Machine Learning Models Expedite Federal Tech Efforts

Instead of creating models from scratch, agencies have swiftly built and deployed ML functionality by adapting pretrained models from cloud service providers.

That approach can benefit agencies that lack robust IT talent and other resources, says Nicol Turner Lee, governance studies senior fellow and director of the Center for Technology Innovation at the Brookings Institution.

“There’s been more legibility when it comes to very dense data sets in recent years,” Turner Lee says. “A lot of machine learning automation of various federal data sets has been supported by companies, enabling functions that were traditionally redundant or inefficient in certain agencies.”

Taking Advantage of Pretrained Models

In 2020, the VBA implemented a custom IBM platform to assist with benefits delivery, Tellez says.

“Veterans deserve the fastest, most accurate delivery of benefits VA can provide,” Tellez says. “That’s why it’s important to embrace automation technology.”

Prepackaged ML libraries that can facilitate model development and training are often provided as part of a cloud platform, according to the General Services Administration’s AI Guide for Government.

Amazon Web Services’ SageMaker AI, for instance, provides pretrained ML models agencies can use to create their own versions. The SageMaker JumpStart ML hub has hundreds of foundation models, which can be employed with broad data sets.

Google offers a Vertex AI ML-based development platform that provides tools for testing, fine-tuning and deploying ML models, including foundational and task-specific models that are ready to use.

“We expect AI to be part of the way we build our products and services,” Turner Lee says. “For companies that do business with the federal government, there will be more appetite for AI to be embedded or have some capability in existing software applications.”

Refining Prompts and Giving Models Detailed Instructions

The National Archives and Records Administration is currently working to produce digital versions of the more than 13.5 billion pieces of paper the agency has preserved, with a goal of digitizing 500 million pages by the end of fiscal 2026.

As items continue to be added from a high-speed scanning center in College Park, Md., and from presidential libraries and other locations throughout the U.S., NARA is conducting a semantic-search pilot to ensure agencies and the public will be able to easily locate relevant resources in the expanded catalog.

RELATED: Agencies must digitize records according to NARA guidelines.

The pilot involves testing open-source and commercially available solutions, such as AWS’s Titan ML model and Google’s Vertex AI ML platform, to perform queries and other functions, says Jill Reilly acting chief innovation officer at NARA.

“We’ve been using mostly pretrained models,” Reilly says. “We’ve also been looking at large language models to explore AI-generated summaries and descriptions. We have several cloud sandboxes set up for different vendors.”

One of the archival or digital access experts on the integrated project team might, for example, enter a prompt into the project interface, asking for a formal title in a document to be identified or a summary of the topics and people involved. The agency would then evaluate the relevance and accuracy of the model’s response.

Comparing the different ML models’ results has been interesting, Reilly says. And the process has highlighted the importance of refining prompts and including detailed instructions for the model.

WATCH: NSF is playing matchmaker with researchers and AI.

In addition to helping users sort through the agency’s sizable and ever-growing volume of data, Reilly anticipates the technology will present staff-related cost reduction and other opportunities.

“It’s been a time-saver to use the pretrained models for our testing and evaluation,” Reilly says. “We wouldn’t be able to provide meaningful access for the American people and our customers without these tools.”

Securing AI/ML Models for Compliance

With built-in cybersecurity and other features, machine learning models from comprehensive cloud providers could potentially offer agencies additional assistance.

“These models should come with some cybersecurity backstops to ensure the technology won’t fail,” Turner Lee says. “Agencies want to think about the cybersecurity benchmarks and what’s embedded, in terms of addressing some of the vulnerabilities that could occur with either a system failure or external breach.”

DISCOVER: Agency transformation calls for AI-driven cyberdefenses.

Some commercially available platforms include amenities that could help agencies meet government standards and compliance needs.

ML model-related information can be shared between AWS’s SageMaker and IBM’s watsonx.governance, for example, to establish customized risk assessment and model approval workflows that make adherence to regulatory and other policies possible.

There are many ways cloud solutions can support ML model development and use; for example, by supplying infrastructure elements such as adequate compute power. Compute resources are necessary for effective use of ML models, but some agencies may be lacking, Turner Lee says.

“There are many fundamental tenets of technology that need to be in place,” she says. “If companies can offer not only the software but also the storage capabilities, it is definitely a win for federal government to embed multiple services from one provider.”