Skip to main content Skip to secondary navigation
Page Content

Prior Hoffman-Yee Grant Recipients

2020 Hoffman-Yee Grant Recipients

Intelligent Wearable Robotic Devices for Augmenting Human Locomotion

Having the ability to navigate around furniture, negotiate stairs and corridors, get on and off chairs or beds, and get in and out of vehicles is essential for an individual to carry out day-to-day tasks and maintain independence. The goal of this project is to create intelligent wearable robotic devices that enable individuals with physical impairments to complete daily tasks by augmenting their capability to move in constrained spaces such as home environments. Our approach is to create an intelligent human agent which in turn trains a wearable device to be intelligent. Simply put, we use AI to teach AI. We expect the intelligent wearable devices to understand human behaviors and intentions in the context of their environments, and make decisions to complement human movement patterns and capabilities. The success of this project will expand the impact of AI to augmenting human physical skills and mobility, improving independence and quality of life for older adults and people with physical impairments.

Karen Liu PI Engineering Computer Science
Steve Collins Co-PI Engineering Mechanical Engineering
Scott Delp Co-PI Engineering Bioengineering, Mechanical Engineering, Orthopaedic Surgery
Leo Guibas Co-PI Engineering Computer Science, Electrical Engineering (courtesy)
VJ Periyakoil Co-PI School of Medicine Primary Care

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

We aim to build an intelligent learning platform that improves how it teaches as it learns more about each student and students overall. Currently, most personalized tutors simply move the student forward or backward among lessons.  No systems leverage the enormous range of techniques and strategies to inspire and enable effective learning: our course might immerse students in discovery learning, orchestrate tutoring, employ teachable agents or offer discursive opportunities depending on what it learns. Using all interactions with each student -- answers, partial work, explorations, side comments -- our course will adjust the learning experience, so students receive customized content and types of instruction. To get there, our team will engage with central questions of AI algorithms and theory. We will build and test new educational tools, targeting online programming and data science learning spaces, to: automate and widen assessments of what students know and can do, facilitate effective and inclusive online peer interactions, enhance instructor efficacy, construct narratives that engage and include students, and automatically identify which activity best supports desired learning outcomes. These threads will blend together to create a joyful new learning experience when students take “the smartest course in the world.”

Chris Piech PI Engineering Computer Science
Emma Brunskill Co-PI Engineering Computer Science
Noah Goodman Co-PI Humanities and Sciences, Engineering Psychology, Computer Science, Linguistics (courtesy)
James Landay Co-PI Engineering  Computer Science
Jennifer Langer-Osuna Co-PI Graduate School of Education  
Dan Schwartz Co-PI Graduate School of Education  

Toward Grounded, Adaptive Communication Agents

Social interactions of all kinds require continual adaptation. Consider how the language and behavior of a medical patient and their caregiver will evolve over the course of their relationship. Initially, the patient will need to rely on detailed instructions like "Please get me the eprosartan mesylate from the downstairs bathroom. I will get some water myself". As the two adapt to each other, such descriptions will simplify to "Time for my meds" -- with a sense that implies coordinated action that is specialized to their relationship. This joint process of grounding (finding the right bottle) and adaptation (what "meds" means) is fast, largely unconscious, and essential for success.

To date, work on artificial agents has largely set these processes aside. However, as we ask these agents to interact with us in more open-ended ways, their lack of physical and social grounding is leading increasingly to poor task performance; an assistive robot acting as an in-home caregiver would be a liability if it failed to adapt to its patient and context. We are addressing these problems via interwoven efforts in characterizing the cognitive and social dynamics of the phenomena and training robots and interactive virtual agents. Our core objectives are to facilitate the development of next-generation intelligent agents and to understand the broader societal effects such technologies will have.

Christopher Potts PI Humanities and Sciences Linguistics, Computer Science (Courtesy)
Judith Degen Co-PI Humanities and Sciences Linguistics
Mike Frank Co-PI Humanities and Sciences Psychology, Linguistics (courtesy)
Noah Goodman Co-PI Humanities and Sciences, Engineering Psychology, Computer Science, Linguistics (courtesy)
Thomas Icard Co-PI Humanities and sciences Philosophy, Computer Science (courtesy)
Dorsa Sadigh Co-PI Engineering Computer Science, Electrical Engineering
Mariano-Florentino Cuéllar Participating Faculty Law   

Curious, Self-aware AI Agents to Build Cognitive Models and Understand Developmental Disorders

Truly intelligent autonomous agents must be able to discover useful behaviors in complex environments without having humans available to continually pre-specify tasks and rewards. This ability is beyond that of today's most advanced autonomous robots.  In contrast, human infants naturally exhibit a wide range of interesting, apparently spontaneous, visuomotor behaviors - including navigating their environment, seeking out and attending to novel objects, and engaging physically with these objects in novel and surprising ways. In short, young children are excellent at playing - ``scientists in the crib'' who create, intentionally, events that are new, informative, and exciting to them. Aside from being fun, play behaviors are an active learning process, driving the self-supervised learning of representations underlying sensory judgments, motor planning capacities, and social interaction.

But how exactly do such young children know how to play, and how can we formalize and harness the  human play capacity to substantially improve the flexibility and interactivity of autonomous artificial agents? Here, we propose using deep neural network software agents, endowed with a mathematically formalized sense of ``curiosity'' and sophisticated perceptual capacities, to naturally generate playful behavior in novel environments. Combining cognitive science ideas with deep reinforcement learning, we seek to make a substantial leap in the fundamentals of AI.  From a cognitive science and clinical perspective, we will use these improved AI systems to build better quantitative models of development and improve our understanding of developmental disorders.

Dan Yamins PI Humanities and Sciences, Engineering Psychology, Computer Science
Mike Frank Co-PI Humanities and Sciences Psychology, Linguistics (courtesy)
Nick Haber Co-PI Education Computer Science (courtesy)
Fei-Fei Li Co-PI Engineering Computer Science
Dennis P. Wall Co-PI Medicine Pediatrics, Psychiatry (courtesy), Biomedical Data Sciences 

Reinventing Government with AI: Modern Tax Administration

Collecting revenue is a core function of government. Taxes fund nearly all public programs, from health care to environmental protection to military defense. The Internal Revenue Service (IRS) relies critically on taxpayer audits to detect under-payment and to encourage honest income reporting, but the process faces considerable challenges. 

The annual tax gap – the difference between taxes owed and paid – is nearing $500 billion. This shortfall starves the government of needed resources, while contributing to growing wealth inequality. The IRS has faced shrinking enforcement resources and a dwindling capacity to audit taxpayers for evasion or fraud, with a 42% drop in the audit rate from 2010 to 2017. Some analysts have suggested that audits excessively focus on lower-income taxpayers, and a more complete analysis and approach to these distributive concerns is important. 

In partnership with the IRS, our team is using AI and new active learning methods to help modernize our country’s system of tax collection and risk prediction. Our work seeks to develop a fair, effective, and explainable AI system for identifying tax evasion and addressing the human-centered challenges of integrating AI in a complex bureaucracy with ~10,000 diverse revenue agents, 150M taxpayers, and 1M audits annually.

Jacob Goldin PI Law


Daniel Ho Co-PI Law Political Science
Guido Imbens Co-PI Business Economics
Anne Joseph O'Connell Co-PI Law Law
Jure Leskovec Co-PI Engineering Computer Science
Rebecca Lester Co-PI Business Business

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

Humans understand the world via concepts, models that express our conception of ourselves and our society. We develop new AI technology to help humanists and social scientists trace how concepts develop and change over time and how concepts differ between groups. Our goal is to build a new kind of microscope for studying the dynamics of concept change and culture, using natural language processing and drawing on online texts from different languages, genres, and time periods, to answer deep humanistic and social science questions. Our work also enriches our understanding of how AI systems like neural network language models represent concepts themselves. Our multidisciplinary team from the humanities, social sciences, computational sciences, and the library ask questions like: How are complex concepts represented (including textual and visual elements) in modern neural networks? How do concepts become moralized over time? How do our conceptions of immigration and immigrants change with successive waves of immigrants? How do conceptions of gender and race vary across time and geography, and what are the implications for sociology and the history of science? What are the legal implications (for example for legal originalism) of conceptual and word meaning varying between groups or over time? This project applies AI to research into the social and historical dimensions of human thought, opening up new perspectives on both human and machine intelligence.

Ran Abramitzky Co-PI Humanities and Sciences Economics
Mark Algee-Hewitt Co-PI Humanities and Sciences English
R. Lanier Anderson Co-PI Humanities and Sciences Philosophy
Dan Edelstein Co-PI Humanities and Sciences French & Italian, History (courtesy)
Julian Nyarko Co-PI Stanford School of Law Law
Dan Jurafsky Co-PI Humanities and Sciences, Engineering Linguistics, Computer Science
Alison McQueen Co-PI Humanities and Sciences Political Science
Londa Schiebinger Co-PI Humanities and Sciences History
Rob Willer Co-PI Humanities and Sciences, Business (by courtesy) Sociology, Psychology (by courtesy)
Jamil Zaki Co-PI Humanities and Sciences Psychology
James Zou Co-PI School of Medicine Biomedical Data Science, Computer Science (by courtesy), Electrical Engineering (by courtesy)
Catherine Coleman Senior Personnel