Army AI Challenges Mounting as Research Portfolio Grows

Army DoD military Defense AI

Brian Sadler, senior scientist for intelligent systems at the U.S. Army Research Laboratory, today discussed how the expanse of artificial intelligence (AI) research activities across the Army is bringing to light even greater challenges.

Sadler, speaking at an Army Science and Technology Symposium hosted by the National Defense Industrial Association, said the impact of AI on Army research has been tremendous, but noted that there are several bumps on that road–including networking needs, limitations of commercially-available technology, resourcing, training limitations, and various unknowns about the technology–which will need to be addressed in order to solve the tremendous problem sets and demanding needs of the Army’s mission.

“It’s safe to say that AI and machine learning is having a huge impact across all of our S&T, and I look at our whole research portfolio at the [Army Research Laboratory], it’s having an impact on everything. That’s not an overstatement,” he said.

But that impact is bringing with it a growing list of emerging challenges.

Networking and Autonomy Combined

Sadler pointed to the current reliance on cellular and wireless networks to fuel AI in practice. As computing needs expand, he said he’s looking at a way that blends autonomous systems seamlessly with networking.

“Since we’re totally relying on our wireless networks and we want to interact with autonomy, this means that we really need to think about combining the two,” he said.

“If we’re going to do distributed autonomy, we need networking. But if we’re going to do networking, we need autonomy,” Sadler continued. “I think we need to get our heads around the idea that we need to combine wireless networking and autonomy into one thing. Traditionally we don’t do this.”

He conceded that this mode of thinking challenges the current practice of development and puts huge constraints on future infrastructure requirements. But he indicated that mode will be an imperative as the military aims to build, among other things, robust autonomous vehicles.

Commercial and Environmental Limitations

“To think that our problems are going to be solved by commercial industry, it’s just simply a fallacy,” Sadler said. While the private sector has been key in crafting AI and machine learning technologies, they’re often far from optimized to suit the “tough, complex environments” the Army faces, he said.

And when the Army connects with the private sector on tackling a challenge? The response Sadler often finds: “That’s really a hard problem.”

Scientists at ARL acknowledge that noisy combat environments make AI training a potential nightmare, and up to this point, the Army isn’t quite moving at machine speed. “We’re really good at slow AI right now. We’re not really good at fast AI,” Sadler said.

In areas like robotics, Sadler said that adaptive machines are going to be a huge area of emphasis, as current models don’t yet have the ability to “think” on the fly.

“We’re not very good at manipulation, where I’m manipulating in the wild. We still solve problems with sort of brittle solutions,” he said. “So there’s a very big need, a very big research base for AI that’s going to enable physical reasoning.”

The Neural Network Guessing Game

Regarding neural networks–the deep learning processing machines whose methods of analysis remain esoteric even to those who create them–Sadler said, “We have no analytic framework. We have no theory which predicts performance.”

Unlike radar–the example Sadler provided–you can’t come up with a predictive model for how deep learning technologies will operate once they’re created. “I don’t have any tools like that for neural networks,” he said.

“Right now, neural networks that are data-driven are dominated by the empirical side. We’re in this guess-and-check,” Sadler continued. “I don’t have a predictive theory. I don’t know how it’s going to perform when I throw it at a problem.”

He pointed to the contentious Project Maven, explaining that any time the environment changes, re-training of the autonomous system is necessary. That’s a resource drain that may not be feasible in the long-term, Sadler said. “That’s just not sustainable. We can’t sustain that kind of effort.”

Neural networks are also fairly static once they’ve been programmed, he noted. “I don’t have any ability to update on the fly. I don’t have anything like an adaptive filter. I train it, it’s done. It’s a function. That is the function. It’s fixed,” he said.

The limitations of today’s Army AI landscape mean that verification and validation are paramount. With that point, Sadler underscored both the promise and unpredictability of the burgeoning field.

“We can expect some pretty dramatic failures along with dramatic successes,” he said.

Recent