Seed Grant Recipients 2022
The Seed Research Grants are designed to support new, ambitious, and speculative ideas with the objective of getting initial results.
-
How might we augment human capabilities by enabling people to practice challenging interpersonal interactions? Workers in an organization, for example, regularly face challenging conversations with their managers and colleagues. Such skills today are typically taught by experts in person or via videoconference, which is time-consuming and costly. Access to trainers who can teach these skills to learners is limited. Can we make such training tools more broadly accessible, developing processes to help prepare people for these challenging interpersonal interactions? Our project will develop interpersonal skill training tools that embody simulated social roles from large pretrained language models such as GPT-3. We will innovate algorithms and interfaces for responsibly eliciting this wide variety of simulated actors from a large language model, and exploring how these large language models can be adapted responsibly to various social roles in different contexts to achieve positive outcomes.
Name
Role
School
Department
Main PI
School of Engineering
Computer Science
Co-PI
School of Engineering
Computer Science
Co-PI
School of Humanities and Sciences
Communication
-
Digital media platforms (e.g. social media, video sharing, online video games, etc.) often feature attention-seeking and attention-absorbing design elements that are powered by AI algorithms to maximize user interaction and time spent engaging with the experiences. Beyond streamlining online experiences, such algorithm-fueled excessive consumption of digital media may constitute addictive behavior patterns and may impose detrimental effects on mental health. With the growing interest from policymakers to regulate such mechanisms, we need better evidence about the addictive risks of AI-powered engagement mechanisms to meaningfully inform legislation to reduce harm. Our work takes a user-centric, real-world approach to analyzing digital addiction. We focus on collecting real-time, first-hand data of usage patterns on smartphones, including usage patterns across specific apps, and the addictive platform features engaged with. This approach will inform human-centered policy solutions to curtail design features driving digital addiction. In addition, to allow for ongoing syndromic monitoring, we seek to develop self-report measures of AI-powered addiction of digital experiences, leveraging the team’s expertise in addiction, survey design and psychometric evaluation.
Name
Role
School
Department
Main PI
School of Medicine
Psychiatry and Behavioral Sciences
Co-PI
School of Humanities and Sciences
Psychology
-
Neuropsychiatric symptoms (NPS) are common but often overlooked in older adults, especially patients with or at risk for dementia. Unintrusive, automatic assessment of triggers, frequency, severity, and associated caregiver distress of NPS would be critical for devising tailored clinical care management plans. Currently, assessing NPS relies primarily on clinician observation and interviews with caregivers, usually within brief and infrequent clinical visits, resulting in highly subjective clinical decision-making and either undertreatment or overtreatment. In this project, we are developing a personalized and clinically interpretable NPS assessment system (referred to as AgeWell) using ambient intelligence technologies, which enables 1) monitoring individual and subsyndromes of NPS (e.g., mood vs. agitation); and 2) detecting early changes in NPS by automatically quantifying and detecting relevant behavioral abnormalities based on both intra- and inter-individual comparisons against published goal standards, Neuropsychiatric Inventory (NPI) and the Mild Behavioral Impairment Checklist (MBI-C).
Name
Role
School
Department
Main PI
School of Medicine
Psychiatry and Behavioral Sciences
Co-PI
School of Medicine
Psychiatry and Behavioral Sciences
Co-PI
School of Medicine
Psychiatry and Behavioral Sciences
Co-PI
School of Medicine
Psychiatry and Behavioral Sciences
Co-PI
School of Medicine
Med/Hospital Medicine
-
AI will influence the future of people from many diverse backgrounds. There is an urgent need to incorporate the perceptions, imaginings, concerns, and creativity of diverse groups into AI developments. In the US, many AI products tend to reflect values and philosophical traditions of European American contexts. The fact that people in Black, Latinx, Asian and Indigenous communities may have different ideas and concerns when it comes to AI and its potential has received much less attention. The rapid advancement of AI can further propagate cultural views of groups that are currently in power, potentially creating a self-reinforcing cycle of inequity.
The lack of a diverse cultural grounding for AI theorists and developers limits the space of the imaginary. Tapping into a variety of cultural ideas can catalyze new categories of AI and create broader societal and environmental benefits. Currently, inclusive AI developments are hindered by a dearth of first-hand knowledge about how underrepresented groups actually imagine and think of AI. We propose to empirically study conceptions of AI held by group members from diverse cultural backgrounds. Our efforts will be a necessary first step to guide critical applications of cultural insights to broadening the intellectual foundations of AI.
Name
Role
School
Department
Main PI
School of Humanities and Sciences
Psychology
Co-PI
School of Humanities and Sciences
English
Co-PI
School of Humanities and Sciences
Psychology
-
Scientists and journalists often communicate research findings to lay people through data visualizations and text (e.g., the positive and negative effects of caffeine intake, the effectiveness of indoor masking in reducing infection, the impact of human activity on global climate, etc.). Well crafted articles allow readers to attend to both text and chart elements to contextualize data, extract factual knowledge, build connections between data and concepts, as well as generate and check hypotheses. For example, in an article about Rising Global Temperatures (Figure 1), the text, might reference an upward trend in global temperature between 2000 and 2019, or the point with the highest value in 2016. At each such text reference, an active reader can choose to examine the chart, identify the aspect of the data being emphasized by the text (e.g. the segment of the time series between 2000 and 2019), and may also see the surrounding context (e.g., an El Niño caused record setting temperature in 2016).
In this case, the text incorporates information that is not present in the visualization to explain the feature it emphasizes. In this project, the team will explore AI-augmented reading experiences that link text with visualizations to facilitate information extraction, processing, and sensemaking. By developing computational models of reader behavior, the team will create mixed-initiative tools for readers of interactive documents. For instance, our tools will guide active reading by automatically highlighting correspondences between text and visualization, will prompt readers to externalize their knowledge through annotations and interactivity, and will support interactive question-answering.
Name
Role
School
Department
Main PI
School of Engineering
Computer Science
Co-PI
Graduate School of Education
Graduate School of Education
-
Electronic retinal implants are a promising solution for restoring vision to those affected by retinal degenerative diseases by stimulating surviving retinal neurons to send artificial visual signals to the brain. However, the limited number of electrodes used in all present-day devices, compared to the 1.25 million optic nerve fibers, restricts their performance. To overcome this limitation, we propose an innovative solution by using AI to better exploit the bandwidth of retinal implants. Our solution leverages the fact that most visual tasks, such as navigating in an indoor environment, only require a tiny fraction of the visual information incident on the retina. This is because most of the detailed content of an image is irrelevant to the task at hand. Therefore, a device that filters out the irrelevant image content and instead represents the visual scene in a simplified format, such as a cartoon sketch of the landscape for navigation, can provide more useful artificial vision with a limited-bandwidth device. This concept applies to various visual tasks, including object identification, facial expression recognition, and reading. We are using modern AI tools for image simplification and abstraction to improve the performance of human observers on various tasks in a simulated low-bandwidth visual environment, with the goal of using the same software in retinal implants.
Name
Role
School
Department
Main PI
School of Medicine
Neurosurgery
-
Cancer cells evade their host’s immune system by negatively regulating T cells via immune checkpoints. By blocking these checkpoint proteins, the ability of the immune system to recognize and kill cancer cells restores. Despite the promising results, individual response rate of checkpoint blockade varies greatly among patients. Furthermore, there exist no good biomarkers to predict an individual’s response to immune checkpoint inhibitors (ICI). In this project, we will develop and validate a novel deep learning strategy for stratification of immunotherapy patients based on the patient’s gene mutation data. We will first develop a transformer neural network-based model with explicitly incorporation of prior biological knowledge about the gene pathways by encoding the protein-protein interactions via graph convolution. A novel method to embed individual’s binary mutation data to encode both PPI and gene occurrence information will be established to facilitate the inference process. Combined with the self-attention mechanism of the transformer model, the proposed framework proposes to yield the interpretable biomarkers for individualized cancer immunotherapy with immune checkpoint Inhibitors.
Name
Role
School
Department
Main PI
School of Medicine
Radiation Oncology
Co-PI
School of Humanities and Sciences
Statistics
-
Patients present to the Emergency Department (ED) with diverse and rapidly progressive disease states for which the necessity and optimal timing of life-saving interventions is difficult to predict. ED patients are routinely connected to sophisticated monitors that continuously measure several dimensions of physiology, which are known to predict a wide variety of disease outcomes. Yet the vast majority of this monitoring data is discarded, and is inaccessible to clinicians. We propose to develop and validate PhysioHub, an AI-powered analysis and prediction platform for patient physiologic monitoring. PhysioHub will synthesize real-time data from the electronic health record and from continuously monitored vital signs and physiologic waveforms, and deploy multimodal deep learning models to enhance the ED physician’s ability to understand her patients’ physiology, clinical risks, and likely responses to therapeutic interventions. We aim to augment the capability of ED physicians to target optimal care at the most critical periods in the clinical course, while reducing the burden of repetitive monitoring tasks and distracting false alarms.
Name
Role
School
Department
Main PI
School of Medicine
Emergency Medicine
-
Unilateral facial paralysis is quite common in the United States and worldwide, with the majority of cases attributed to Bell's palsy. Current methods of assessing unilateral facial palsy require that physicians manually score disease severity, which is both time-consuming and inherently subjective. Additionally, existing efforts to automate this process are limited to analyzing still-frames from patient videos. In response to these limitations, we, in partnership with The Stanford Facial Nerve Center, have developed a video-based algorithm that dynamically assesses facial symmetry using computer vision and deep learning. Our algorithm has been trained on patient videos scored with the eFACE metric, a clinician-validated grading scale that assesses palsy severity. It uses accurate facial landmark detection to track dynamic unilateral asymmetry and produce discrete measurements. Our aim is to create a standardized and automated system of scoring unilateral facial palsy. In doing so, we hope to not only improve the care patients receive and decrease clinicians' workload, but to also increase access for those unable to receive care from a specialist by embedding the algorithm into a mobile application.
Name
Role
School
Department
Main PI
School of Medicine
Otolaryngology - Head & Neck Surgery
Co-PI
School of Medicine
Biomedical Data Science
-
Illegal fishing is a global issue that threatens our food security and livelihoods. A key to preventing illegal fishing is seafood traceability, but a common practice of transshipping catches between vessels at sea could obscure the source of the catches if it is not properly monitored. By contrast, transshipment in ports is considered to offer more tractable monitoring and enforcement, making them a promising approach for enhanced traceability. However, there is currently no systematic tool to observe in-port transshipments, and consequently, we still have a poor understanding of broad patterns of transshipment in ports. Our project aims to create a new capacity to track in-port transshipments from satellite imagery by harnessing recent AI developments toward better fisheries monitoring and enforcement.
Name
Role
School
Department
Main PI
School of Sustainability
Center for Ocean Solutions
Co-PI
School of Humanities and Sciences
Statistics
Co-PI
School of Sustainability
Center for Ocean Solutions
Co-PI
School of Medicine
Biomedical Data Science
-
The overall goal of this project is to enable robots to creatively solve physical puzzles by manipulating or combining objects in novel ways. While robots are more and more capable of manipulating a variety of objects, humans distinguish themselves in their capacity to generalize their manipulation skills to novel objects, scenarios, or goals - often done in only a handful of attempts of trial and error. Currently, robots lack this creativity, flexibility, and fast adaptation which ultimately will be required when deploying them into environments they have never seen before or into environments that their creators did not anticipate (e.g. disaster zones or other planets). The research proposed here develops and tests computational models of physical problem solving, serving the twin goals of (1) equipping robots with this capacity, and (2) better understanding the underlying cognitive mechanisms that support this capacity in humans. The core research question addressed in this proposal is how an agent can effectively reason about the infinitely many ways of interacting with novel objects in order to achieve various goals in previously unseen scenarios.
Name
Role
School
Department
Main PI
School of Engineering
Computer Science
Co-PI
School of Humanities and Sciences
Psychology
-
There is a growing interest in neuromodulation techniques for their potential to treat various mental conditions. In particular, repetitive transcranial magnetic stimulation (or TMS) of the dorsolateral prefrontal cortex (DLPFC) is an FDA approved treatment for patients affected by drug-resistant major depression. Several studies have demonstrated the benefits of personalized TMS targeting that accounts for inter-patient brain structure and head shape variability rather than the current standardized approach. For clinicians, there are trade-offs between the targeting precision and the time and complexity of the targeting protocol. Here we leverage evidence that there is a correlation between head features and brain anatomy in humans, to train a compact AI model to predict brain target regions based on a patients’ head features. To assure better model performance, the supervised training phase will be guided by the specific information contained on the MRI scans following a student-teacher approach. As an initial step, the developed model will be focused on localizing the DLPFC for treating depression symptoms. A similar design can be used at a later stage to improve the external targeting of other cortical regions.
Name
Role
School
Department
Main PI
School of Medicine
Radiology
Co-PI
School of Medicine
Psychiatry and Behavioral Sciences
Co-PI
School of Medicine
Psychiatry and Behavioral Sciences
-
Online advertising largely sustains the news we consume everyday. The use of AI-powered algorithmic bidding and targeting in advertising affects the revenue generated by different news websites, thereby incentivizing the creation of different types of content. However, despite concerns that the digital advertising business model inherently amplifies harmful content, the presence of a bias towards publishing such content based on its potential for monetization by ads has not been empirically investigated. To address this gap, our project investigates whether the advertising-based business model of mainstream digital platforms amplifies partisanship and misinformation. Specifically, we will examine 1) whether advertising on misinformation and partisan content performs better than advertising on other content, and 2) which types of ads are more likely to be algorithmically placed on misinformation and partisan websites. In examining the factors underlying machine behavior in this setting, we will uncover how algorithmic advertising shapes the incentives of which news gets monetized and sustained.
Name
Role
School
Department
Main PI
Dean of Research
Human-Centered Artificial Intelligence/Digital Economy Lab
-
A major challenge in containing wildfires is monitoring and preventing spot fire spread. Spot fires are an increasingly prevalent route of wildfire spread in which the main fire launches burning pieces of wood called firebrands downwind that ignite new fires upon landing. Spot fires are notoriously unpredictable and can ignite as far as a mile or more from the main fire. Physics-informed machine learning could be a powerful tool for analyzing spot fire spread, but a lack of understanding of the rich physical dynamics governing firebrand transport prevents the development of effective algorithms. To address this, we are performing laboratory experiments to examine the mechanisms that guide firebrand transport. We model firebrand transport in a water flow-tunnel by releasing particles (firebrands) in a heated plume (smoke column) in a canopy (forest) subject to a current (wind). By utilizing standard fluid mechanics modeling approaches we are able to examine the impact of various physical features such as forest structure, wind speed, and flame intensity on the variability of firebrand trajectories. These features are typically already known from existing wildfire data sets or can be easily obtained for future fires, so the results from our experiments will enable physics-informed modeling of spot fire spread. Achieving better spot fire spread models should, in turn, lead to more effective mitigation strategies.
Name
Role
School
Department
Main PI
School of Sustainability
Civil and Environmental Engineering
Co-PI
School of Sustainability
Civil and Environmental Engineering
-
Abstract coming soon.
Name
Role
School
Department
Main PI
School of Engineering
Mechanical Engineering
Co-PI
School of Engineering
Computer Science
Co-PI
School of Medicine
Surgery
-
Our team's primary research aim is to study how medical AI algorithms should be evaluated and regulated. While AI algorithms have made significant strides in the research world, only a fraction of them have received regulatory approval, and even fewer actually reach the patients they are intended to serve. An imminent challenge in medical AI lies in ensuring its safe and effective deployment in healthcare settings. Our team's primary research goal is to study the gap between research and deployment through the lens of clinical trials and regulation. Our research will focus on a) advancing a data-driven understanding of the regulatory landscape for medical AI in the US and b) developing computational methods for assessing and reducing performance disparities in evaluation and deployment.
Name
Role
School
Department
Main PI
School of Medicine
Biomedical Data Science
Co-PI
School of Law
Law School
-
In this proposed research we will pilot an approach to: 1) identify ethical issues that may emerge with development and multi-site deployment of a healthcare AI (AI-HC) application for risk assessment of pulmonary embolus (PE); and 2) develop consensus on how to address these ethical issues, once identified. We will also 3) develop consensus on an ethics “label” to communicate ethical constraints. In doing 1, 2 & 3 we will refine a generalizable approach for identifying and addressing ethical challenges with an AI-HC as well as provide guidance for how to communicate ethical concerns with future AI-HC.
No systematic approach has yet emerged regarding how to survey the landscape of AI-HC conception, development, calibration, implementation, evaluation, and oversight. Bereft of any conceptual map of this landscape, the identification of ethical concerns arising from this emerging, complex, cross-disciplinary technology that potentially affects many aspects of healthcare has thus far been reactive, ad hoc, and fragmented. This is problematic, especially for so-called “wicked” problems, which unlike more straightforward and “tame” technical problems, typically defy a singular formulation of the problem, are nested within systems that have interrelated problems, and have social values woven into their fabric such that solutions are not simply true or false but rather better or worse. In this research, we aim to enhance our ability to identify – proactively, systematically, and in a more thoroughgoing and integrated manner – the variety of ethically relevant decisions and their ethically relevant consequences regarding AI-HC (through this initial case study of AI-HC for PE).
Name
Role
School
Department
Main PI
School of Medicine
Anesthesiology
Co-PI
Law School
Law School
Co-PI
Law School
Law School
Co-PI
School of Medicine
Biomedical Informatics
Co-PI
School of Engineering
Management Science and Engineering
-
Large language models (LLMs) have increasingly been incorporated into real-world settings, including high-stakes use cases that involve value judgments such as content moderation and toxicity detection. Such use cases, however, may require language models to make value judgments that are shaped by cultural and societal norms that can differ vastly across different parts of the world. If an LLM is biased toward a particular value system or culture (e.g., Western or American), it could result in adverse downstream effects such as imposing hegemonic worldviews and homogenizing users’ perspectives and belief systems. Therefore, we propose to study the cultural views and values reflected by LLMs, and how these representations may shape human perception of different cultures.
Name
Role
School
Department
Main PI
School of Engineering
Computer Science
-
The symbolic AI models of the mid-twentieth century were designed to leverage valuable insights about causal structure, ranging from commonsense reasoning to advanced scientific knowledge. These insights were expressed directly in hand-designed structures, and this ensured that model behaviors were systematic and human-interpretable. Unfortunately, these models tended also to be brittle and narrowly specialized, limiting their utility. By contrast, present-day AI models are general-purpose and data-driven, and they can flexibly acquire a range of complex behaviors, which has opened up many new avenues for productive applications of AI. However, the trade-offs are now evident: these models often find opaque, unsystematic solutions, and they are completely dependent on available data. As a result, a model that performs well in testing can turn out to be dangerous and unpredictable once deployed in the wider world. To address this, we have begun to lay the foundation for allowing AI models to incorporate important insights about causal structure while still learning in a data-driven fashion. In this method, we define high-level causal models, usually in symbolic terms, and then train neural networks to conform to the structure of those models while also learning specific tasks. The central technical piece of this proposal is interchange intervention training (IIT), in which we swap internal representations in the target neural model in a way that is guided by the input–output behavior of the causal model. IIT objectives are easily defined, fully differentiable, and provide a formal guarantee that, if the loss is minimized, the target model is a causal abstraction of the neural network. For the present project, we will use IIT to induce causal structure in massive foundation models and to distill large language models, and we will more robustly quantify the extent to which IIT leads models to robustly acquire capabilities that are only sparsely represented in the training data.
Name
Role
School
Department
Main PI
School of Humanities and Sciences
Linguistics
Co-PI
School of Humanities and Sciences
Philosophy
-
Cancer immunotherapy represents one of the most important advances in modern medicine. Specifically, the discovery of immune checkpoints led to the regulatory approval and clinical use of a number of therapeutic agents for cancer treatment. Immunotherapy has dramatically improved long-term survival for some patients and is now the standard of care for treating most advanced cancers. Unfortunately, only a small proportion of patients respond to immunotherapy and would experience clinical benefit. Moreover, these drugs may cause serious side effects and are substantially more costly. Therefore, it is crucial to identify, on a personalized basis, which patients will respond to and benefit from immunotherapy. In this project, we propose to develop interpretable AI systems by integrating multi-modal radiology, pathology, and clinical data for predicting immunotherapy response in cancer patients. Because this information is obtained from data that are routinely available, the proposed model could be readily integrated into clinical workflows and have a positive impact on cancer patients.
Name
Role
School
Department
Main PI
School of Medicine
Radiation Oncology - Radiation Physics
Co-PI
School of Medicine
Radiation Oncology - Radiation Therapy
-
Over the last few years, millions of children have had short or long term disruptions to in-person formal school. While AI educational tools could be a key part of the solution, ensuring such software is effective is often challenging. In part this is because the outcomes of interest, such as passing a class or graduation, are often delayed, which makes it hard for adaptive methods like reinforcement learning to learn quickly enough to be practical. Other methods, such as static causal discovery, often are designed for a single environment and may not transfer to different student distributions or when new pedagogical content or interventions are introduced. In collaboration with our partner War Child, that supports children in conflict regions, we will use and advance new methods to develop automatic ways of identifying short-term proxy variables that can be used for data-driven AI software optimization that can transfer.
Name
Role
School
Department
Main PI
School of Engineering
Computer Science
-
Emerging adulthood—the transitional period from late teens through mid-twenties—is a particularly important time for studying social, economic, and health risk factors. During this period, emerging adults experience multiple changes to their daily lives, such as leaving their family home, gaining independence in activities of daily living, and increased social, financial, and academic responsibilities. As a result, emerging adulthood is a pivotal point for health over the life course, yet traditional epidemiologic methods are severely limited in their ability to measure human behavior during dynamic periods. Traditional epidemiologic data collection often relies on self-reported behavior and has long gaps between collection waves. This project seeks to leverage the ubiquity of smartphone ownership among emerging adults, combined with modern algorithmic inference, to quantify dynamic human behavior and exposures at fine spatiotemporal resolution. We seek to understand how comparable metrics based on traditional data collection are to those derived from novel, digital phenotyping data collection and build AI-based models to compare how predictive these metrics are of different health behaviors. Importantly, a primary goal is to understand the experiences of participants themselves, especially among traditionally excluded groups such as lower-income or racially minoritized participants, to understand how smartphone data collection can be used to improve both data quality and gather data from more generalizable samples.
Name
Role
School
Department
Main PI
School of Medicine
Epidemiology and Population Health
Co-PI
School of Humanities and Sciences
Communication
Co-PI
School of Humanities and Sciences
Psychology
-
Our team will explore applications of artificial intelligence in the study of artistic creative process. We will leverage technical resources from fields such as machine learning and computer vision, integrate contemporary research perspectives from the humanities, and focus on artists and artworks connected to Stanford’s Asian American Art Initiative. Representative technical aims include the use of shape fingerprinting and manifold learning methods to characterize the geometry of an artist’s trajectory in creating new forms over time; application of mixed-modal clustering to glean new insights from digital artists’ archives; and investigation of principled approaches to surfacing stereotypes, stemming from or projected onto art objects, that may be embedded in large generative models.
Name
Role
School
Department
Main PI
School of Humanities and Sciences
Applied Physics
Co-PI
School of Humanities and Sciences
Art and Art History
Co-PI
School of Humanities and Sciences
Art and Art History
-
Geometric principles are the basis of engineering, architecture, and other human activities. Furthermore, geometry as a discipline has been used for centuries to scaffold the development of rigorous mathematical thinking in students. How do humans learn to generalize geometric concepts and prove geometric propositions? Can AI systems for geometric reasoning augment human learning during formal education? We aim to answer these questions by developing a framework for constructing geometric principles, studying human cognition compared to computer vision algorithms, and creating neuro-symbolic systems to model geometric reasoning. These components will allow us to understand human sensitivities to geometry and develop AI systems to solve increasingly complex, real-world geometric programs — achieving our goal of leveraging AI to challenge and tutor human geometry students.
Name
Role
School
Department
Main PI
School of Humanities and Sciences
Psychology
Co-PI
School of Engineering
Computer Science
-
Abstract coming soon.
Name
Role
School
Department
Main PI
School of Humanities and Sciences
Psychology
Co-PI
School of Humanities and Sciences
Linguistics
-
The human eye exhibits remarkable capabilities, but in many situations it would be desirable to have vision and imaging capabilities that greatly surpass those of our eye. In our research, we aim to address this important challenge by creating new optical technologies that are inspired by human intelligence and can augment our own vision capabilities. We will aim to realize such optical elements in a compact and flat form factor, so that they can easily be integrated into regular eyewear. With this type of optics, humans will be able to experience and learn more about the world around us. In our opinion, the current state-of-the-art in creating enhanced vision and imaging capabilities has barely scratched the surface of what could be possible.
Name
Role
School
Department
Main PI
School of Engineering
Materials Science and Engineering
-
Genome-Wide Association Studies (GWAS), Polygenic Risk Scores (PRS), and phenotype prediction from genotype data (geno-to-pheno) are providing data-driven solutions to predict, understand, and screen for a wide range of diseases using DNA sequence data. Such techniques provide tools for physicians to augment their diagnoses and personalize their treatment plans and can also give a clearer picture of the causal relationships between DNA and traits to aid in the search for drug targets. However, most of the existing genomic studies and models only incorporate white (European-descent) individuals and in the majority of cases these models do not transfer well to other ancestries. Additionally, such single ancestry models cannot be applied to individuals who belong to no single model group. This category includes over one third of the children born today in the United States, who are of multiple ancestries. In our increasingly diverse society, disease risk prediction models that fail for such multi-ancestry groups are not tenable. We want to address the lack of well-established and accurate algorithmic solutions for incorporating population structure and admixture information along the genome to perform association studies through developing techniques that work across ancestries using self-supervised learning (SSL), transformers, and large neural network models trained on large biobanks.
Name
Role
School
Department
Main PI
School of Medicine
Biomedical Data Science
Co-PI
School of Medicine
Biomedical Data Science
-
The transmission of an infectious disease is fundamentally a social event. In general, for an infection to be transmitted, a susceptible individual needs to be brought into contact with an infectious individual. The likelihood that these two agents will interact in a way that will cause disease transmission depends on the behavior of the agents, the norms and institutions that shape their behavior, and the social structures in which they are embedded. Despite the central importance of social structure and human behavior, remarkably little research has focused on how such behavior should be represented in mathematical and computational models of infectious-disease transmission dynamics. This is a real problem since these models are fundamental tools for developing effective interventions for highly-consequential societal phenomena. Our project will adopt the tools of "Scientific Machine Learning" (SciML) to address this important problem. We will use machine learning to train neural networks on a wide range of possible functional forms for the interaction terms. These neural networks can then serve as "universal approximators"---basically, extremely general, data-informed functions---that we can include in standard transmission-dynamics models. AI tools have heretofore been of limited utility at the outset of epidemics---when we need forecasting the most---because of the sparseness of data as outbreaks get started. SciML, by training AI on mathematical models, provides a novel opportunity to augment our current capacities and harness the power of AI for social good. SciML allows us to train from the small data characteristic of the outset of outbreak, extrapolate outside of the original data sets, as is essential as outbreaks become epidemics and/or pandemics, and interpret our results back into general biological and social mechanisms, as is necessary for developing more generalizable insights.
Name
Role
School
Department
Main PI
School of Sustainability
Earth System Science
Co-PI
Graduate School of Business
Graduate School of Business
Co-PI
School of Humanities and Sciences
Political Science
Co-PI
School of Humanities and Sciences
Biology
-
Machine Learning has had substantial impact on diverse sectors of the industry, science and society. However, despite ML growth pervasive growth, the participation of historically under-represented groups in the ML workforce and research is extremely low. Studies have suggested that the problems of racial bias in ML systems arise due to the lack of diversity in the environment that they are shaped in including the non-representative nature of the people who develop these systems. These issues are amplified in the realm of Scientific Machine Learning (SciML) which is generally described as Machine Learning applied to data generated by physical experiments and simulations. Physical sciences students from under-represented groups are affected by limited access to SciML teaching and training. We will initiate a joint SLAC-ICME program, funded by the HAI seed grant, that makes Scientific Machine Learning accessible to students from historically under-represented groups. In addition to SciML teaching through workshops, the program will offer follow-up training and internships to enable students to apply their knowledge under the guidance of leading experts. Finally, it will provide mentorship and community building, so that students can have fruitful careers and become change-agents.
Name
Role
School
Department
Main PI
School of Engineering
Mechanical Engineering
Co-PI
SLAC
Machine Learning Initiative
Co-PI
SLAC
Machine Learning Initiative
-
The operating room is a highly demanding environment in which surgeons must process vast amounts of information to make timely and informed decisions that can greatly impact surgical outcomes. However, prolonged working hours coupled with the well-established effects of fatigue on cognitive and technical performance can lead to a significant potential for error, with detrimental consequences to patient health. To mitigate these human failings, we focus on developing SuTr, a foundational vision model for surgery with the ability to provide real-time intraoperative guidance and analysis. By leveraging the latest advancements in artificial intelligence and machine learning, SuTr aims to gain a holistic understanding of the operative scene to provide actionable insights into enhancing surgical precision and techniques, with the goal of decreasing error rates and complications, and paving the way towards autonomous robotic surgery.
Name
Role
School
Department
Main PI
School of Medicine
Cardiothoracic Surgery
-
The Technology & Racial Equity Field Incubator will convene Stanford faculty to map the problem space at the intersection of technology and racial equity and identify key interventions to build this nascent field of research. We understand race and technology as vectors that cut across academic disciplines. Given the ubiquity of technology and the fact that racial inequality runs through the fabric of our society, these issues will inevitably intersect with every field of study. The Technology & Racial Equity Field Incubator is designed to bring new voices into the conversation and uncover research questions that are not currently being addressed. By convening faculty from the Schools of Engineering, Humanities & Sciences, Medicine, Business, Education and Law, this program will catalyze interdisciplinary partnerships and provide training, resources, and thought-partnership to empower faculty so they can become leaders in the emerging field of race and technology.
Name
Role
School
Department
Main PI
School of Humanities and Sciences
Center for Comparative Studies in Race and Ethnicity
Co-PI
School of Humanities and Sciences
English
Co-PI
School of Humanities and Sciences
Center for Comparative Studies in Race and Ethnicity
Co-PI
Management Science and Engineering
School of Engineering
Co-PI
School of Humanities and Sciences
History
-
Sharing neuroimaging data is a scientific imperative to enhance transparency and reproducibility in neuroimaging research and accelerate discoveries on how the brain functions in both the healthy and the disease states. Yet, at the same time, data sharing requires rigorous privacy measures to safeguard subjects’ identities and other sensitive information. Existing methods, such as partially removing or blurring facial features in structural MR images, along with deleting direct individual identifiers from metadata, have been considered to provide sufficient privacy protection for neuroimaging data. However, recent studies have shown that it is possible to reidentify subjects by matching facial images reconstructed from structural images with photos of each individual using advanced face recognition. Our project aims to develop a technical measure to counteract this novel threat to privacy and examine the impact of the increased risk to data privacy on neuroimaging research and open science practice. We will first design a new defacing technique that can fool the face recognition algorithm by adding small adversarial perturbations on the face surface of structural MR images. We will then investigate the ethical and regulatory implications of the increased reidentification risk in neuroimaging data sharing and examine policy strategies to better protect neuroimaging data. This interdisciplinary project will inform neuroscientists, ethicists, policymakers, and the public, as potential subjects, about the privacy risk associated with neuroimaging data posed by advanced AI algorithms. Further, this project will lay the groundwork for developing technical and regulatory safeguards to address the risk.
Name
Role
School
Department
Main PI
School of Humanities and Sciences
Psychology
Co-PI
School of Engineering
Computer Science
-
Contracts are a ubiquitous legal artifact. When we purchase health insurance, move homes, accept a Terms of Service agreement, or start a new job, we sign contracts along the way. Contracts are encased in natural language. Accordingly, contractual terms are inherently vague and context-dependent. This is both a bug and a feature. Disputes may arise between contracting parties due to divergent interpretations of vague terms such as reasonable, good faith, or pollution. Regardless of how such disputes are resolved, and even if a dispute never makes its way to court, both sides may incur costs in the form of time, energy, and money. In practice, most people lack the legal resources necessary to effectively draft, interpret, and analyze natural language contracts. Computational law offers the promise of democratizing access to such resources through the development of widely-available assistive analytic tools. Beyond such tools, a further class of solutions known as “computable contracts” holds the promise of reducing the incidence of contractual disputes by encoding contractual terms in unambiguous formal computer code rather than natural language. However, both assistive computational tools and computable contract infrastructures are, at present, hindered by a limited understanding of the prevalence and role of linguistic vagueness in contract drafting and interpretation. Researchers in this space have only scratched the surface when it comes to reformulating vague natural language in computable code.
Our project will support the development of a) the next generation of computational tools that can assist in the drafting, interpretation, and analysis of contracts; and b) computable contract infrastructure capable of addressing linguistic ambiguity beyond the structural level. We envision open-source tools that simulate how humans use and interpret vague natural language, including via the strategic use of vagueness when social conditions permit. Our project also aims to empower individuals – particularly those without access to conventional legal resources – to make informed legal decisions with the assistance of responsible and carefully produced computational aides. This project will not only be an interdisciplinary venture, but also one that is possible through industry collaborations with AXA, DLA Piper, Lex Machina, and Casetext.
Name
Role
School
Department
Main PI
School of Humanities and Sciences
Linguistics
Co-PI
School of Humanities and Sciences
Linguistics
Co-PI
School of Engineering
Computer Science
Co-PI
School of Humanities and Sciences
Philosophy
Co-PI
Law School
Law School