Google DeepMind\'s AI Program Learns Human Navigation Skills

Oren Etzioni is a smiling bear of a guy. A computer scientist who runs the Allen Institute for Artificial Intelligence in Seattle, he greets me in his bright office wearing jeans and a salmon-­colored shirt, ushering me in past a whiteboard scrawled with musings about machine intelligence. (“DEFINE SUCCESS,” “WHAT’S THE TASK?”) Outside, in the sun-drenched main room of the institute, young AI researchers pad around sylphlike, headphones attached, quietly pecking at keyboards.

Etzioni and his team are working on the common-sense problem. He defines it in the context of two legendary AI moments—the trouncing of the chess grandmaster Garry Kasparov3 by IBM’s Deep Blue in 1997 and the equally shocking defeat of the world’s top Go player by DeepMind’s AlphaGo last year. (Google bought DeepMind in 2014.)

3 In 1996, Kasparov—then the best chess player in the world—beat Deep Blue. During a rematch a year later, Kasparov surrendered after 19 moves. He later told a reporter: “I’m a human being. When I see something that is well beyond my understanding, I’m afraid.”

“With Deep Blue we had a program that would make a superhuman chess move—while the room was on fire,” Etzioni jokes. “Right? Completely lacking context. Fast-forward 20 years, we’ve got a computer that can make a superhuman Go move—while the room is on fire.” Humans, of course, do not have this limitation. His team plays weekly games of bughouse chess, and if a fire broke out the humans would pull the alarm and run for the doors.

Humans, in other words, possess a base of knowledge about the world (fire burns things) mixed with the ability to reason about it (you should try to move away from an out-of-control fire). For AI to truly think like people, we need to teach it the stuff that everyone knows, like physics (balls tossed in the air will fall) or the relative sizes of things (an elephant can’t fit in a bathtub). Until AI possesses these basic concepts, Etzioni figures, it won’t be able to reason.

With an infusion of hundreds of millions of dollars from Paul Allen,4 Etzioni and his team are trying to develop a layer of common-sense reasoning to work with the existing style of neural net. (The Allen Institute is a nonprofit, so everything they discover will be published, for anyone to use.)

4 Microsoft cofounder and philanthropist Paul Allen donated billions to science, climate, and health research, as well as to Seattle causes. He died of complications from cancer on October 15 at age 65.

The first problem they face is answering the question, What is common sense?

Etzioni describes it as all the knowledge about the world that we take for granted but rarely state out loud. He and his colleagues have created a set of benchmark questions that a truly reasoning AI ought to be able to answer: If I put my socks in a drawer, will they be there tomorrow? If I stomp on someone’s toe, will they be mad?

One way to get this knowledge is to extract it from people. Etzioni’s lab is paying crowdsourced humans on Amazon Mechanical Turk to help craft common-sense statements. The team then uses various machine-learning techniques—some old-school statistical analyses, some deep-learning neural nets—to draw lessons from those statements. If they do it right, Etzioni believes they can produce reusable Lego bricks of computer reasoning: One set that understands written words, one that grasps physics, and so on.

Yejin Choi, one of Etzioni’s leading common-­sense scientists, has led several of these crowdsourced efforts. In one project, she wanted to develop an AI that would understand the intent or emotion implied by a person’s actions or statements. She started by examining thousands of online stories, blogs, and idiom entries in Wiktionary and extracting “phrasal events,” such as the sentence “Jeff punches Roger’s lights out.” Then she’d anonymize each phrase—“Person X punches Person Y’s lights out”—and ask the Turkers to describe the intent of Person X: Why did they do that? When she had gathered 25,000 of these marked-up sentences, she used them to train a machine-learning system to analyze sentences it had never seen before and infer the emotion or intent of the subject.


The WIRED Guide to Artificial Intelligence

At best, the new system worked only half the time. But when it did, it evinced some very humanlike perception: Given a sentence like “Oren cooked Thanksgiving dinner,” it predicted that Oren was trying to impress his family. “We can also reason about others’ reactions, even if they’re not mentioned,” Choi notes. “So X’s family probably feel impressed and loved.” Another system her team built used Turkers to mark up the psychological states of people in stories; the resulting system could also draw some sharp inferences when given a new situation. It was told, for instance, about a music instructor getting angry at his band’s lousy performance and that “the instructor was furious and threw his chair.” The AI predicted that the musicians would “feel fear afterwards,” even though the story doesn’t explicitly say so.

Choi, Etzioni, and their colleagues aren’t abandoning deep learning. Indeed, they regard it as a very useful tool. But they don’t think there is a shortcut to the laborious task of coaxing people to explicitly state the weird, invisible, implied knowledge we all possess. Deep learning is garbage in, garbage out. Merely feeding a neural net tons of news articles isn’t enough, because it wouldn’t pick up on the unstated knowledge, the obvious stuff that writers didn’t bother to mention. As Choi puts it, “People don’t say ‘My house is bigger than me.’ ” To help tackle this problem, she had the Turkers analyze the physical relationships implied by 1,100 common verbs, such as “X threw Y.” That, in turn, allowed for a simple statistical model that could take the sentence “Oren threw the ball” and infer that the ball must be smaller than Oren.

Another challenge is visual reasoning. Aniruddha Kembhavi, another of Etzioni’s AI scientists, shows me a virtual robot wandering around an onscreen house. Other Allen Institute scientists built the Sims-like house, filling it with everyday items and realistic physics—kitchen cupboards full of dishes, couches that can be pushed around. Then they designed the robot, which looks like a dark gray garbage canister with arms, and told it to hunt down certain items. After thousands of tasks, the neural net gains a basic grounding in real-life facts.

“What this agent has learned is, when you ask it ‘Do I have tomatoes?’ it doesn’t go and open all the cabinets. It prefers to open the fridge,” Kembhavi says. “Or if you say ‘Find me my keys,’ it doesn’t try to pick up the television. It just looks behind the television. It has learned that TVs aren’t usually picked up.”

Etzioni and his colleagues hope that these various components—Choi’s language reasoning, the visual thinking, other work they’re doing on getting an AI to grasp textbook science information—can all eventually be combined. But how long will it take, and what will the final products look like? They don’t know. The common-sense systems they’re building still make mistakes, sometimes more than half the time. Choi estimates she’ll need around a million crowdsourced human statements as she trains her various language-parsing AIs. Building common sense, it would seem, is uncommonly hard.

Source :

How to Teach Artificial Intelligence Some Common Sense
How Google’s AI Beat a Human at ‘Go’ a Decade Earlier Than Expected
Why data is the new oil: What we mean when we talk about “deep learning”
Artificial intelligence made by Google mimics human brain cells used for navigation — and takes shortcuts
Meet Google DeepMind's AlphaGo, A Program That Beats Professional Player at Go
What most people don’t understand about AI and the the state of machine learning
DeepMind’s New AI Taught Itself to Be the World’s Greatest Go Player
What is Deep Learning and how does it work?
Google’s AI beats a professional Go player, an industry first
What do games tell us about intelligence?