DARPA ottaa HTM:n käyttöön

"ORBIT's automated, 3D, object-recognition capability will help eliminate the time analysts spend manually identifying objects," said Dr. Peter Bilazarian, ORBIT program manager, Lockheed Martin Advanced Technology Laboratories (ATL). "We think ORBIT will reduce analysis time of one square kilometer of imagery from 1,300 hours to less than 10 hours. Faster turnaround time for analysts means more timely and accurate mission planning."

Central to ORBIT are three complementary brain-inspired and machine- learning approaches to object recognition: Numenta's Hierarchical Temporal Memory technology, which performs invariant pattern learning based on a model of the brain's neocortex; a standard model of the brain's visual cortex for spatial recognition; and a computer-vision technology, which mimics the process of human vision and recognition.

ORBIT's object-recognition algorithms will recognize images as humans do -- based on low-level and complex patterns, object shape and texture, and features common to a class of objects. ORBIT will also recognize objects based on position in the landscape -- called contextual spatial analysis. For example: An object in the middle of the street is not a mailbox because mailboxes are not found in the middle of streets.

Bilazarian said that ORBIT will determine if an area is of interest, propose potential recognitions of the area, and use previous recognitions to improve the accuracy of object recognition. Analysts could then use previously identified objects to track changes in landscape over time or show possible hostile activity: For example: An image associated with trash may contain an improvised explosive device.