Air Force turns to Qylur for AI that monitors autonomous vehicles 

In an increasingly connected world of autonomous vehicles and edge devices, armed forces around the world are seeking to improve the coordination and performance of complex systems. To this end, the U.S. Air Force has signed a contract with tech contractor Qylur Intelligent Systems for an AI-based “Collaborative Autonomous System” that would help maintain the data layer and performance of groups of autonomous vehicles over time. 

The contract will fund research and development into Qylur’s “Social Network of Intelligent Machines (SNIM)” AI — “a patented, core technology for ongoing management of autonomous intelligent devices and for maintaining the long-term superiority of their AI performance,” according to Qylur’s news release. But the company is aiming for commercial applications as well. 

“This is a core technology that we’re putting inside our own systems,” said Qylur CEO Lisa Dolev in a phone call with VentureBeat. “We’re working to go into this world of defense and be helpful as we can to win any advantage for our country. On the commercial side of it, [the technology] can be applied in autonomous cars, autonomous agriculture machines, home robotics — even in medical nano machines.” 

Qylur’s software stakes its claims on solving the challenges associated with the deployment of on-device AI. SNIM AI provides a performance-monitoring layer to the equipment found on the edges of the network, such as industrial robotics for private companies or drones for the Air Force. 

Founded in 2005 by Dolev, Qylur was previously in the business of venue and event security technology, producing the Q Entry Experience, a honeycomb-shaped bag scanner.

Small batteries and small datasets: challenges to overcome

Qylur’s equipment was deployed at the 2016 Rio Olympics and San Francisco’s Levi’s Stadium, providing insights that allowed the company to discover an obstacle often faced when deploying remote-sensing devices and mobile equipment: small data sets available to train models. Qylur’s initial products in the security space sought to detect guns and explosives, but the actual event of someone trying to hide weapons happened very rarely. It needed a solution.

Much like the more familiar online social networks, SNIM AI connects groups of related devices which then use the same set of shared data. Qylur says these pools of resources optimize the accuracy of decision-making and speed up real-world adaptations of the models. These features are relevant to both combat arenas and industrial use cases, as either can be fast-moving, changing environments.

Edge devices are limited by battery power and low processing ability when compared to more centralized infrastructure. Qylur’s SNIM AI seeks to alleviate those impediments by tailoring models to specific use and deployment cases. “[SNIM] allows you to have these specialized and customized mission-specific models, so you don’t have to have everything there all the time,” said Dolev.

Early experience revealed AI model drift

An AI model’s behavior can also change over time, making it necessary to address “AI drift” in ongoing operations. “If you have something that you think is working nearly perfect, in a few weeks, it could not be working perfect — it could be going a little haywire,” said Dolev.

The SNIM mitigates AI drift by automatically detecting it, and in response, retraining custom-boosted models and redeploying them to the edge devices.

“We’ve understood very early on that this drift happens, and you have to manage it all of the time. Not just once or twice — all the time,” said Dolev. “Managing it can be extremely expensive if you need a whole bunch of data scientists and an ML ops. So [SNIM AI is] an automatic way to do that.”

The goal of the partnership with the military is to allow Qylur to adapt the SNIM AI technology into fulfilling the Air Force’s use cases. Qylur also sees ongoing commercialization of SNIM AI as critical next steps to equipping AI-enabled devices in the field. 

“The easiest thing would be to say this is kind of like an AI for the AI; the guardrails for AI — like a gardener for it,” said Dolev.