David Brown is a research engineer in the Defense & Intelligence Solutions Division at Southwest Research Institute where he is the lead engineer for cognitive electronic warfare system research and development.
What is the Southwest Research Institute?
Our tagline is deep sea to deep space. We have multiple applied research programs going on across the spectrum—everything from chemical to petroleum to mechanical engineering to planetary and deep space science. We’re a nonprofit independent applied research organization not affiliated with any university. SwRI was started by a philanthropist in the 1940s and was initially focused on research related to the automotive and petroleum industries. We’ve since evolved to focused research across most scientific domains. My group within Southwest Research is specific to defense and intelligence systems. Most of our work is sensitive or classified.
What brought you to the Southwest Research Institute?
I dreamed of flying F-15 fighters when I was a teen. When I enrolled at Georgia Tech as an engineering undergrad, I immediately joined the National Guard because they had an F-15 unit. When I graduated I needed an engineering job while waiting to be picked up to go fly F-15s. Southwest Research Institute hired me directly out of school and a year or two later, I did get picked up to go fly in the National Guard, but the F-15 unit had changed aircraft to the B-1B, so I became a B-1B “wizzo” electronic warfare officer while working with the Southwest Research Institute on the same electronic warfare systems. It was a neat combination—performing research and development for the systems that I was flying into combat.
Southwest Research has always pushed the envelope on electronic warfare (EW) and using the latest technology to improve our processes. In the past few years, technology improved so that our systems started generating more data than could be processed using traditional signal processing methods so our teams started looking at different ways of handling the data using newer technologies. We looked at compressive sensing and several other methodologies to handle that volume of data. We realized we were getting very good at throwing away data in an attempt to hone in on the pieces that were most relevant to us. But, the data that we were throwing away had useful information, we just didn’t have the capacity to process it. A few years ago, we started delving deep into how we could apply Artificial Intelligence (AI) to some EW problems that come from having to handle a large amount of data. That has since become a large internal research program that is now partially funded by the Department of Defense.
What is cognitive electronic warfare?
Let’s start with electronic warfare. It’s a broad spectrum that covers most radio frequency (RF) emitters out there today—we’re talking about things like radar and communication systems as well as everyday devices such as Bluetooth and wireless internet. However, most of my focus is on radar signals. You can think of radar as a technology that is trying to get a picture of the space around itself, and it’s usually used to see aircraft or other vehicles moving within that space. Electronic warfare focuses on ensuring that those radar systems cannot perform their job as intended. We’re on the defensive side of this so if you think about a US Air Force or Navy aircraft flying in a hostile area, for example, our intent is to keep US aircraft safe by making sure the adversary can’t develop a track on that aircraft for a targeting and firing solution.
As for cognitive EW, there are two definitions we could use. One is very loose, and it refers to any AI process that is applied to the EW problem set. But I like a stricter definition where you’re looking at a system that can autonomously sense its environment, make decisions based on what it sees within the environment, and then affect the environment based on the decisions that it makes. Maybe it senses the emitters within the environment, makes a decision about what jamming technique to use on those emitters, and then autonomously generates that response and an estimate of its effectiveness so that it can control what it’s doing.
You’ve been researching and using electronic warfare systems for decades. What’s changed over the past 10-20 years that has led to the need for cognitive systems?
If you go back to the 1980s, the technology that we had was very constrained. If we wanted to update a radar from one frequency to another or change the operating parameters such as the timing that it uses it was a lengthy process that might have taken our adversaries years. They had to develop new hardware, that hardware had to be tested, and you had very little digital control and processing capability within the system. And during that process, as our adversaries were developing a new system, our intelligence community would collect information on what they were doing, how they were radiating, what technologies were being employed, and so we’d have some indication years in advance that they were developing a new threat system that we had to react to. But our systems were constrained by the same technology problem and lengthy development process. Well, over time the acquisition process for these systems started speeding up. In addition, new hardware technologies have significant performance margins relative to the control process, meaning there is significant flexibility in digital software control. Thus, hardware isn’t changing nearly as fast as our ability to control it with software. So, if you can imagine that an adversary is using a radar in a particular mode—using particular timing and modulation forms—and they detect we have a way of responding to that operating mode to introduce error into their system, they can very quickly—within weeks, days, or hours—change what they’re doing. We must keep up with that rate of change and in many cases, the only viable way to respond to highly adaptive threat systems is to incorporate cognitive processes into our EW system. In addition, the idea with cognitive radar is that the radar system itself can sense changes in the RF environment and cognitively change how the radar is operating. Our adversaries are continuously updating how they send out their pulses to adapt to the dynamic environment, and we have to be faster at updating on our end to predict what they’re going to do next and determine the right response.
Earlier this year SwRI was awarded a nearly $6.5M contract from the USAF to do R&D work on cognitive electronic warfare systems. What are you researching as part of this contract?
We have a strong research collaboration agreement with a major chip manufacturer who is developing an advanced neuromorphic processor. One of our distinctions is that we’re one of the few research organizations that has research agreements on using neuromorphic processing for cognitive EW. There are a few places in the US working on the general problem of applying AI techniques to this domain, but what makes our work stand out, I think, is that we don’t approach neuromorphic computing for cognitive electronic warfare as an AI problem first. Instead, we approached it as an EW problem and then looked for the best tool available to solve our challenges. One of the newer technologies that is available is neuromorphic processing with spiking neural networks. As we apply that technology, we’re seeing incredible improvements and advantages within the electronic warfare context.
What makes neuromorphic architectures such a good solution for cognitive electronic warfare?
Neuromorphic architectures are much more generalizable than more traditional AI approaches. Radar signal data is very messy, which creates challenges around training the AI to make inferences on radar data. We started at the I/Q level, which is a direct measurement of the RF spectrum itself. That is the most information-rich form of data, but it’s also the messiest and most complex. With traditional signal processing techniques, we smooth I/Q data using transforms so that we can pull out features that are understandable to humans, but in that process, we’re actually obscuring data that’s in the I/Q stream. When we apply AI to that original I/Q data using traditional methods, we can make it work on a specific set of I/Q data, but it doesn’t generalize over various operating and environmental conditions. It’s more specific to one scenario. We find that it generalizes much better with spiking neural networks. We’re able to use it over a wider operating envelope and apply AI to more cases than we can with a traditional method.
The second thing is that it uses much less power than traditional AI approaches. A lot of our EW systems are implemented on aircraft or small autonomous vehicles that have very limited power. In some of our tests, we are seeing performance out of our neuromorphic processor that’s equivalent to a bank of GPUs. Each of those GPUs is pulling quite a bit of power, but they also have to be cooled and that’s actually often the bigger challenge. But when we go to a neuromorphic processor, we’re using at least 3 orders of magnitude less power so we’re able to put these processors in a much more constrained environment.
What are the biggest challenges with using neuromorphic computing for cognitive electronic warfare?
One of the biggest challenges is obtaining suitable data to train the AI. We are constantly running into that issue. One way of mitigating this challenge is to augment real data with synthetic data. The problem with the real data taken from a radar is that the data was only collected in a limited number of environmental and operational conditions. There are only so many radar systems I can collect data from, and it turns out that most of our adversaries are not very cooperative when we do our testing. But with synthetic data, I can generate an almost unlimited number of scenario combinations and represent a wide range of environments and operating characteristics. The challenge is that synthetic data tends to be mathematically precise, which essentially allows the AI to pick up on features that aren’t available in a real data set. So, the AI is looking at precision that’s not there when we are flying in a real environment. The hardest problem that we’re working on solving right now with the DoD is ensuring that our algorithms are not biased by that synthetic data so we can graduate the algorithm to a real data set in an operational environment.
What does the future of electronic warfare look like to you?
What I’m seeing is the use of AI within specific domains. AI has largely been a domain unto itself where you had AI experts that did not necessarily understand the application domain such as EW. Now we’re growing in AI knowledge and the AI experts are growing in the EW domain knowledge. We’re seeing a real acceleration of what we’re realizing we can do with this technology. For decades we’ve had these paradigms where we’re used to solving problems in a traditional linear way using non-AI methods, but I would expect that over the next five to ten years, you’re going to see an acceleration as AI becomes more ubiquitous in the domain. Ten years from now people who are starting work in the electronic warfare domain are just going to understand these systems intrinsically. Forty years ago we would say we’re going to use a “digital controller” on an electronic warfare system. Now nobody says they’re going to use a digital controller; it’s just assumed. Similarly, in ten years, it’s going to be assumed that you have cognitive in your electronic warfare systems.