An Arlington, Va.-based company called Deep Learning Analytics has developed a deep machine learning automatic target recognition (ATR) program prototype for DARPA that will help assist aircraft pilots in finding targets.
Deep Learning Analytics is nine months into a 40-month contract for DARPA’s Target Recognition and Adaptation in Contested Environments (TRACE) program, according to company founder and data scientist John Kaufhold. TRACE seeks to develop an accurate, real-time, low-power target recognition system that can be co-located with the radar to provide responsive long-range targeting for tactical airborne surveillance and strike application.
DARPA, on its website, said adversaries have an advantage of using sophisticated decoys and background traffic n target-dense environments to reduce the effectiveness of existing ATR solutions. Airborne strike operations against movable targets require pilots to fly close enough to obtain confirmatory visual identification before weapon release, putting manned platforms at extreme risk.
Radar, DARPA said, provides a means for imaging ground targets at safer, and far greater, standoff distances, but the false-alarm rate of both human and machine-based radar image recognition remains unacceptably high. Existing ATR algorithms also require impractically large computing resources for airborne applications. Thus, current approaches for inserting ATR into tactical applications either move the processing to remote ground stations or drastically reduce performance to fit legacy airborne platform computing capabilities, according to DARPA.
Kaufhold said Deep Learning Analytics’ prototype runs on a commercial-off-the-shelf (COTS) NVIDIA [NVDA] TX1 processor. Kaufhold said the processor can process about 200 frames per second, which he said was not available anywhere else.
Kaufhold said pilots currently spend a lot of time locating targets, which is complicated by having to manually maneuver a joystick and zoom in, all while flying at aircraft speeds, which puts pilots “in another county” by the time they make a decision. Kaufhold said the goal of TRACE is: better, faster and adaptive.
Better, he said, means spending less time looking at a part of an image where there aren’t targets, while lowering the false alarm rate. Faster, Kaufman said, means running the program on a three point device that runs on 10 to 20 watts power, an order of magnitude better than a legacy product that would have to use a down link to a ground control station and further more to a device that maybe weight tens of pounds and requires higher power.
Kaufhold said adaptive means leveraging advantages of deep learning to improve targeting in limited amounts of data. He said one problem with machine learning is that it needs training data, and pilots in might not have training data for all types of targets, especially in contested airspace. Kaufhold said Deep Learning Analytics is leveraging deep learning to better, and faster, identify platforms like tanks. Kaufhold said machine learning is all about finding patterns in data.
Kaufhold said Deep Learning Analytics won this DARPA contract in a competitive procurement. Though he didn’t know while speaking to Defense Daily at the Pentagon’s DARPA Day how much the contract was worth, the company’s website says it was worth $6 million.
The contract has a unique schedule as Kaufhold said DARPA will provide more challenging tests more challenging environment sand target types as the program develops. He said DARPA will make the program more difficult after nine months to see how performance may degrade, then re-evaluate after another nine months.