An official website of the United States government
A .mil website belongs to an official U.S. Department of Defense organization in the United States.
A lock (lock ) or https:// means you’ve safely connected to the .mil website. Share sensitive information only on official, secure websites.

News | July 20, 2020

Artificial Intelligence Deployment Requires Diverse Image Data

By C. Todd Lopez

Ensuring that technology powered by artificial intelligence will work anywhere requires that AI is "trained" on a diverse data set that readies it for deployment anywhere in the world. 

That's something the Joint Artificial Intelligence Center is well aware of as it pushes forward with the Defense Department's AI strategy, the center's director said.

Diversity in training models is a challenge, Nand Mulchandani said during the VentureBeat Transform 2020 virtual conference on artificial intelligence July 17.

The faces of five individuals are shown in a video conference.
Group Photo
Shuchi Rana, with the technology company Headspin, chairs a discussion on artificial intelligence during the VentureBeat Transform 2020 virtual conference, July 17, 2020. Included in the discussion were Nand Mulchandani, director of the Joint Artificial Intelligence Center; Anthony Robbins of the Nvdia technology company; Stacey Shulman of Intel; and Josh Sullivan of Booz Allen Hamilton.
Photo By: Screen capture
VIRIN: 200717-D-ZZ999-001Z

"When you think of sort of an oxymoron like mass customization, think of a single system that's deployed worldwide, globally," he said. "We've trained the model on a particular training data set. But that data set is not representative of say, global terrain, or global information, or even things like faces. So when you think of the diversity of ... humankind out there, ... if you're doing something like facial recognition or something, the training data set from a testing and representative perspective is so important."

From the testing and evaluation side, Mulchandani said, it's important for the JAIC to be able to ensure an AI system is trained in a diverse enough way that it can be deployed globally and work anywhere.

A rifle, sighting devices and drones.
Weapons Package
An M4 Carbine rifle equipped with the Smart Shooter sighting device lies next to drones to be used as moving targets for the 9th Reconnaissance Wing at Beale Air Force Base, Calif., Aug. 14, 2019. The sighting device attaches to the weapon and then locks on and fires to neutralize its target with or without movement. The device is also being used to limit friendly fire, as the weapon cannot be fired unless it is purposely locked on.
Photo By: Air Force Tech. Sgt. Alexandre Montes
VIRIN: 190814-F-BN304-0013

"What we're finding is ... we're still in the early days of AI where the ability for a single dataset to perform in multiple different environments and applications is incredibly important," he said.

Mulchandani also addressed concerns that DOD might be difficult to work with for smaller companies involved in artificial intelligence research and development.

Information graphic
Artificial Intelligence
An information graphic summarizes the goals of the Defense Department’s artificial intelligence strategy.
Photo By: Erik Sams, DOD
VIRIN: 190212-D-ZZ999-0211C

"From the outside, there seems to be this idea that the DOD, the Pentagon, has a very hard time liaising and working with tech startups and even large tech companies," he said. However, he added, there's a lot of change going on at DOD as the department partners and works directly with large and small tech companies.

The JAIC now has projects going on with startup companies that have as few as seven employees, he noted.

"The ability for us to have those direct conversations, direct work with them — the environment has never been better," he said. "And there's huge changes going on in terms of how acquisition gets done, how we actually acquire, procure and deliver software inside the DOD from a cloud perspective, and other things."