DARPA to Bring Chattiness to Computers
Give researchers about five years, and people will be communicating with machines—a far cry from today’s clicking, swiping and typing, says one scientist working with the Defense Department’s futuristic research arm. While improvements in algorithms helped develop artificial intelligence over the past several years so vast amounts of data from videos, signals and human intelligence could be deciphered quickly, the progress amounts to little more than fancy math.
In spite of all the technological advances and everything computers can do for people, the devices used today do not actually understand language. People increasingly rely on computers and machines to learn, to help navigate, to stay connected with family, to travel—to make life easier. In reality, people do not communicate with machines, and the relationship is one-sided. The two parties are not partners—yet.
“We have almost no way to communicate with computers these days,” says Paul Cohen, a program manager in the Information Innovation Office at the Defense Advanced Research Projects Agency (DARPA). “Sure, you can point and click and swipe, or you can type keywords into a text box, … but there is nothing like having a conversation with a computer or working with a computer to solve a problem or really having a computer be a full partner in your activities.” Today’s interactions with computers are asymmetric.
“We view computers, even today, the same way we view power tools,” explains Cohen, who joined DARPA as a program manager in September 2013 from the University of Arizona, where he is a professor and founding director of the School of Information: Science, Technology and Arts. “They do our bidding. We use them for very specific purposes, but we don’t ask their opinions. We basically treat them as, well, I guess as servants of one kind or another, but certainly not as partners.”
DARPA researchers want to achieve with machines the human-to-human standard of sharing ideas through language. “What would it take to unlock the problem-solving abilities of machines—which are considerable—and to take full advantage of them as partners?” Cohen asks. “Machines are awfully good at some things, but to fully exploit what they’re good at, it’s actually necessary to communicate with them.”
For example, cellphones today might have more computing power than NASA systems used to launch the first space shuttle, but in the end, the embedded alarm clock still is just an alarm clock. It can wake the user only at a preselected time. “I’d like to say to my cellphone: ‘I have to catch a 10 a.m. flight. What time should I get up?’ The moment the machine tries to help me, it has to ask me some questions, and I have to be able to answer them. And we like to do that in language because that’s the natural mode for us,” Cohen explains.
The advancements are far from simply asking Siri to search the Web for an answer to a query or building computers that can act like humans. The research tiptoes right up the fine line of not only putting complete trust in computers to precisely and continuously make the right predictions for people, but also to teach them to genuinely think on their own. To that end, DARPA is funding three case studies in its Communicating with Computers (CwC) program, which seeks to enable symmetric communication between people and computers so that the machines serve as more than receivers of instructions. Cohen envisions machines as collaborators that truly can understand a full range of communication methods, from language to gestures and facial expressions. “If we can communicate with machines, then we could expect to unlock their problem-solving ability,” he says.
The most challenging study is the effort to teach computers to develop stories with humans, adapting an old pastime of having storytellers take turns contributing a sentence each to weave a tale. “The strongest test of whether we can communicate with computers … is that a human and a machine will collaboratively tell a story, one sentence at a time, taking turns,” Cohen expresses. “One of the reasons that is a very hard test is the common sense versus knowledge problem. Since we’re not constraining what the story could be about, the machine really has to know about all sorts of things that humans know about, and that’s going to be very, very challenging. It not only has to know all of the common sense stuff, but it has to know what makes for a good story.”
The Internet today is an enormous repository of readily available information—the most ever recorded in human history—and it is growing every day. “I’m hoping this will promote a lot of research on understanding the enormous amount of human knowledge that is encoded in text that’s already out there on the Web,” Cohen continues. “I think machines can go a long way to knowing what people know if they could just understand what’s out there on the Web. That means we have to make some pretty significant progress in language understanding.”
In reality, language is rather ambiguous. “Language rarely specifies ideas perfectly, and we rely very heavily as humans on context to help us understand what ideas actually mean,” Cohen explains. “That’s another area that has been completely neglected in artificial intelligence and cognitive science. But it will get a lot of play in this program.”
The expression of ideas can be taught, Cohen notes. But how do you teach common sense? Researchers at Carnegie Mellon University are trying to answer the question and have invested years in a program called Never-Ending Language Learning (NELL), an effort that uses machine-learning technologies to research and extract information from hundreds of thousands of Web pages to teach computers to read. Since 2010, NELL has run continuously to “read” the Web from hundreds of millions of pages. People interested in tracking the program’s progress can follow on Twitter at @cmunell.
Perhaps the easiest to achieve of the CwC cases studies is what researchers call Blocks World, where a human and a machine collaboratively will build something out of wooden blocks, talking about what they are doing and how to do it throughout the process. “That’s a case where we’re taking state-of-the-art computer visions, state-of-the-art robotics, state-of-the-art language understanding, and we’re tying it all together into this ‘let’s build something together on a tabletop’ kind of communication,” Cohen says.
The experiment will factor in tone, gestures, eye movements, facial expressions, pointing and body posture. “They all go a long way to helping a person understand what you mean,” Cohen says. “Although language is the principal mode for communication, it is by no means the only one, and we really want to get at that in the Blocks World challenge problem.”
Sandwiched between the storytelling experiment and Blocks World is a case study in which humans and machines collaborate on a project called “biocuration,” or building models of complicated biological processes. It borrows heavily from another DARPA effort called Big Mechanism, in which computers mine and correlate big data. A focus area is cancer research, where computers mine for data in numerous medical studies, break down the information and its constituent parts, and build a model that could point to a possible cure.
Big Mechanism, started nearly a year ago and involving roughly 100 researchers from dozens of groups, steps in at a point where humans have reached the limits of the ability to understand complicated things. “The reason we started Big Mechanism is that humans are awfully good at interfering with really complicated systems that they don’t understand very well,” Cohen concedes. “And that would include the climate and the economy and things like that. These systems are very complicated, and the reasons that things happen in these systems are often hard to explain. Should we lower interest rates? Should we throw iron filings into the Pacific Ocean to promote algal blooms? ... The real reason for the Big Mechanism program is that, in this day and age, our survival depends on really, really complicated systems, and no human being really understands any of those systems in its entirety.”
The researchers’ work does not ignore the potential social ills that could emerge from programs left unchecked or overlook ethical challenges. Generally, the agency relies on internal review boards and external consultants to ensure advances adhere to ethical standards, DARPA Director Arati Prabhakar said during a news media briefing held to highlight agency projects for the coming years. “Because we find that our job is to push these boundaries [of breakthrough technologies], we often stumble into these really vast societal questions about how these technologies ultimately will get used,” Prabhakar said. The agency does play a role in sorting out societal issues, but,“It’s very important that we not shy away simply because there are uncomfortable questions. … We have some important choices to make about where we put our next dollar in each of these fields. In the research enterprise, it’s usually fairly clear what is and isn’t OK legally and from a regulation point of view.”
DARPA relies on advice from peers and graduate students in various fields for guidance on “ethical issues and societal implications that is part and parcel of thinking about the research that they do,” according to Prabhakar.
DARPA’s recent broad agency announcement seeks a good mix of program participants from academia, industry and military service laboratories. “We are looking for people who are really good at language understanding, who know something about the structure of stories and rhetoric,” Cohen shares. “We are looking for people who are able to detect gestures and … subtle body language cues. We’ll be looking for people who are really good at machine learning, people who are experts at how to represent knowledge in a way that machines can understand it.
“I think it’s quite likely that, if we succeed, even a little bit, at building machines that can communicate with people, people will like those machines and like them genuinely,” he predicts.
