17:44 GMT +327 October 2016
    General Atomics MQ-9 Reaper

    DARPA’s New 'Neural' Microchip Could Let Drones Think Like a Human

    © Wikipedia/
    Get short URL

    DARPA, the Pentagon’s cutting-edge science division, is developing a neural network to enhance “deep learning” on the battlefield.

    Modern warfare requires more than simple brute strength. The Information Age has revealed new ways to communicate and access information, and these new rules have also affected the way wars are fought and won.

    Imagery gathered by satellites and drones must make its way to units on the ground for data analyzation. While a drone can take a photograph, it cannot tell you what it’s looking at, or what to do about it, if anything. A human brain is necessary to spot a terrorist stronghold, or to recognize that a missile launcher in one photo has been moved, when compared with an older photo.

    "Full exploitation of this information is a major challenge," officials with the Defense Advanced Research Projects Agency (DARPA) wrote in a 2009 brief on "deep learning."

    "Human observation and analysis of [intelligence, surveillance and reconnaissance] assets is essential, but the training of humans is both expensive and time-consuming. Human performance also varies due to individuals’ capabilities and training, fatigue, boredom, and human attentional capacity."

    Working with a team of researchers at MIT, DARPA is hoping to take all of that human know-how and shrink it down into processing unit no bigger than your cellphone, using a microchip known as "Eyeriss." The concept relies on "neural networks;" computerized memory networks based on the workings of the human brain.

    A palm-sized neural network chip could be installed in drones or satellites, allowing these units to conduct their own learning in real time, without the need for human analysis.

    Instead of a team of individuals combing through imagery looking for a single target, a drone could simply alert soldiers on the ground once the drone has identified a target.

    The technology could also work in disaster zones, allowing drones to spot and identify people in need, and then communicate location and other data to aid workers.

    Current deep learning technology requires a large number of servers and the energy necessary to run those computers. Data can be sent to warehouses containing the computers for analyzation, but that requires an Internet connection, which isn’t always readily available in combat situations and, when it is, is not always secure.

    But Eyeriss could change the way the game of war is played. Packing more processing power into a much smaller space, the microchip could allow our handheld devices to become even smaller, and allow drones and satellites to operate without a need for massive server warehouses or hundreds of human analysts.


    100s of Those Killed by US Drone Strikes in Pakistan Remain Unidentified
    Absolute Anti-Drone Shield? Company Creates UAV-Killing System
    At Least 20 Major US Drone Accidents ‘Worst Annual Toll Ever’ in 2015
    neural networks, deep learning, artificial intelligence, drones, Defense Advanced Research Projects Agency (DARPA), MIT, Pentagon, United States
    Community standardsDiscussion
    Comment via FacebookComment via Sputnik
    • Сomment

    All comments

    • michael
      I will be interested to see what they come up with, but at the moment some of their terms are a little fuzzy (as in the logic).
    • Ivan Buckeye
      Only a piece of human thought. What piece, you might ask? The piece of human thought involving aiming and shooting. After the fact thinking.
    • JPHin reply tomichael(Show commentHide comment)
      Neural network pattern recognition actually removes most of the programming and replaces that with training a network on an extensive database of patterns to be recognized. At this point of time this approach has become the favorite approach to achieving autonomous self driving cars.
      Do realize that a lot of human labor depends on some kind of pattern recognition and acting upon that. So DARPA is trying to apply this commercially available technology (see NVIDIA's Titan and Drive PX 2 ) to the military arena. Target detection and identification is such an obvious application required for autonomous systems like Russia is experimenting with too.
      More info:
    • michaelin reply toJPH(Show commentHide comment)
      JPH, thanks for the links and I will be checking them out! I realise that pattern recognition is the key here, I am more interested to see how the db can be used by the system to adapt to unforeseen circumstances or conditions effectively. In other words, how it shall avoid errors. Thanks :)
    Show new comments (0)

    Top stories