The Defense Advanced Research Projects Agency made headlines last fall when it announced that it was pledging $2 billion for a multi-year effort to develop new artificial intelligence technology.
Illustration: Getty
Months later, DARPA’s “AI Next” program is already bearing fruit, said Peter Highnam, the agency’s deputy director.
DARPA — which has for decades fostered some of the Pentagon’s most cutting-edge capabilities — breaks down AI technology development into three distinct waves, he said during a meeting with reporters in Washington, D.C.
The first wave, “describe,” focused on developing platforms that employed a rules-based system. These AI platforms are the basis of commercial products such as TurboTax, he noted.
The second wave is “recognize,” and is made up of the machine learning systems that are prevalent today. Such systems can classify objects of interest and take the burden off of human analysts who often must pore through mounds of data and turn it into actionable information. However, while the theory behind these second wave technologies was established in the 1970s, much more work still needs to be done to mature them, he added.
The third wave, “explain,” is where the future of AI is headed and focuses on adding context and trust to artificial intelligence platforms, Highnam said.
“We haven’t completed the second wave of technologies, … there’s a lot more to do,” he said. “‘Explain’ is where we’re going now.”
DARPA’s AI Next program has three thrusts: to increase the robustness of second wave AI technologies, to aggressively apply second wave systems to new applications and to further examine third wave technologies, he said.
Second wave AI technologies still lack robust underpinnings, he noted.
“We have a lot of really good examples of successes, but the notion of being able to use a second wave technology in a safety critical situation on its own isn’t there yet,” he said.
“We have a lot of robustness work to do, a lot of basic theory, a lot of AI system engineering to be developed.”
The agency has been employing a variety of contracting methods to get at the development of new second and third wave systems, he said. It has already released 10 broad agency announcements for AI-related programs that will require major research proposals and multi-year work.
But alongside that are what are known as AI exploration, or AIE, awards which are meant to fast-track new and emerging technologies, Highnam said.
“We don’t know what the third wave is going to look like yet, so at that intersection, we have … AI exploration awards,” he said. The agency posts a topic of interest on its website and interested parties are given 30 days to respond. Within 60 days of that deadline, DARPA awards an other transaction authority agreement, which has become an increasingly popular contracting method as acquisition officials look to cut through bureaucratic red tape. The process is completed within 90 days.
“If you deal with government acquisition, you know this is fast,” Highnam said. “But it’s being done very intentionally.”
AIE awards are worth up to $1 million each and last between 18 months and two years. DARPA has already awarded about $50 million since September, he said.
This is a particularly appealing approach for those in academia, Highnam said.
“We have huge interest in the research community because if you’re a graduate student, you can think about … entering a master’s program and being involved in a proposal, getting the funds to proceed, [purchasing] new equipment and so on, within three months of starting,” he said. “You can actually do [that] within the life of your one- or two-year degree.”
Small businesses have also been keen on the AIE awards, he said during a hearing on artificial intelligence in March before the Senate Armed Services subcommittee on emerging threats and capabilities.
AIE program topics include a variety of efforts such as examining how to apply physics and social sciences to artificial intelligence, Highnam said.
Neural technology is another area of interest, he added. For example, AI can be employed when conducting mapping when neurons fire signals in the brain to sensors in a system controlling an arm and back again.
“Those are machine learned interfaces and they have to be recalibrated continuously,” he said. Utilizing artificial intelligence is “one way of doing it.”
Highnam noted that about one-third of DARPA’s 250 programs include building new AI technologies, making such systems more robust or applying them to new areas.
“It’s not a casual thing for us really across the board now,” he added.
As DARPA develops new technology alongside its industry partners, one key aspect it considers is how to transition a product to the armed services, Highnam said.
“We want industry to make these technologies as they’re proven to work, to be commercial, to be incorporated into other products, that then the [Defense] Department can buy back,” he said.
The recent establishment of the Pentagon’s Joint Artificial Intelligence Center will make it even more likely that AI technology developed at DARPA will move forward and be transferred to the services, he said.
AI technologies often have many sustainment and deployment challenges, he said. “I personally, as an R&D guy, am really happy to now have the JAIC stand up as a partner to take on … that engineering, deployment, sustainment tail because … I expect it will make transitions … more straightforward.”