Stanford AI Projects Greenlighted in National AI Research Resource Pilot

On May 6, the U.S. National Science Foundation and the Department of Energy awarded grants to 35 research teams for access to advanced computing resources through the National Artificial Intelligence Research Resource (NAIRR) pilot. This initial wave of awarded projects includes scholars from across the U.S. who are working in clinical medicine, agriculture, biochemistry, computer science, informatics, and other interdisciplinary fields. Two Stanford AI projects — from the School of Engineering and School of Medicine — were selected to participate in the pilot.

Part of the 2023 Executive Order on the Safe, Secure, and Trustworthy Development and Use of AI, the NAIRR pilot launched in January 2024 with four stated goals: spur innovation, increase diversity of talent, improve capacity, and advance trustworthy AI. Stakeholders in academia, industry, and government see this program as a critical step toward strengthening U.S. leadership in AI and democratizing AI resources for public sector innovation.

“The NAIRR pilot is a landmark initiative that supports applied AI research and will benefit the entire nation,” said Stanford Institute for Human-Centered AI Deputy Director Russell Wald. “No AI scholar should be constrained by the high cost of compute resources and access to data to train their models.”

Most of the awarded projects are given computational time on NSF-funded supercomputer systems at the University of Illinois Urbana-Champaign, University of Texas at Austin, and Pittsburgh Supercomputing Center; additionally, the DOE will allocate resources at its Summit supercomputer at Oak Ridge National Laboratory and AI Testbed at Argonne National Laboratory to a few of the research teams.

Taking Reinforcement Learning into Visual Environments

A team from the Stanford Intelligent and Interactive Autonomous Systems Group (ILIAD), led by HAI Faculty Affiliate Dorsa Sadigh, an assistant professor of computer science and of electrical engineering, submitted a proposal to continue groundbreaking work in the domain of human-robot and human-AI interactions. The project will focus on learning effective reward functions for robotics using large datasets and human feedback.

Reward functions are key to a machine learning technique called reinforcement learning, which works by training a large language model to maximize rewards. When humans provide feedback as part of the training process, the model learns how to make decisions that are aligned with human priorities. Stanford computer science PhD student Joey Hejna says that applying this technique to real-world robotics presents new challenges because it requires understanding the visual world, which is captured by modern visual-language models. Another challenge is that it’s not enough for the model to get the right result; how it arrives at that answer also matters. Researchers will want to make sure the robot operates safely and reliably around people, and they may need to personalize how certain robots interact with humans – in a home-care setting, for example. 

“Training robot models that can work in the real world will require a massive amount of compute power," Hejna explains. “High-performing VLMs usually have at least 7 billion parameters. This project would not be possible without access to the GPU hours from the National Science Foundation.”

Autonomous Patient Monitoring in the ICU

The second Stanford project to receive NSF support comes out of the School of Medicine’s Clinical Excellence Research Center (CERC), dedicated to reducing the cost of patient care. Part of a multiyear initiative to enhance healthcare environments by integrating smart sensors and AI algorithms, the awarded project seeks to develop computer vision models that can collect and analyze comprehensive video data from ICU patient rooms to help doctors and nurses better track patients’ health.

A key aspect of the research is to address potential biases in the AI models used for predicting patient status and monitoring clinical activities. By analyzing demographic data from electronic health records, the team aims to identify and correct algorithmic biases that might affect predictions across different ethnicities and sexes. “The ultimate goal is to develop bias-free algorithms and propose interventions to ensure fair and accurate patient monitoring and care in ICUs,” said the team’s lead scholar, Dr. Kevin Schulman.

Leadership from Stanford HAI

HAI’s leadership team has been a driving force behind the creation of a National AI Research Resource since the founding of the institute in 2019. Co-Directors Fei-Fei Li and John Etchemendy started to organize universities and tech companies in 2020, and they initiated the call for a government-led task force to establish the program.

“From our earliest conversations with universities, industry executives, and policymakers, we felt that American innovation was at stake,” said Li. “We knew that support from Congress and the president could have a meaningful impact on the future of AI technology.”

According to Etchemendy, “The start of this pilot program marks a historic moment for U.S. researchers and educators. It will rebalance the AI ecosystem by supporting mission-driven researchers who want AI to serve the public good.”

Reflecting on the years of strategic planning and dedication that have led to this milestone, Wald added, “John and Fei-Fei’s vision, combined with the extraordinary support of the Stanford community and our country’s policymakers, is leading to greater access to AI research not just at Stanford but at all of America’s universities.”

Immediately following the May 6 announcement of initial awards, the NAIRR pilot opened the application window for a second wave of projects. With contributions from industry partners, a wider range of technical resources are going to be available for applicants this round, including access to advanced computing systems, cloud computing platforms, foundation models, software and privacy-enhancing tools, collaborations to train models, and education platforms. 

Researchers and educators can apply for access to these resources and view descriptions of the first cohort projects on the NAIRR pilot website.

Stanford HAI’s mission is to advance AI research, education, policy and practice to improve the human condition. Learn more.