What Keeps AI Experts Up at Night?


The Defense Advanced Research Projects Agency's (DARPA) $2.9 billion budget is funneled into many programs, grouped around complex military systems, the information explosion, biology as technology, and expanding the technological frontier.


To this end, DARPA gathered 1,000 scientists, engineers, and innovators in St. Louis recently. On one panel, two veterans of artificial intelligence, Dr. Paul Cohen and Dr. Tom Dietterich, infused their serious examination of the space with a bit of insider geek humor. Cohen, who joined DARPA as a program manager in 2013 from the University of Arizona, sees a future where—due to collaborative computing—humans and machines can work alongside each other in mutual understanding, curing cancer and perhaps even creative works of music and art.

"But machines that can truly communicate will be uncomfortably human," Cohen said with a smile. "Did Turing imagine that AI machines will also test us?"
Dietterich, who serves as president of the Association for the Advancement of Artificial Intelligence and teaches computer science at Oregon State University, has concerns.
"I don't believe we should build completely autonomous AI for high stakes scenarios," he told the assembled crowd, saying that too much can go wrong without human oversight. When someone in the audience mentioned being in a store where they were told "We can't process your card because the machine won't let me do that," Dietterich quipped, "That's why I carry cash." This got a big laugh, which quickly died, when many considered the implications of an AI expert saying such a thing.

Dietterich saw the future of AI continuing in two distinct branches: Tool AI (Siri, Cortana, Google Now) and Autonomous AI (self-driving cars, automatic weaponry, smart power grids).

When someone asked when the Singularity was due, Dietterich suggested that "Tool AI is already smarter than most humans." He pointed to the rapid progress in deep neural networks that underpin Google speech recognition, an AI system has gone from a 23 percent word error in 2013 to just 8 percent in 2015.

"There's not a librarian in the world that can give you as much information as Google today," he pointed out.

Google might have all the answers, but someone has to put it there. To that end, robotics expert Yolanda Gil from USC is building intelligent scientist assistants.

"Scientists are so specialized now that they rarely communicate across disciplines, which all use differing sets of assumptions and knowledge bases," she said during a later panel.

To rectify this, her team at USC focuses on AI systems that can assist, say a biologist, by retaining knowledge and the latest data across multiple scientific fields, including neuroscience and advanced chemistry, while understanding the algorithms underlying and cojoining them.

Gil appeared with Cornell's Hadas Kress-Gazit, who is focusing on robot controllers for complex, high-level tasks. Simply put, "I'm making systems that anyone can program by selecting different components," she said.

She's also building robots that can talk back. "These are robots that can communicate directly to humans using computational linguistics. For example, being set a task and saying,'I can't do this because my sensors need replacing' and so on."

The Next Generation
As part of the event, DARPA spent half a day evaluating more than 50 promising early career scientists to identify DARPA Risers on whom the agency will be keeping close tabs for 2030 and beyond.

One of those Risers was Anupama Lakshmanan from Caltech, who spoke about her research into cellular agents for non-invasive brain disorders. The agents enter the brain and then spontaneously self destruct at end of surgery.

"Imagine something a little like a continuous movie showing the inside of the brain and how cells interact," she told the crowd. Her work—if it comes to fruition and then commercial exploitation—may prove to be instrumental in developing peptide-based therapeutics for amyloid-associated diseases, like Alzheimer's.

In the DARPA demos arena, meanwhile, Vietnam Vet Fred Downs, who was wounded in combat, showed his prosthetic arm, attached to his short left stump. Wearing a harness with a battery pack and sensors, Downs makes movements with his feet and torso to cycle through the sequence of hand grips and motions (video below).

"This arm enables me to live normally and tinker in my workshop, where I work with wood mostly, making stuff," he told PCMag. "In fact this prosthetic arm sometimes works better than my right arm, which I also nearly lost in Vietnam."

As the day wound down, back in the demo hall, there was an impromptu sweet photocall. All the disaster relief robots from the DARPA Robotics Challenge, which took place in Pomona, California earlier this year, were moved into a line next to each other with their end effectors touching as if gathered for a team photo.

Of course the robots had they done this themselves; team members from various countries posed them together. So much for dastardly competition between nation states and/or academic institutions. But that was the purpose of DARPA's event. Only through collaboration and ensuring a steady stream of smart research programs can the U.S. retain military supremacy and national security, after all.

Source: pcmag.com
Share on Google Plus
    Blogger Comment

0 comments:

Post a Comment