AI technology that would allow us to build machines like this. And it’s basically because machines today don’t have common sense.”
"A housecat has way more common sense than the smartest machines."Common sense? But wouldn’t a machine make a better decision in many situations than a human? Machines must have common sense—or do they? LeCun explains why they don’t: “We don’t have the ability to get machines to learn all the enormous background knowledge: the enormous background knowledge about the world that we acquire in the first few weeks and months of life as humans—and that a lot of animals also acquire.”
And because of this, some of our simplest assumptions about robots don’t hold water—not even dishwater. “We cannot have dexterous robots,” LeCun contends. “We cannot have household robots that can fill up our dishwasher and empty it. That’s beyond today’s state-of-the-art in robotics, and it’s not because we can’t build the robots. It’s because we don’t know how to build their brains. We don’t know how to train them to know enough about how you grasp things, how you get around obstacles and how you put things in.”
He adds: “A housecat has way more common sense than the smartest machines.”
Given LeCun’s place in bringing artificial intelligence to life, that might sound like a flip (if lighthearted) way to dismiss things. But he also holds out tremendous enthusiasm—wonder, even—when he considers a bright AI future approaching at lightning speed, in fields such as medicine.
“With medical image analysis, we can train convolutional nets to detect tumors in CT scans or MRIs, or detect melanoma from skin pictures,” he says. “I think this is going to have profound effects on radiology.”
No matter his opinions, Yann LeCun retains the playfulness of a twentysomething reporting for the first day of work at Bell Labs. If you follow him via Twitter (@ylecun), you’ll find out firsthand, just checking out his tweet that summed up his shared research of potential connections between deep neural networks and the “glassy systems” of solid-state physics.
LeCun posted it on July 8, the day he turned 58: “Deep Nets are classy and glassy.”
Learn more about how companies are leveraging AI today.
CREDITS: Facebook
About the Author
Forbes Insights is the strategic research and thought
leadership practice of Forbes Media. By leveraging proprietary databases of
senior-level executives in the Forbes community, Forbes Insights conducts
research on a wide range of topics to position brands as thought leaders and
drive stakeholder engagement.
No comments:
Post a Comment