Best Practices for Building an AI Development Platform in Government

John P. Desmond, AI Trends Editor

According to Isaac Faber, chief data scientist at the U.S. Army AI Integration Center, the AI ​​stack defined by Carnegie Mellon University is fundamental to the approach the U.S. Army takes in developing its AI platform. World Government AI The event was held in person and virtually in Alexandria, Virginia last week.

“If we want to move the Army away from legacy systems to digital modernization, one of the biggest challenges I’ve seen is the difficulty of abstracting away the differences in applications,” he said. “The most important part of digital transformation is the middle tier, the platform that makes it easy to work in the cloud or on-premises.” The desire is to be able to migrate your software platform to another platform with the same ease that a new smartphone migrates a user’s contacts and history.

Ethics permeates every level of the AI ​​application stack: at the top is the planning stage, followed by decision support, modeling, machine learning, big data management, and at the bottom the device or platform level.

“I advocate for us to look at the stack as the underlying infrastructure and way to deploy applications, rather than as a siled piece of our approach,” he said. “We need to create a development environment for a globally distributed workforce.”

The Army is working on the Common Operating Environment (Coes) software platform, which was first announced in 2017. It is scalable, flexible, modular, portable and open design for DoD work. “It is suitable for a wide range of artificial intelligence projects,” Faber said. For completing these efforts, he said, “The devil is in the details.”

The Army is working with CMU and private companies on a prototype platform, including Vizimo of Coraopolis, Pennsylvania, which offers artificial intelligence development services. Faber said he prefers to collaborate and coordinate with private businesses rather than buy products off the shelf. “The problem is that you are locked into the value that is provided by one vendor that is not typically designed for DoD networking needs,” he said.

The Army is training a number of technical teams to use AI

The Army is involved in developing the AI ​​workforce for several teams, including: executives, specialists with advanced degrees; technical personnel undergoing training to obtain a certificate; and AI users.

Tech teams in the Army have a variety of focus areas, including general purpose software development, operational data processing, deployments that include analytics, and a machine learning operations group, such as the large team needed to build a computer vision system. “As people come to work, they need a place to collaborate, build and share,” Faber said.

Types of projects include diagnostic, which can integrate streams of historical data, predictive, and prescriptive, which recommend a course of action based on the forecast. “At the far end is artificial intelligence, don’t start there,” Faber said. The developer will have to solve three problems: data engineering, an AI development platform, which he called the “green bubble,” and a deployment platform, which he called the “red bubble.”

“They are mutually exclusive and all interconnected. These teams, made up of different people, must be coordinated programmatically. Typically a good project team will have people from each of these bubble areas,” he said. “If you haven’t already, don’t try to solve the green bubble problem. There’s no point in doing artificial intelligence unless you have an operational need.”

When asked by a participant which group was the hardest to reach and train, Faber responded without hesitation: “The hardest group to reach is the executives. They need to understand what value the AI ​​ecosystem can provide. The biggest challenge is how to communicate that value,” he said.

Panel discussion discussed the most promising use cases for AI

In a discussion on the foundations of new AI, moderator Kurt Savoie, director of the Global Smart Cities Strategies program at research firm IDC, asked which new use case for AI has the most potential.

Jean-Charles Ledet, autonomy technology advisor for the Air Force Office of Scientific Research, said, “I would point to the benefits of decision-making at the front end, supporting pilots and operators, as well as decisions at the back end, for mission planning and resources.”

Krista Kinnard, director of the Department of Labor’s Emerging Technologies Office, said, “Natural language processing is an opportunity to open the door to artificial intelligence at the Department of Labor,” she said. “Ultimately we are dealing with data about people, programs and organizations.”

Savoie asked what big risks and dangers the panelists saw in the implementation of AI.

Anil Chaudhry, director of federal artificial intelligence implementation at the General Services Administration (GSA), said that in a typical IT organization using traditional software development, the impact of a developer’s decision goes very far. With artificial intelligence, “you have to consider the impact on an entire class of people, participants and stakeholders. By simply changing algorithms, you could delay benefits for millions of people or make the wrong conclusions at scale. That’s the biggest risk,” he said.

He said he asks his contract partners to keep “people informed and people informed.”

Kinnard echoed this, saying: “We’re not looking to exclude people from the process. It’s really about empowering people to make better decisions.”

She emphasized the importance of monitoring AI models after they are implemented. “Models can change because data drives change,” she said. “So you need a level of critical thinking to not only complete the task, but also to evaluate whether what the AI ​​model is doing is acceptable.”

She added: “We have developed use cases and partnered with the government to ensure we implement responsible AI. We will never replace people with algorithms.”

The Air Force’s Lehde said: “We often have use cases where the data doesn’t exist. We can’t look at 50 years of war data, so we use simulation. The risk in training an algorithm is that you have a ‘simulation-to-real gap,’ which is a real risk. You’re not sure how the algorithms will map out in the real world.”

Chaudhry emphasized the importance of a testing strategy for artificial intelligence systems. He warned developers “who get carried away with the tool and forget the purpose of the exercise.” He recommended that the development manager develop an independent verification and validation strategy. “Your testing is where you need to focus your energy as a leader. Before committing resources, a leader needs to keep in mind how they will justify whether the investment was successful.”

The Air Force’s Lede talked about the importance of explainability. “I’m a technologist. I don’t make laws. The ability of an AI function to explain in a way that a human can interact is very important. AI is a partner with whom we have a dialogue, not an AI that comes to a conclusion that we have no way to verify,” he said.

Find out more at World Government AI.