Engineering AI for the public good

The rapid adoption of artificial intelligence (AI) is reshaping nearly every industry as it accelerates automation and increases efficiency.

At the same time, the advent of AI has strained natural resources and revealed the need for a workforce that knows how to harness its capabilities and maintain its infrastructure.

Maryland Engineering is developing creative, pragmatic solutions to these challenges while equipping professionals with the skills they need to lead in the AI economy.

Improving self-driving vehicles via Mario Kart simulations

A University of Maryland (UMD) team is working to provide regulators with a roadmap to certifying AI technologies in the autonomous fleets that are increasingly popping up in cities across the U.S. and around the world.

The National Highway Traffic Safety Administration sets safety standards for aspects of traditional cars, and drivers are tested before they are issued a license. In contrast, no broadly applicable regulatory framework exists for evaluating the safety of autonomous vehicles—even though they already operate on public streets.

“It’s kind of the Wild West out there right now,” says Mumu Xu, associate professor of aerospace engineering. “We don’t really have a process to figure out, ‘Is this autonomous car going to be safe when driving?’”

Xu is working to fill that gap with a process that relies on the classic Nintendo game Mario Kart. Her team downloaded an online version of the game and trained a computer to play it, using reinforcement learning: They augmented the code with a reward structure—a set of “if-then” statements—to grant Mario points for finishing a lap and not hitting the wall. Over time, after trying different actions and learning how to maximize its reward, the computer’s driving behavior improved. Xu then extracted the map of Mario’s driving path and speed and verified whether it met the standards she had set, effectively evaluating whether the computer had learned to drive safely.

The video game is the perfect tool for developing a safety-testing process, because it’s a simple analog for the simulators that will be used to test actual vehicles. Xu’s research is funded by the Naval Air Warfare Center Aircraft Division, which may apply the basic process she’s developed with its own higher-fidelity simulators. She anticipates that regulatory agencies for autonomous vehicles will develop a similar process for testing vehicle safety. After the vehicles’ software has been tested for safe driving in simulation, regulators would analyze the outputs—just as Xu is evaluating Mario’s driving record—to determine whether the vehicle is operating safely. If the software performed well in simulation, it would then be tested in a real car.

Xu’s work also could be applied to biomedical devices with AI components, such as surgery tools with computer vision sensors: The tools could be tested in simulation, and evaluators could review the outputs and adjust them to improve safety.

Data centers: Metric-based building standards for long-term success

As data centers proliferate, UMD engineers are ensuring these enormous structures are built to last, with fewer environmental impacts.

With their immense size, heat generation, water usage, and noise production, data centers are unlike any other infrastructure. But the data-center construction industry is so young that no building standards or maintenance protocols exist.

The current construction boom is akin to a gold rush, says Nii Attoh-Okine, chair of the Department of Civil and Environmental Engineering: “Everybody is focused on the output from these data centers, and that’s important—but we’re overlooking the civil and the maintenance aspect of these huge buildings. If the building itself will have problems in a few years, that threatens the investment in the computer technology.”

Attoh-Okine is developing a data center “report card” based on metrics (such as water use per square foot) that companies can use for self-evaluation and improvement. He’s also partnering with industry to develop maintenance protocols, adaptable to regional variations in climate, water availability, soil structure, and seismic activity.

Such guidelines can be used to train the data center technicians who maintain this new type of infrastructure: the foundation, walls, and roof, as well as the computing equipment.

“We need a workforce that can handle very complex infrastructure maintenance,” Attoh-Okine says. “Right now they are largely learning it on the job, as they go.”

Last fall, Attoh-Okine and Professor of Civil and Environmental Engineering and Chair of Civil Empowerment Birthe Kjellerup led the inaugural gathering of experts from academia, industry, government, and local trade unions to discuss sustainable design, strategic location, and future job creation related to data centers. The conversation led to the launch of an information hub that will be a resource for the mid-Atlantic region.

“Data centers aren’t going away—on the contrary,” Kjellerup says. “We can design them well from the start if we plan now. At the University of Maryland, we have the resources, we have the knowledge, and we can be the ones that have the innovative solutions.”

Reducing server load with efficient tuning of large language models

UMD engineers develop a streamlined method for tuning large language models (LLMs) that cuts energy use and boosts performance.

The development and deployment of LLMs require vast amounts of energy; communities where data centers are located have seen strains on their electric supply and spikes in energy prices.

Assistant Professor of Electrical and Computer Engineering Sanghamitra Dutta is devising more efficient ways to design language models in order to save energy and precious natural resources. Her work is applicable to models customized for specific tasks, such as a bank’s chatbot that provides customer service. In contrast to general-purpose LLMs like ChatGPT, customized models are generally deployed in environments with more constrained storage and memory capabilities.

One project involves knowledge distillation, the process of training a smaller language model (called a student) from a larger model (called a teacher). Dutta’s work makes the knowledge distillation process more efficient by using a strategic set of data points to tune the model.

Take, for instance, a feature on a bank’s website that can tell a customer whether they’d qualify for a loan. Instead of training the model with data-point pairs of loan applications and their result—acceptance or decline—her method gets more specific by using contrasting data points called “counterfactuals,” designed to help the model understand and generate “what-if” scenarios that assess how outcomes changed with small differences in past conditions.

By using counterfactuals, Dutta’s method cuts the total number of data points needed for training in half, and also yields performance improvements. Less training time means less energy used: “We can train these student models much faster, and the student models are more faithful to the teacher models,” Dutta says. “It’s a win-win situation.”

Maryland engineers the solution >> Keep reading to get the full story.