NEW YORK—The Defense Advanced Research Projects Agency is exploring unconventional processing methods to analyze mission-critical intelligence, surveillance and reconnaissance (ISR) data in the analog realm.
The agency has launched a program called UPSIDE, which stands for the Unconventional Processing of Signals for Intelligent Data Exploitation. Darpa’s motivation is to find processing techniques that do not have the same power and speed limitations associated with digital processors used in today’s ISR platforms.
The program employs probabilistic inference as its fundamental computational model to drive down power consumption. Inference can be implemented directly in approximate precision by traditional semiconductors as well as by new kinds of emerging devices. This approach aims to compute inference directly and more naturally through a number of interconnected analog nodes.
“A single inference operation performed in digital CMOS could involve hundreds of operations. However, the inference operation can be more naturally performed directly by a group of coupled analog devices, reducing both power and the number of discrete operations,” said Darpa program manager Dan Hammerstrom. “Utilizing coupled analog devices to represent data in a non-Boolean form for high-order computation is expected to provide orders of magnitude improvements in the power efficiency of data processing.”
According to an agency announcement, the definition of emerging devices is broad to allow developers to propose a wide range of parts. But they cannot be CMOS-based, at least not without additional specialized processing steps, and cannot be on the mainstream ITRS http://www.itrs.net/about.html roadmap. Analog-based devices such as memristors and spin torque oscillators could possibly be used to replace CMOS-based digital processors to accelerate video and imagery analysis in ISR systems, according to Hammerstrom.
The MIT Age Lab recently worked with Monotype Imaging in an investigation of effects of typeface on the demand of human-machine interactions during a simulated in-vehicle point-of-interest (POI) menu selection task. Their study showed that people could perform a menu selection task faster and with more accuracy depending upon the text type used as well as white text on a black background vs. the standard block text on a white background. Differences were clearly noted.