Air Force wants smart system that can learn how to detect threats from sensor data

The Air Force wants smart software that can help analysts identify targets from a disparate patchwork of high and low resolution imagery data, a solicitation reveals. The hope is that the technology can be used by counterinsurgency teams to detect improvised explosive devices and ground threats.

The Pentagon is specifically interested in “machine learning” technology -- artificial intelligence capabilities that allow computers to recognize complex patterns and make intelligent decisions with repeated use. (The most famous example of machine learning technology at work would be Watson, an IBM-built computer system that answers questions and has formidable Jeopardy skills.)

Defense envisions that the systems will get smarter with time if taught how to model the way in which humans detect anomalous targets. “The activity detection system has an opportunity to learn from the detections provided by the users,” the solicitation says.

By outsourcing some of the work to computers, military officials expect analysts will be able to avoid mistakes and make faster decisions, such as calling for a high-resolution high-revisit of data after some low-resolution imagery has indicated potential threats.

Defense wants the machines to be able to fill in the blanks from what they’ve “learned” over time when data proves spotty. The machines also will be expected to “monitor network performance under varying sensor output and user requests, learn to predict future bottlenecks, and develop proactive network management and prioritization policies,” the contracting document says.

The Air Force is soliciting proposals between Aug. 27 and Sept. 6. The algorithms created for the software need to factor in the varying quality of sensor imagery and a limited set of data from which machines can learn from, according to the solicitation. Scientists that make it to the final phase of funding will deploy the prototype with ground stations.