Artificial intelligence is more than a new way of doing things: it’s a data-heavy computing process and that requires infrastructure.

To achieve its AI vision, the military will need to gather, move and process vast tracts of data. That means new network and compute infrastructures, much of it delivered at the tactical edge. The promise of predictive analytics will require an evolved hardware and software architecture.

Big vision

Responsibility for that infrastructure falls first to the recently formed Joint Artificial Intelligence Center (JAIC) and its emerging JAIC Common Foundation (JCF). That effort aims to develop data acquisition tools, unified data stores, libraries, standards and reusable tools in support of AI.

“As data grows, JCF will provide the supporting back-end infrastructure, tools and processes necessary for conducting analysis at scale,” said Elissa Smith, a Department of Defense spokesperson.

A cloud-based environment will deliver those resources to the war fighter. “It will provide the necessary platform capabilities for processing batch and stream data in an environment where the data can easily be discovered and shared across AI projects,” Smith said.

That’s a rough sketch. Industry experts say that “cloud” is only an approximate goal. The military will need to take a sophisticated and nuanced approach to the cloud in order to make AI work for war fighters on the edge.

While the big cloud providers, such as Amazon Web Services, can crunch massive loads of data to inform AI algorithms, it takes time to send that data from the myriad intelligence, surveillance and reconnaissance sensors in the field back to the cloud for processing — perhaps too much time.

“If you are in the Navy running AI on a ship, that ship is producing data on an hourly basis,” said Cameron Chehreh, Dell Technologies federal chief technology officer and vice president.

“Rather than send it to the cloud and get an answer back using limited bandwidth, you want to do the compute on the ship in near real time. You produce the data locally and you do the analytics locally.”

This requires robust local networking, large-scale compute power and ample storage, all at the tactical edge. In essence, this setup is a small-scale mirror of the large commercially available cloud offerings.

“If you have semi-autonomous drones in the battlefield, you need to be able to run the algorithm at the edge so that it can take defensive maneuvers immediately if necessary,” Chehreh said.

“The platform has to make some real-time decisions and you cannot afford the latency required to send it out to the cloud and back. It has to be able to do that computing at the edge where it sits.”

Some describe this as a “smart edge” infrastructure: a compute architecture that empowers sensors, mobile devices and drones to inform the AI algorithms in real time.

“This reduces the time from data to decision, and it decentralizes that decision authority,” said Ki Lee, a vice president in Booz Allen Hamilton’s strategic innovation group. Ideally, this system would dovetail into a large network infrastructure, making it possible to send AI findings back to the commercial cloud for further processing.

“In the instances where you don’t have the bandwidth, you process on the end, and then you will sync up with the cloud after operations, because that data is still imperative for refining the AI algorithms,” Lee said.

To make the most of this edge AI infrastructure, military planners will need to take a deep dive into data management. Data sits at the core of the AI effort, and industry experts say the Pentagon will need to be thoughtful about how it is handled, ensuring they have the architecture in place to make relevant data easy to share.

“If you have algorithms running at multiple locations, they will be processing different data. You need a global fabric that allows us to sync that data and those algorithms, a distributed architecture that ties it all together,” Lee said.

This issue of sharable data is a key consideration in discussions of AI infrastructure.

“The department first needs to figure out how to share data and break down data silos. Traditionally, the DoD and the intelligence community have held their data very close, but AI implementations require a steady stream of data to train models and improve inference engines,” said Steve Orrin, federal chief technology officer at Intel Corp.

This is of critical importance to the C4ISR community, which likely will be a prime generator of AI data.

“Sensor fusion, in particular, will require updated infrastructure to realize its benefits. The DoD will need to invest in architectural frameworks, such as peer-to-peer and mesh networks with gateways and aggregators,” Orrin said.

This modernized infrastructure does not come in a box, but rather is the product of new methodologies.

“When you look at AI there are some basic things you need to get right: data tagging, data [extract, transform and load]. When you think about combining data sets, tagging them so they can be understood, cleansing the data to get the noise out of it those are all steps that need to be done to make the data available to AI,” said Kyle Michl, managing director at Accenture Federal Services. “It’s not as exciting as algorithms, but it is critical.”

Securing the infrastructure

The next element in this infrastructure is security.

In this AI-enabled future, data must be secure and reliable. “If we actually do expand AI to the edge, we are dramatically increasing our attack surface,” Lee said.

“It will be critical to be able to secure things all the way to that edge device.”

Defense will need infrastructure not just to protect AI from garden-variety hacking, but also from enemy AI.

Consider the rise of Adversarial Machine Learning, or AML, “which, until recently, has mostly been explored in academic “circles,” said Scott Scheferman, senior director of global services at BlackBerry Cylance.

“That field has now matured sufficiently to present near-term material risks to AI systems that are not designed to withstand AML or be able to detect AML attacks.”

As the Department of Defense moves to address these security concerns, along with myriad other infrastructure needs, some say military leaders should look to industry to play a leading role.

Much of the development currently unfolding in the commercial realm could also be leveraged to support the military’s emerging AI infrastructure.

“If the Defense Department can take 80 percent of a solution from industry and spend only 20 percent of the time customizing it for the DoD’s unique use cases, they can speed development and avoid starting from scratch,” Orrin said.

Lee points to industry efforts in support of higher processing speeds and smaller, lighter computing devices as two areas ripe for industry engagement. “These algorithms will require more processing, and that will require computers with new efficiencies. At the same time, we have to be very mindful of size, weight and power: We need higher compute with lower wattages. Those are places where industry can help,” he said.

Arguably, the military may also need to look at its ethical infrastructure around AI. It’s less tangible than networking and storage, but no less significant, officials have said.

“If AI were leveraged for vehicle collision prevention technology, what does AI do in the event of an imminent crash in which avoiding one way will kill the driver of car A, and avoiding another way will kill the driver of car B?” said Edris Amiyar, a senior systems engineer at NetCentrics.

More In Artificial Intelligence