The Defense Department is looking to make artificial intelligence a central part of its mission to stay on top of emerging threats.
But as DoD prepares to meet its AI-readiness goals over the coming years, the department should set realistic expectations for what AI can do to improve performance, advised Rob Albritton, vice president and AI practice lead at Octo, which was recently acquired by IBM.
During Federal News Network’s DoD Cloud Exchange 2023, Albritton said industry as a whole has “sold a bill of goods to DoD and the government that isn’t true when it comes to AI” — especially that AI is a silver bullet to its problems.
“Those AI solutions require massive computational resources to process. Oftentimes, the AI solutions are designed only to be deployed on cloud resources, and the DoD — specifically, the Marine Corps, Army, Special Operations Command — those kinds of folks are operating around the world in places where they don’t have good comms bandwidth,” he said. “They don’t have computational resources. They’re not able to push data back to the cloud to crunch on it. So AI solutions need to be developed to work at the edge on low size, weight and power devices. And that’s not what’s being sold to the government.”
Proving out the use of AI at the edge
Some real-world tests are beginning to help address that challenge, said Mark Johnson, vice president of federal technology for IBM. He noted that the value proposition of AI for DoD is “all about filtering the information that the warfighters at the edge need.”
The emerging field of AI-assisted fighter jets, Humvees and ships is helping resolve a longstanding challenge: “How do you get the right information into the cockpit at the right time?” Johnson said
“Now, instead of just a bandwidth question, and a display question, we have the capability to filter that with an AI machine learning tool that does a much better job of both sending data to the edge and bringing data back from the edge — the right information,” he said.
While AI can help DOD enable data-driven decision-making, the department relies on a wide spectrum of automation tools to meet its mission.
“Not all problems are best suited to be solved with AI,” Albritton said. “Some are just a high-performance computing problem, where you’re just speeding up the analytics. Some are AI problems, where you’re doing predictive analytics and things like that. But customers have to understand their problem.”
He said Octo is able to walk agencies through what it takes to create an AI solution and whether a less sophisticated automation tool is better suited to streamline the task at hand.
“It takes a lot of data to train a machine learning model. So we take the customer through that journey of understanding whether or not they have the resources to actually solve their problem with AI,” he said. “Sometimes the answer is no, and they need to invest more in their infrastructure — in their pipelines, data mesh, data fabrics, things like that — to get data into their own ecosystem to train their models.”
Fine-tuning an agency’s AI approach
Johnson said IBM relies on input and feedback from its federal customers to provide the best AI and automation solutions for the right job.
“It’s a team sport. Whether it’s AI, getting the data out to the edge, security, all these things take a large ecosystem,” he said. “And from the technology side, we do work with the big providers — for DOD and for the federal government — to make sure that we’re listening to what the warfighter needs and then adapting our technologies.”
An agency’s capability to field AI solutions depends in large part on the state of its network infrastructure. DoD, by the nature of its missions, is best served by a hybrid cloud architecture, Johnson said.
“Any big enterprise, really, will always be hybrid cloud. Thinking that everything is going to go into the cloud, everything is going to be on the edge, everything is going to be in on-premise data center, it’s just not the modern world,” he said.
Although DoD has highly classified secure environments that belong in on-premise data centers, Johnson pointed it out that the department also has a considerable public-facing workload.
The decision on what can move the cloud and what can’t is “about a balance between the security necessary, the compute power necessary and the ease of access,” he said.
Albritton said AI and machine learning solutions “need to be done as close as possible to the edge” of the network. Without automation at the edge, DoD must spend months getting data off the battlefield and using it to train a machine learning model.
“It’s a multimonth process that’s far too long for our warfighter,” he said. “We have to be able to iterate on our machine learning models as quickly as possible, meaning we’ve got to get ML ops as close as possible to the edge.”
Using cloud as an AI testbed
Johnson added that the cloud also gives DoD a sandbox environment to train and experiment with AI solutions before they’re ready to be deployed.
“We have to be able to spin up environments, test things out, try it, see what results happen and then tear them down very quickly and put new ones in place,” he said.
Although AI is a key part of DoD’s information technology transformation journey, agencies are also taking steps to ensure their automations are secure, reliable and free of bias.
As DoD pours resources into the next-generation of explainable AI, agencies are seeking transparency into how the algorithms arrive at their decisions — and whether they will be able to vouch for the accuracy of those answers, Albritton said.
“That is critical. We need to be able to tell what that machine learning model is doing in the background. How’s it arriving at that end answer and, then, are there issues with it?” he said. “All machine learning models drift over time. They all degrade over time. There’s not one single one that doesn’t. We need to be able to monitor those when we’re using them. And a nontechnical user has to be able to understand when that model is not working properly.”