Skip to main content Skip to secondary navigation
Page Content

Meet 12 Asteroid Shots in AI

Stanford scholars explore advances in foundation models, explore the next-generation chip, and study causal models at the recent Hoffman-Yee Symposium.

Image
Night sky with meteors with dark trees in the foreground

What’s the next big thing in artificial intelligence? Over the past three years, the Stanford Institute for Human-Centered AI’s Hoffman-Yee Grants have funded asteroid shots — big, exciting ideas on the edge of possibility — that leverage AI to address significant scientific or societal challenges aligned with Stanford HAI’s mission.

Research teams of faculty, postdocs, and graduate students spanning Stanford schools of business, engineering, humanities and sciences, law, and medicine compete for up to $2.5 million in funding each over the course of two years. 

“These projects were selected for their boldness, ingenuity, and potential for transformative impact,” said HAI Vice Director of Research James Landay. “We believe the results of these projects could play a significant role in defining future work in AI from academia to industry, government, health care and civil society.”

Meet these 12 teams and learn how they’re pushed the boundaries of AI’s abilities. 

An AI “Time Machine” To Explore the Social Lives of Concepts

This team uses AI to explore the social lives of concepts to help humanists and social scientists trace how concepts change over time in an effort to better understand culture. In one study, this team tracked political speech regarding immigration over decades, finding more positive attitudes today than at any point in history (but with more partisan divide). Another study analyzed research innovation at Stanford over the past 50 years, revealing funding patterns, most successful collaboration teams, and even the impact of marketing descriptions on future revenues. The team, which includes scholars from philosophy, economics, political science, sociology, computer science, history, medicine, English, and the libraries, is now examining whether large language models (LLMs) can predict the complexity of human behavior in experiments, which could help social scientists identify effective interventions in policy or public health much more cheaply and efficiently.   

Intelligent Wearable Robotic Devices for Augmenting Human Locomotion

Falls are a common health care problem that can lead to a hospital stay or worse. This team is building an AI-enabled exoskeleton that understands human motion and intention and can learn to recognize when someone is about to fall and then actually prevent it. To build a successful exoskeleton, though, the team must first collect robust data on human falls. To that effort, they developed a device that safely simulates and induces falling, while measuring variables such as the motion of the fall (through cameras) or the muscle activity required to recover (through sensors). Their dataset will be used to train the predictive human model for the exoskeleton. 

EAE Scores: A Framework for Explainable, Actionable, and Equitable Risk Scores for Health Care Decisions

About 1.5 million people in the U.S. live with type 1 diabetes, but care and access can vary widely. Today, more tools exist to monitor and improve this condition remotely: continuous glucose monitors, insulin pumps, activity monitors, and more. This team developed the Stanford 4T Study (teamwork, technology, targets tight control) to improve health outcomes using these tools. In tests, they were able to start glucose monitoring within the first two weeks of a patient’s diagnosis, get people on automated insulin delivery, and build an algorithm to help control insulin dosing. With engineering and diabetes experts, they built the TIDE platform (timely interventions for diabetes excellence), which alerts clinicians when families need help between visits. The team is developing AI to prioritize patients for review, highlight clinically meaningful insights, personalize interventions, estimate treatment effects, and even craft messages for patients to improve outcomes. “We really have an opportunity to improve care at the patient level, improve the process for the clinician, and improve the quality of life for families,” said Stanford pediatrics professor David Maahs.

The team is partnering with Children’s Mercy Hospital in Kansas City and nonprofit Tidepool, a mobile app for automated insulin dosing.

Foundation Models: Integrating Technical Advances, Social Responsibility, and Applications

Foundation models, or systems trained on massive amounts of multimodal data that can be adapted for multiple uses, have gone mainstream — think ChatGPT, Stable Diffusion, and more. This team studies the applications, technical advances, and social impact of these exciting tools. 

In applications, they have developed new resources in law, health care, and robotics. For example, the team created a 256 gigabyte open-source legal dataset built from court opinions, contracts, administrative rules, and legislative records. In health care, they developed RoentGen, a text-to-image model that generates synthetic chest X-rays with Stable Diffusion. In robotics, they built Voltron, trained on human videos, and diversified household robotics data from households across the world.

In technical advances, the team developed a new suite of techniques that improve upon state-of-the-art methods for model performance. DoReMi utilizes distributional robust optimization to automatically determine domain weights. Using DoReMi, the team was able to train a model 2.6 times faster compared with the standard heuristic data selection approach. Hyena is an alternative, faster, architecture to the transformer, enabling longer context; pretraining optimization tool Sophia is shown to be 2 times as fast as the state-of-the-art (Adam); and Direct Preference Optimization (DPO), an alternative to reinforcement learning from human feedback (RLHF), is as simple as doing supervised learning. 

And in the social responsibility space, the team is working toward improving transparency. They developed HELM, or Holistic Evaluation of Language Models, benchmarking 30 prominent language models across scenarios and risks, and Ecosystems Graphs, which evaluates the downstream uses of foundation models. “This has profound implications on the policy world, and this work is extremely urgent,” said lead investigator Percy Liang.

Tuning Our Algorithmic Amplifiers: Encoding Societal Values into Social Media Algorithms

Values at the core of social media algorithms are focused on the individual user experience —what will interest them and keep them engaged? But that comes at a cost; consider increased polarization across social media during the last U.S. presidential campaign.

But what if we could embed social values into social media? Can we imagine a different relationship with social media than what we have today?

This team aims to develop an algorithmic library of societal values tied to communities and cultures. Decades of research in the social sciences have led to an already established library of values, which this team will use to build the algorithmic model. They will also deploy a large-scale empirical study to see if these measures impact human behavior and attitudes. 

In an early paper, the team looked at political polarization in sociological research and translated it into a feed-ranking algorithm to up-rank pro-democratic attitudes. Among 2,000 U.S. voters, they were able to reduce the partisan animosity without decreasing engagement or prompting perceived threats to free speech. 

Curious, Self-Aware AI Agents To Build Cognitive Models and Understand Developmental Disorders

This team is examining how humans learn —through self-supervised, open-ended contexts — to build a collection of agents that can learn to play and play to learn, with the aim of training a simulated robotic environment to eventually implement into real-life robotics. 

To gather data, the team created BabyView — high-resolution cameras affixed to child-sized helmets. Ten families enrolled in the pilot, capturing one to two hours a week for hours of usable footage. Now the team is using that data to develop self-supervised and unsupervised learning algorithms deploying deep neural embeddings to develop agents that act more like humans. They’ve also developed the concept of counterfactual role modeling, or models trained to predict the future from partial previous information that can also make counterfactual predictions about what might happen under different circumstances.

Dendritic Computation for Knowledge Systems

Deep neural networks have radically expanded the limits of AI, but they have also created a major demand for computation resources. This team seeks to make more efficient computer chips by borrowing the idea of a dendrite, the protrusions that brain neurons use to detect signals, and to eventually make chips so efficient that a giant model could run on a cell phone. This would rein in AI’s unsustainable energy costs, distribute its productivity gains equitably, transform its users’ experience, and restore their privacy.

To do this, they’re creating specialized accelerators for memory cells called Mulitgate FeFET that change how voltage flows within the cell without increasing heat and melting the cells. These “nanodendrites” are a new kind of computing device that operates on spatial temporal patterns instead of traditional vectors. The team is also studying how these nanodendrites can accelerate information retrieval. So far, the team’s Dendro-SPLADE has beaten current benchmarks on speed, recalling information with only microseconds of latency. “It’s really pushing the frontier of what’s possible by fusing algorithms and hardware,” said Saarthak Sarup, a Stanford PhD student in electrical engineering.

Matching Newcomers to Places: Leveraging Human-Centered AI To Improve Immigrant Integration

In this moment in time, there are 281 million immigrants in the world, 108.4 million displaced by geopolitical events or climate disasters, and 35.3 million refugees. Government relocation agencies seek to place these refugees in locations where they’ll excel — find jobs, integrate into society, and develop new lives. 

But how can you find the best possible matches between immigrants and locations? 

This team developed GeoMatch, a tool that dynamically learns best synergies over time and provides placement officers with personalized recommendations for newly arriving refugees. 

The team interviewed refugees and placement officers on what variables should be included in the model — characteristics like well-paying jobs, co-ethnic national origin community, affordable housing, English language lessons, and personal transportation. They then launched the tool with Lutheran Refugee and Immigration Services and helped 2,000 refugees find placements. 

It’s key to maintain human oversight and inclusion, said team lead and professor of political science Jens Hainmueller. “The placement officer always has the final say.” The algorithm must also continuously be re-evaluated on the latest data; anything from a global pandemic to a changing job market will impact the results.

The next steps include advancing theoretical foundations around market design and estimation and launching in new countries.

MARPLE: Explaining What Happened Through Multi-Modal Simulation

Why did a car get in an accident? Perhaps we saw the driver on a cell phone. Maybe we heard a screech of tires and two loud crashes. We are confronted with all kinds of evidence from different modalities when we reason what happened. The scholars behind MARPLE hope to create a computational framework that uses multimodal evidence like sound and vision to produce human understandable explanations of what happened and why. One of the team’s projects looks at human motion — if you remove the objects a person interacts with like a chair or table, can you still reason what he’s doing? MARPLE’s use is two-fold: On the practical side, it could provide explanations for concrete events like car crashes and, more theoretically, it could probe the function of explanatory thought and what makes a good explanation. Potential applications include explanation assistants that could support legal fact finders or forensic architecture, or even in-home robotics that can investigate what’s happening and alert the owners to issues. 

AI Tutors To Help Prepare Students for the 21st Century Workforce

Stanford scholars from education, psychology, and computer science partnered to develop AI for effective, inspiring education that is both accessible and scalable. The team is creating new AI systems that model and support learners as they work through open-ended activities like writing, drawing, working on a science lab, or coding. The research monitors the learners’ motivation, identity, and competency to improve student learning. The team is testing their tools in code.org, brick-and-mortar schools, virtual science labs, and more.

Toward Grounded, Adaptive Communication Agents

How can we develop next-generation, language-based virtual agents capable of collaborating with us on meaningful, challenging tasks? Think beyond ordering a pizza — how about helping us care for patients. This research team is considering impactful assistive technologies, where a human’s behavior and language use will change over repeated interactions with a personal agent. In their work, the scholars called for a Shibboleth Rule for artificial agents, so that consumers could know when they were talking to an AI, and also examined liability when agents lie or defame. 

Reinventing Government with AI: Modern Tax Administration

Public-sector AI offers massive opportunities — who has more data and provides more services than a government overseeing 330 million people? This Hoffman-Yee team demonstrates how AI-driven, evidence-based learning can benefit U.S. government agencies through efficiencies and better delivery of services. For example, the United States Internal Revenue Service has a $400 billion gap between taxes owed and paid. The team proposed an active-learning system that uses an AI algorithm to decide which tax returns should be prioritized for auditing for more effective and fairer tax collection. The scholars on this team also found disparities in the algorithms that the IRS uses — Black taxpayers receive audit notices 2.9 times more often than non-Black taxpayers.  And this research has implications for a wide range of other governmental contexts, including environmental and health compliance.

Watch the most recent Stanford HAI Hoffman-Yee Symposium to hear from these scholars firsthand. 

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

More News Topics