Poisoned datasets, Trojan horses, and ever-changing cyber threats could derail the Army's plans to widely adopt AI. However, the new 100-day plan aims to eradicate bottlenecks and prepare the service to extend third-party models and algorithms.
The plan, announced in March by the Army's assistant secretary for acquisition, logistics and technology, allows the Army to emulate the way commercial developers work so services and industry partners can deliver features faster. The announcement comes after the publication of a software directive aimed at supporting.
“The Army recognizes that we can't develop algorithms and models and train better than industry. We want industry to do it. And they do it very well. So we want to adopt a lot of that. One of the hurdles to adoption is how do we consider the risks around AI?” Young Bang told reporters on Monday. “What are the problems with contaminated datasets, adversarial attacks, Trojan horses… And development is easier if you develop in a controlled, trusted environment, owned by the Department of Defense or the Army, for example. Become.”
These efforts will ultimately align with the Army's ongoing work to simplify how AI is used, particularly in intelligence systems, through a program called Project Lynchpin.
As the Army develops its lynch pin environment, “many programs are now developing, training, deploying and testing their own models,” Bunn said.
Lynchpin is designed to be a digital environment with vetted tools, infrastructure, standards, potential use cases for AI, and relevant data, said Linchpin, director of the Army's Project Lynchpin and Sensor AI program. said one Bharat Patel. And once he brings them all together in one place, project managers should be able to easily incorporate artificial intelligence into programs that can take advantage of it.
LynchPin is currently focused on defining these use cases and preparing the data associated with them, said Patel, who is also product lead for the intelligence, electronic warfare and sensors program office.
“We're doing basic model training. But it's not for performance or anything. We're trying to understand what the process is, what the governance is, what the heck. [are] Our standards allow us to do this much faster. So our objective is really to kind of learn things and see if we're ready to execute after the 500-day implementation plan and contracts and everything else is in place,” Patel said. said.
And as Linchpin matures, we plan to feed algorithms from other programs and third parties so program managers can pick and choose features from which to build. Mr Bunn said details of Lynchpin's readiness would be announced after a 100-day plan, likely to be completed this summer, followed by a 500-day plan.
“Each theater is different,” Patel said. “You can't think of a model” [European Command] Ready to use [Indo-Pacific Command]. Trees and biospheres are different… That’s why understanding the use case is so important.and where is it [area of responsibility] Specifically. So we're watching it very closely. ”
