learning abilities and human instruction to turn bot tots into supersmart A.I. agents “that [can] learn and be taught like a human.”

Darpa, the Pentagon’s far-out research arm, wants robots that can outdo (or at least match) human smarts, from data analysis to C3P0-esque language translation. Advanced A.I. was also the overarching goal of their Biologically Inspired Cognitive Architectures (BICA) program, which sought to mimic the physiological and neurological elements of the human mind.

But given the agency’s lofty A.I. goals, they also needed a better way to determine just how well their ‘bots mimicked human functioning in the first place. Enter Dr. Shane Mueller, a senior research scientist at Applied Research Associates, Inc. and an expert in cognition, perception and memory.

In 2008, Mueller and a team of researchers developed a program on devising and testing ‘bots as part of Darpa’s BICA program. Their paper, which was released online last week after a Freedom of Information request, wants to see military scientists start young — literally — by rearing robots with the capabilities of an average toddler.

“There were many motivations for this target, but one central notion is that if one could design a system with the capabilities of a two-year-old, it might be possible to essentially grow a three-year-old, given realistic experiences in a simulated environment,” Mueller writes.

To test how well the bots stacked up against their pint-size human peers, Mueller designed a decathlon of evaluations that expand on the Turing test,which was developed in 1950 as a means of testing the linguistic capacities of an A.I. agent compared to an intelligent human.

Of course, A.I. agents have yet to ace the original Turing test — and it’s applicability is largely considered irrelevant in modern A.I. initiatives, because the test demands much of a robot’s linguistic capacities without evaluating a wider range of abilities (like visual analysis or motor skills).

Mueller’s testing schema is more more involved, and includes categories for visual recognition, search abilities, manual control, knowledge learning, language and concept learning and simple motor control (including eye movement). And while the initial expectation is that A.I. agents operate much like a tot — capable of, for example, finding a red ball amid a room of objects — they’d gradually learn from both surroundings and an instructor, advancing to more advanced cognitive capacities.

For example, a ‘bot would gradually learn to move “[by] replicat[ing] hand movements of an instructor (with identical embodiment), including moving fingers, rotating hands, moving arms, touching locations, etc.”

Darpa’s BICA program never took off, but that hasn’t stopped Mueller and others from continuing to refine the cognitive duathlon approach, sharing progress on a web-based forum. Now, after testing several components of the battery, Mueller’s published his latest paper. The results? Prepare for ‘bots that’ll outdo you, even when they make mistakes: humans boast myriad strengths over ‘bots, but as training improves, Mueller notes that biologically-inspired ‘bots will one day mimic human behavior…flaws and all.

Take human eyes, for example: a lack of calibration could be “sidestepped by an artificial system.” Or, one day, strategically imitated:

“But the lack of calibration turns out to be incredibly useful because we are able to perform shape-based matching to objects regardless of size or orientation…AI systems that have taken this biologically-inspired skill seriously have exhibited impressive performance…This phenomena is likely to recur across a wide range of human skills: when consistent errors or imprecisions are found in human performance, it may go hand-in-hand with a flexibility and robustness, such that if our performance were better, it would also be more brittle.”

Photo: UConn

See Also:

• Darpa: Heat + Energy = Brains. Now Make Us Some.
• Darpa Wants Self-Guiding, Storytelling Cameras
• Darpa Wants Brainy Machines to Replace Bored G.I.s
• Darpa Plan: Turn Warzone Data into Simple Stories
• Air Force Looks for ‘Core Algorithms’ of Human Thought
• Inside the Rise of the Warbots

Authors: Katie Drummond

to know more click here