Fei-Fei Li

Fei-Fei Li is a Chinese-born American computer scientist whose research spans computer vision, machine learning, and visual intelligence. She led the development of ImageNet, a large-scale collection of labeled images, and helped establish its visual recognition challenge. These collaborative projects provided training data and benchmarks that supported advances in deep learning and automated image recognition. (en.wikipedia.org)

A professor at Stanford University, Li served as director of the Stanford Artificial Intelligence Laboratory from 2013 to 2018 and became a founding co-director of Stanford’s Institute for Human-Centered Artificial Intelligence in 2019. Her work also encompasses AI-assisted healthcare, inclusive computing education through AI4ALL, and spatial intelligence through World Labs, which she co-founded. Her career connects research on how machines interpret visual environments with institution-building, industry research, and public discussion of the development and uses of artificial intelligence. (profiles.stanford.edu)

Early life and scientific education

Fei-Fei Li was born in Beijing, China, and grew up in Chengdu, Sichuan. In 1992, she immigrated to New Jersey with her family and subsequently attended Parsippany High School. Her parents encouraged her interest in science and technology, while she supplemented the family’s income through part-time work in a Chinese restaurant. Her early scientific interests included fundamental physics: at Princeton’s 2024 Alumni Day, she recalled working in the restaurant during the day and thinking about special relativity at night. This recollection places her enthusiasm for scientific explanation before her university training in computing and perception. (cs.stanford.edu)

Li attended Princeton University from 1995 to 1999, earning a B.A. in physics with high honors. Alongside her major, she completed certificates in engineering physics and computational and applied mathematics. Her undergraduate education thus combined physical theory with mathematical and computational methods, rather than following a conventional computer-science degree program. During these years, she also helped run her family’s dry-cleaning business in Parsippany, returning home on weekends while maintaining her university studies. In 1999, she received Princeton’s Kusaka Memorial Prize in Physics. (cs.stanford.edu)

Between her undergraduate and doctoral studies, Li spent 1999–2000 in Tibet conducting research on Tibetan medicine, supported by Princeton’s Martin Dale fellowship. She then entered the California Institute of Technology in 2000 for graduate study in electrical engineering, receiving a master’s degree in 2001 and a Ph.D. in 2005. Her graduate training brought together computational approaches to recognition and experimental investigation of human vision. Pietro Perona was her dissertation adviser; her autobiographical account also identifies Christof Koch as an adviser, and Caltech records Koch among the members of her doctoral committee, alongside Irving Biederman and Yaser S. Abu-Mostafa. (cs.stanford.edu)

Her 2005 dissertation, Visual Recognition: Computational Models and Human Psychophysics, examined how people and machines recognize objects and scenes. Its human-vision studies used controlled experiments to investigate what observers could perceive in briefly presented photographs and how recognition depended on attention. Its computational studies addressed learning object categories from very few examples and categorizing natural scenes. These were complementary research methods: behavioral experiments measured human perceptual performance, whereas computational models tested explicit procedures for learning and recognition. The dissertation’s organization documents her transition from an undergraduate foundation in physics to research at the intersection of computer vision, machine learning, and human perception, while her doctorate remained formally in electrical engineering. (thesis.caltech.edu)

Academic appointments and laboratory leadership

Li began her faculty career in August 2005 as an assistant professor in the Department of Electrical and Computer Engineering at the University of Illinois Urbana-Champaign, with an associated appointment in psychology. In January 2007 she joined Princeton University as an assistant professor of computer science, again holding an associated psychology appointment. She moved to Stanford University in June 2009 as an assistant professor of computer science, became an associate professor in August 2012, and advanced to professor in January 2018. These appointments placed her research on visual recognition within both engineering and computer-science departments while maintaining formal connections to the study of human perception. (cap.stanford.edu)

At Stanford, Li directed the Stanford Vision Lab, whose research combined computational approaches to seeing with cognitive and neuroimaging experiments concerning biological vision. This was a focused research group, distinct from the Stanford Artificial Intelligence Laboratory (SAIL), which she directed from 2013 to October 2018. SAIL is a broader, multi-faculty organization: its research groups encompass machine learning, natural language processing, robotics, and other areas alongside computer vision. Leading the vision laboratory therefore concerned a particular research community and agenda; directing SAIL involved a larger institutional setting supporting multiple faculty-led groups. Stanford’s subsequent research-group directory identifies the Stanford Vision and Learning Lab as a joint faculty laboratory involving Li, Juan Carlos Niebles, Silvio Savarese, and Jiajun Wu. (fsi.stanford.edu)

In June 2019 Li became Stanford’s inaugural Sequoia Professor of Computer Science. The named professorship was a faculty appointment, separate from the SAIL directorship that had ended the previous year. Her academic work also included classroom instruction: in the winter quarter of 2015, she and Andrej Karpathy were listed as the instructors for CS231n, Convolutional Neural Networks for Visual Recognition. Justin Johnson, Yuke Zhu, Brett Kuprel, and Ben Poole served as teaching assistants. The course treated deep learning methods for visual recognition, drawing background material from the ImageNet Challenge. Its assessment allocated 45 percent to three assignments, 15 percent to a midterm, and 40 percent to a final project, combining structured exercises with an extended student investigation. (cap.stanford.edu)

Graduate supervision constituted another role, distinct from teaching a scheduled course or administering a laboratory. Karpathy’s dissertation, Connecting Images and Natural Language, identifies Li as his primary adviser; his account of their work also credits both with designing CS231n and identifies him as its primary instructor. In a September 2016 account of doctoral research, Karpathy described a concrete practice in Li’s group: an internal deadline two weeks before a conference submission, requiring a draft of at least five pages and an internal review using conference-style review forms. This practice linked supervision to experimental planning and manuscript criticism, rather than merely to the organization of lectures or laboratory membership. (cs.stanford.edu)

Object recognition and the study of human vision

Li’s early computer vision research examined how a system could learn a new object category from very few examples despite variation in appearance and spatial arrangement. In the 2003 paper A Bayesian Approach to Unsupervised One-Shot Learning of Object Categories, written with Rob Fergus and Pietro Perona, the researchers tested four categories: human faces, airplanes, motorcycles, and spotted cats. They learned three categories from hundreds of images, extracted general statistical knowledge, and used it to learn the fourth from one to five examples. This form of transfer learning addressed a question motivated by human recognition: whether experience with familiar categories could reduce the evidence needed to recognize unfamiliar ones. (cs.stanford.edu)

Their mechanism combined Bayesian inference with a probabilistic “constellation” model of object parts. Probability distributions described parts’ appearances and relative positions, allowing variation rather than requiring an exact template. Previously learned categories supplied a prior distribution over model parameters; new examples updated it to a posterior using variational Bayesian estimation. Recognition incorporated uncertainty over parameters instead of relying on a single best-fitting estimate. Although the title called learning “unsupervised,” training images were identified as containing the target category; the system was not given the object’s location, size, or part correspondences. (cs.stanford.edu)

Li and Perona extended probabilistic learning to whole-scene recognition in their 2005 paper A Bayesian Hierarchical Model for Learning Natural Scene Categories. Here the task was to classify an image as, for example, a bedroom, forest, or highway, rather than identify an individual object within it. Their hierarchy represented a scene through local image patches, discrete visual “codewords,” and intermediate latent “themes.” Clustering patch descriptors produced the codebook; the model learned distributions of codewords within themes and mixtures of themes associated with scene categories. These intermediate representations did not require people to label individual patches as sky, foliage, or other semantic components, although scene-category labels were used for training. Experiments used grayscale images from 13 categories, with 100 images per category for training and the remainder for testing. A test image received the category whose model assigned it the highest likelihood. (people.csail.mit.edu)

Human perception was also an experimental subject in Li’s research, not simply an analogy for machine learning. Her 2005 doctoral dissertation, Visual Recognition: Computational Models and Human Psychophysics, combined computational category learning with psychophysics: controlled measurements of how people respond to visual stimuli. It brought together work on recognizing objects from limited examples, learning scene representations, and detecting meaningful categories when attention was occupied elsewhere. These investigated related but distinct questions—how representations can be learned and what information human observers can extract under constrained viewing conditions. (thesis.caltech.edu)

A specific example was the 2002 study Rapid Natural Scene Categorization in the Near Absence of Attention, conducted with Rufin VanRullen, Christof Koch, and Perona. Participants performed a demanding central letter-discrimination task while judging whether briefly presented peripheral photographs contained animals or vehicles. They could perform these categorization tasks concurrently, whereas control tasks involving discrimination between T and L shapes or differently arranged two-color disks were substantially impaired. The authors interpreted the contrast as evidence that some high-level visual categorization could proceed with very little focal attention—not that all visual recognition was attention-free. The experiments measured behavioral performance rather than establishing that human vision implemented the Bayesian algorithms developed in the computational work. (allpsych.uni-giessen.de)

ImageNet: dataset construction and visual recognition benchmarks

Fei-Fei Li’s work on ImageNet shifted the scale of resources available for visual object recognition. The project’s 2009 presentation at the IEEE Conference on Computer Vision and Pattern Recognition was a collaborative publication by Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. It described a database containing 3.2 million images across 5,247 categories in twelve hierarchical subtrees. Rather than introducing a single recognition algorithm, the researchers sought to build extensive, human-verified training data through which different computational methods could learn and be compared. (image-net.org)

The organizing structure came from WordNet, a lexical database whose “synsets,” or synonym sets, represent distinct concepts and connect them through semantic relationships, including more general and more specific categories. ImageNet associated photographs with these concepts. Its construction pipeline retrieved candidate images from internet search engines using expanded queries and multiple languages, then employed human annotators through Amazon Mechanical Turk to verify whether each image depicted the intended concept. Agreement requirements adapted to the category, rather than assuming that every concept was equally easy to recognize. This combined automated collection with human verification; search results alone did not constitute accepted annotations. (image-net.org)

The ImageNet Large Scale Visual Recognition Challenge (ILSVRC), launched in 2010, used selected categories and images rather than the entire database. Its 2012 classification subset contained 1,000 classes, 1,281,167 training images, 50,000 validation images, and 100,000 test images. Classification required assigning category labels; single-object localization additionally required a bounding box around an instance of the target category. Object detection, introduced in 2013, required finding and labeling instances across 200 categories. Standardized evaluation included top-five classification error, bounding-box overlap criteria for localization, and mean average precision for detection. Participants submitted predictions for a test set whose annotations were withheld, allowing organizers to compare systems under common conditions. (arxiv.org)

The 2012 competition supplied a particularly influential result for deep learning. Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton developed the AlexNet system, whose winning submission achieved a top-five test error of 15.3%, compared with 26.2% for the runner-up. Top-five error counted an image as incorrect when its reference label was absent from the system’s five highest-ranked predictions. Their convolutional neural network contained five convolutional and three fully connected layers; training a network took approximately five to six days on two NVIDIA GTX 580 graphics processing units. The result joined large-scale labeled data, network design, and GPU computation. Li and the ImageNet collaborators supplied the dataset and benchmarking infrastructure, while the Toronto researchers developed the winning network and training implementation. (papers.nips.cc)

Dataset construction itself subsequently became an object of investigation. In research released in December 2019 and published at FAT* in 2020, Kaiyu Yang, Klint Qinami, Li Fei-Fei, Jia Deng, and Olga Russakovsky examined ImageNet’s “person” subtree. They identified 1,593 of its 2,832 categories as potentially inappropriate labels and proposed their removal. Of the remaining 1,239 categories, only 158 met the study’s imageability threshold. This distinguished whether a label was acceptable from whether photographs could meaningfully establish it: “philanthropist,” for example, cannot reliably be inferred from appearance. The researchers also analyzed perceived gender, skin color, and age distributions and proposed image-subset balancing without publicly releasing sensitive image-level annotations. These proposals concerned category suitability and representation in the broader database; the paper reported that its filtering would leave the standard 1,000-class challenge subset unaffected. (arxiv.org)

Image descriptions, scene understanding, and embodied intelligence

Li’s collaborative research extended computer vision beyond assigning category labels to images toward representing objects, their attributes, and their relationships within scenes. A representative project was Visual Genome, introduced in 2016 and published in the International Journal of Computer Vision in 2017 by Ranjay Krishna, Yuke Zhu, Justin Johnson, Li, and other collaborators. Its 108,077 images received crowdsourced annotations connecting localized descriptions with objects, attributes, and relationships. These annotations supported scene graphs: structured representations in which nodes identify objects and edges describe relationships between them. Unlike a list of detected objects, such a representation distinguishes the entities in a scene from how they are spatially or semantically connected. Region-level annotations could also be combined into a graph representing the entire image. (link.springer.com)

With Andrej Karpathy, Li investigated the correspondence between image regions and natural-language expressions in Deep Visual-Semantic Alignments for Generating Image Descriptions, first posted in December 2014 and presented at CVPR in 2015. The system combined a convolutional neural network operating on image regions with a bidirectional recurrent neural network processing sentences. A shared embedding space brought corresponding visual and linguistic representations together, allowing the model to infer which regions supported particular words or expressions from training examples pairing images with descriptions. A separate multimodal recurrent network used the inferred alignments to generate new descriptions of regions or complete images. This was multimodal learning rather than classification into a fixed vocabulary of object categories: its outputs were sequences of words describing visual content. Evaluations included image–sentence retrieval on Flickr8K, Flickr30K, and MS COCO, as well as generated descriptions compared with retrieval-based baselines. (arxiv.org)

In the 2018 CVPR paper Image Generation from Scene Graphs, Justin Johnson, Agrim Gupta, and Li studied the reverse direction: synthesizing an image from an explicit specification of objects and relationships. Their model passed information along graph edges using graph convolution, then predicted a bounding box and segmentation mask for each object. These predictions formed a spatial layout that a cascaded refinement network converted into an image. Training used two adversarial discriminators to encourage realistic outputs. The scene graph therefore conditioned both the objects to be depicted and their relationships, rather than supplying only an unordered set of labels. Experiments on Visual Genome and COCO-Stuff examined multi-object image generation through qualitative comparisons, ablation experiments, and user studies; the outputs were two-dimensional images, not interactive environments. (openaccess.thecvf.com)

Li also collaborated on embodied intelligence, where agents learn through observations and actions within an environment. iGibson 1.0, described in a December 2020 preprint by Bokui Shen, Fei Xia, Chengshu Li, Roberto Martín-Martín, Li, Silvio Savarese, and other collaborators, provided 15 interactive home-sized scenes comprising 108 rooms. It combined simulated visual and depth sensors with movable and articulated objects, motion-planning tools, and interfaces for collecting human demonstrations. Randomizing object appearance and physical properties supported experiments on generalization. Unlike interpreting a static photograph, a robotic task could require selecting movements, encountering obstacles, and changing object configurations before receiving the next observation. The researchers used the environment to study navigation, manipulation, and imitation learning. (arxiv.org)

The collaborative iGibson 2.0 project, introduced in 2021, expanded the simulated tasks beyond movement and physical contact by representing object properties such as temperature, wetness, cleanliness, and sliced or switched-on states. Logical predicates translated simulator variables into task-level conditions such as “Cooked” or “Soaked”; a sampling mechanism generated physical configurations satisfying specified conditions. A virtual-reality interface enabled people to demonstrate household activities for imitation-learning experiments. These mechanisms made task success depend on changes in an environment’s state, rather than solely on recognizing an object or predicting its location in an image. (arxiv.org)

Ambient intelligence and AI-assisted healthcare research

Li and physician Arnold Milstein co-lead the Partnership in AI-Assisted Care at Stanford University, linking computer-science research with the Clinical Excellence Research Center. The collaboration applies computer vision to observing how care is delivered: sensors passively record clinical environments, and machine-learning systems identify activities involving patients and staff. Its research includes monitoring hand-hygiene dispenser use, recognizing patient mobilization in intensive care units, and characterizing older adults’ daily activities. These projects concern behavior and care processes rather than the interpretation of diagnostic scans. (med.stanford.edu)

In this work, ambient intelligence means sensing and computationally interpreting human activity within physical environments, including hospitals and daily living spaces. Li, Albert Haque, and Milstein outlined this research program in their September 2020 Nature review, “Illuminating the dark spaces of healthcare with ambient intelligence.” Unlike wearable monitoring alone, contactless sensing does not require a person to carry a device. Depth sensors represent distance from the sensor, producing silhouettes and spatial information rather than conventional color images; thermal imaging can obscure identifying visual details, although sensor choice also limits which activities can be distinguished. The researchers treated clinical validation and privacy protection as requirements for translation into large-scale care, not as consequences automatically established by successful activity recognition. (nature.com)

A 2020 study by Amit Singh, Haque, and colleagues, with Li and Milstein as co-last authors, evaluated automated hand-hygiene observation at Lucile Packard Children’s Hospital Stanford. Sixteen ceiling-mounted depth sensors collected data above dispensers outside patient rooms from March through August 2017. A convolutional neural network was trained on 111,080 annotated images; evaluation against simultaneous human observation used a separate set of 718 images. Human–algorithm agreement was 96.8%; against adjudicated annotations, algorithm sensitivity was 92.1% and specificity 98.3%. The measured task was detecting movements consistent with dispenser use, not assessing the completeness of hand cleaning. The study did not calculate overall compliance rates or demonstrate fewer hospital-acquired infections. Its authors described it as a single-unit proof of concept, with uncertain generalization to other settings. (academic.oup.com)

Research on daily living included “Vision-Based Gait Analysis for Senior Care,” a 2018 workshop paper by David Xue and collaborators, including Li and Milstein. Its pipeline identified walking segments in video, estimated skeletal joint positions, and analyzed changes in the distance between the feet to detect steps and calculate gait measurements. Testing used both the public UTKinect-Action3D dataset and recordings from a partner senior-care facility. The system counted steps correctly in 85% of UTKinect walking sequences; average foot-placement timing error on the senior-care recordings was 0.125 seconds. These were measurements of technical performance, not evidence that the system diagnosed neurological disease or improved patient outcomes. (arxiv.org)

Privacy, consent, and data governance were addressed explicitly in a collaborative paper led by Nicole Martinez-Martin, published online in December 2020 in The Lancet Digital Health, with Li and Milstein among its authors. The paper distinguished automated analysis during deployment from development, when human annotators may inspect recordings. It discussed restricted access, encryption, retention periods, secondary use, and residual reidentification risks. Consent discussions, the authors argued, must distinguish research recording from real-time clinical monitoring: stored sensor data may not be reviewed promptly or provide warnings of patient problems. Ethical review also depended on the project’s classification; the hand-hygiene study reported Stanford institutional review board assessment and designation as nonhuman-subjects research. (rcastoragev2.blob.core.windows.net)

Industry research and World Labs

From 2017 to 2018, Li served as a Google vice president and chief scientist of artificial intelligence and machine learning at Google Cloud. This was an industry leadership appointment undertaken during leave from Stanford University, rather than a university research position. Her responsibilities included developing AI capabilities for commercial users through cloud computing. On September 10, 2018, Google announced that she would return to her Stanford professorship and become an AI/ML adviser to Google Cloud, with Andrew Moore joining to lead Google Cloud AI at the end of that year. (profiles.stanford.edu)

A specific example of the product-development work during her tenure was Cloud AutoML. In a July 24, 2018 announcement, Li described the public beta of AutoML Vision and new AutoML offerings for natural language and translation. AutoML Vision allowed customers to adapt image-recognition capabilities to categories relevant to their own businesses, occupying a middle ground between using a fixed, pretrained service and constructing a custom model from scratch. These were Google Cloud team products, not individual research contributions attributable to Li alone; their stated purpose was to make customized machine learning accessible to organizations without extensive specialist expertise. (cloud.google.com)

In 2024, Li co-founded World Labs with Justin Johnson, Ben Mildenhall, and Christoph Lassner. The company’s founding announcement identified all four as co-founders, bringing together expertise in computer vision, machine learning, and computer graphics; Stanford’s biographical profile also identifies Li as its co-founder and chief executive. World Labs described its intended research direction as “large world models”: systems designed to perceive, generate, and interact with three-dimensional environments. Its founding program therefore concerned a company-wide research and engineering effort, rather than a continuation of Li’s university laboratory under a commercial name. (radical.vc)

The company uses spatial intelligence to describe the ability to model objects, places, and interactions across space and time. Its stated ambition extends beyond producing plausible pictures to representing environments in which viewpoints can change and actions have consequences. On December 2, 2024, World Labs presented “Generating Worlds,” describing early progress toward persistent, navigable three-dimensional scenes that users could explore in a browser. The emphasis on persistence and navigation distinguished the demonstration’s intended output from a single generated image: a scene was presented as a space to explore rather than merely a fixed composition to view. The announcement characterized this as early progress, not a completed system for general physical reasoning. (worldlabs.ai)

On November 12, 2025, World Labs made Marble generally available. According to its release documentation, this world model could generate environments from text, images, video, or coarse three-dimensional layouts, and supported editing, expansion, and composition of generated worlds. Outputs could be exported as Gaussian splats, meshes, or videos. Its experimental Chisel interface separated structural control from visual styling: users arranged simple shapes or imported assets to establish a layout, then supplied text to guide the generated appearance. These documented functions concerned creating and manipulating spatial content. The company’s broader proposed applications—including robotics, design, and visual-effects production—were ambitions and application directions, distinct from evidence that the released system could reliably simulate physical dynamics or control robots. (worldlabs.ai)

Human-centered AI institutions and public policy

Li developed the Stanford Institute for Human-Centered Artificial Intelligence (HAI) with John Etchemendy, a philosopher and former Stanford provost. In their March 2019 account of its origins, they traced the collaboration to a conversation in summer 2016 about bringing humanities and social-science perspectives into technological development. During 2017, Li involved computer scientists James Landay and Christopher Manning in developing the proposal and asked Etchemendy to become co-director after he stepped down as provost. A public initiative was announced in October 2018, and the institute formally launched at Stanford University on March 18, 2019. (hai.stanford.edu)

HAI began as a university-wide organization rather than a laboratory devoted to a single technical specialty. Its launch announcement listed 200 participating faculty members from Stanford’s then seven schools, bringing computing and engineering into collaboration with humanities, social sciences, medicine, law, business, and education. Its mechanisms included interdisciplinary research grants, fellowships, faculty recruitment, and meetings involving academic researchers, industry, government, and civil society. By the launch, it had supported approximately 55 interdisciplinary research teams, including projects concerning refugee resettlement, intensive-care delivery, and the governance and infrastructure implications of autonomous vehicles. These were activities through which HAI sought to advance its institutional mission, not evidence that AI had already produced the social benefits it proposed to pursue. (news.stanford.edu)

Li and Etchemendy articulated three principles for their approach to human-centered AI: studying and anticipating AI’s effects on people, developing applications that augment rather than replace human capabilities, and pursuing machine intelligence with greater human-like subtlety and versatility. Their proposal treated social inquiry as an input to technical development, rather than merely an assessment performed after deployment. Humanities and social sciences would help identify consequences; economics and education would inform responses; and law, policy, and ethics would contribute to governance. These principles expressed the founders’ research and institutional agenda, rather than a universally accepted definition of responsible AI. (hai.stanford.edu)

Li’s federal public-service work included membership in the National Artificial Intelligence Research Resource Task Force, announced on June 10, 2021, by the White House Office of Science and Technology Policy and the National Science Foundation. The task force was charged with preparing a roadmap to broaden access to AI research resources and educational tools. On September 14, 2023, she testified before the Senate Committee on Homeland Security and Governmental Affairs at its hearing, “Governing AI Through Acquisition and Procurement.” Her written testimony advocated evaluating government-purchased systems for accuracy, fairness, and reliability under real-world conditions; requiring vendors to disclose information about data collection, annotation, risks, and mitigation; and increasing public investment in AI research. The testimony explicitly stated that she spoke in an individual capacity, not on behalf of affiliated organizations. (hai.stanford.edu)

In 2023, Li was also named one of seven external scholars on United Nations Secretary-General António Guterres’s Scientific Advisory Board. The board’s mandate extended beyond AI: it was established to provide independent advice on scientific and technological developments and their relationships with ethics, governance, and sustainable development. Its structure combined external scholars with senior scientific representatives of UN organizations, the Secretary-General’s technology envoy, and the rector of the United Nations University, supported by a wider network of scientific institutions. Li’s appointment placed her within this science-policy advisory mechanism rather than in an intergovernmental negotiating or regulatory role. (india.un.org)

AI4ALL and access to AI education

Li’s work in AI education outreach began with the Stanford Artificial Intelligence Laboratory’s Outreach Summer program (SAILORS), established in 2015 with Olga Russakovsky and Rick Sommer, executive director of Stanford Pre-Collegiate Studies. The precursor to AI4ALL initially served high-school girls entering tenth grade. Its first cohort comprised 24 students selected from 213 applicants for a free, two-week, nonresidential program at Stanford University. Faculty and graduate-student volunteers introduced participants to research while providing contact with mentors and peers interested in artificial intelligence. (spcs.stanford.edu)

AI4ALL was formally established as a nonprofit in 2017, rather than in the year of the Stanford pilot. Its organizational history identifies Rab Govil as a founding board member and Melinda French Gates/Pivotal Ventures and Jensen and Lori Huang as early financial supporters. Tess Posner joined as founding chief executive, serving from 2017 to 2021. These roles distinguished the founders’ educational initiative, board governance, philanthropic support, and executive management. Li subsequently served as board chair; in an October 2021 announcement, she credited Posner with guiding the organization’s strategic direction and development. (ai-4-all.org)

The educational model combined technical instruction, supervised projects, mentoring, and discussion of AI’s social applications. At Stanford in 2016, mornings included lectures and tutorials, while other sessions offered mentoring, meetings with women working in engineering, and research activities. Student projects explored natural-language processing for disaster relief, computer vision for hospital safety, machine-learning approaches to genomic cancer data, and autonomous driving. These were teaching projects, not demonstrations of established medical or transportation capabilities. Students presented their work at the end of the program. This structure connected computational methods with concrete problems while making researchers and professional role models accessible to participants. (news.stanford.edu)

As the initiative expanded, its intended audience extended beyond the Stanford program’s initial focus on girls. AI4ALL’s history describes summer programs serving groups historically underrepresented in AI, including Black, Hispanic and Latinx, and Indigenous students, women, and high-school students with demonstrated financial need. A similar program opened at the University of California, Berkeley in 2017. The organization records growth from one university summer-program partnership in 2015 to sixteen in 2022; those figures describe that historical network, not a current partnership total. AI4ALL also launched Open Learning in 2019, offering free, adaptable curricula and resources for high-school teachers across subjects, and College Pathways in 2020, extending its work to undergraduates. (ai-4-all.org)

The initiative’s stated goals—widening participation, strengthening confidence, and developing future AI leaders—are distinct from demonstrated educational outcomes. A 2016 paper coauthored by Li and the Stanford program’s organizers reported statistically significant improvements in surveyed technical knowledge, career interest, and confidence after the pilot. Because this was an organizer-authored evaluation of a small, selected cohort, it was not an independent demonstration of long-term changes in participation across the field. AI4ALL’s 2019 annual report described broader evaluation through pre- and post-program questionnaires and alumni follow-up, tracking technical exposure, motivation, mentoring, subsequent study, and employment. These measures documented participant experiences and trajectories rather than establishing that the programs alone caused later educational or career choices. (cs.stanford.edu)

Publications, memoir, and professional recognition

Li publishes scholarly work under the name “L. Fei-Fei,” a convention explicitly identified in her Stanford University biography. Her publication record includes collaborative research papers, dataset and benchmark descriptions, and interdisciplinary reviews. Early examples include “A Bayesian Approach to Unsupervised One-Shot Learning of Object Categories,” written with Rob Fergus and Pietro Perona and presented at ICCV in 2003, and “A Bayesian Hierarchical Model for Learning Natural Scene Categories,” written with Perona for CVPR in 2005. These publications represent distinct strands of her early computer vision research: learning object categories from limited examples and modeling the categories of whole scenes. (profiles.stanford.edu)

A selective bibliography also distinguishes research infrastructure from recognition and generation methods. “ImageNet: A Large-Scale Hierarchical Image Database” (CVPR, 2009), by Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li, documented the collaborative construction of ImageNet. “Deep Visual-Semantic Alignments for Generating Image Descriptions” (CVPR, 2015), with Andrej Karpathy, connected image regions with language and used those correspondences to generate descriptions. “Image Generation from Scene Graphs” (CVPR, 2018), with Justin Johnson and Agrim Gupta, instead investigated generating images from structured representations of objects and their relationships. These papers respectively concern a shared data resource, multimodal learning, and conditional image generation, rather than interchangeable versions of a single contribution. (cs.stanford.edu)

Her memoir, The Worlds I See: Curiosity, Exploration, and Discovery at the Dawn of AI, was published on November 7, 2023, by Flatiron Books under its A Moment of Lift Books imprint. The 336-page book combines her account of immigration, family responsibilities, and scientific education with an explanation of developments in artificial intelligence. Its opening chapter places her preparations for congressional testimony alongside recollections of her immigrant childhood and responsibilities toward her parents. Those scenes convey Li’s retrospective account of her experiences and motivations; they are autobiographical testimony, distinct from the publication dates, institutional appointments, and other chronology documented in external records. The book is therefore both a personal narrative and a participant’s account of AI research, not a technical monograph or an independently authored history of the field. (us.macmillan.com)

Li was elected to the National Academy of Engineering in 2020 for her contributions to large knowledge bases supporting machine learning and visual understanding. Also in 2020, she was elected to the National Academy of Medicine, whose citation recognized her work in vision-based AI, including research on recognizing clinician and patient behavior. In 2021, she was elected to the American Academy of Arts and Sciences. These were separate elections by three institutions, rather than successive stages of one award. (nae.edu)

The 2025 Queen Elizabeth Prize for Engineering recognized modern machine learning and was shared by seven recipients: Yoshua Bengio, Geoffrey Hinton, John Hopfield, Yann LeCun, Jensen Huang, Bill Dally, and Li. The prize organization distinguished their complementary contributions: neural-network foundations and methods, computing hardware, and datasets for training and evaluation. Its account specifically associated Li’s recognition with high-quality datasets and ImageNet. The award thus honored an interconnected set of engineering contributions; it was neither an individual prize to Li alone nor an attribution of all modern machine-learning advances to any one recipient. (qeprize.org)

References

1. Li Fei-Fei: Biography; Vision; Princetoncs.stanford.edu
2. The Godmother of AIalumni.princeton.edu
3. At Alumni Day, Princeton honors computer scientist Fei-Fei Li ’99 and ornithologist John Fitzpatrick *78alumni.princeton.edu
4. Li Fei-Fei: Resume; Vision; Princetoncs.stanford.edu
5. Fei-Fei Li's Profile | Stanford Profilesprofiles.stanford.edu
6. Visual Recognition: Computational Models and Human Psychophysics — CaltechTHESISthesis.caltech.edu
7. Fei-Fei Li: Curriculum Vitaecap.stanford.edu
8. Fei-Fei Li | FSIfsi.stanford.edu
9. About – Stanford Artificial Intelligence Laboratoryai.stanford.edu
10. Research Groups – Stanford Artificial Intelligence Laboratoryai.stanford.edu
11. Stanford University CS231n: Convolutional Neural Networks for Visual Recognitioncs231n.stanford.edu
12. Connecting Images and Natural Languagecs.stanford.edu
13. Andrej Karpathy Academic Websiteai.stanford.edu
14. A Survival Guide to a PhDkarpathy.github.io
15. A Bayesian Approach to Unsupervised One-Shot Learning of Object Categoriescs.stanford.edu
16. Publicationsrobots.ox.ac.uk
17. A Bayesian Hierarchical Model for Learning Natural Scene Categoriescs.stanford.edu
18. Rapid natural scene categorization in the near absence of attentionallpsych.uni-giessen.de
19. ImageNet: A Large-Scale Hierarchical Image Databaseimage-net.org
20. ImageNet Large Scale Visual Recognition Challengearxiv.org
21. ImageNet Classification with Deep Convolutional Neural Networkspapers.nips.cc
22. Towards Fairer Datasets: Filtering and Balancing the Distribution of the People Subtree in the ImageNet Hierarchyarxiv.org
23. ImageNetimage-net.org
24. Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotationslink.springer.com
25. Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotationsarxiv.org
26. Deep Visual-Semantic Alignments for Generating Image Descriptionsarxiv.org
27. Deep Visual-Semantic Alignments for Generating Image Descriptionscv-foundation.org
28. Image Generation From Scene Graphsopenaccess.thecvf.com
29. Image Generation From Scene Graphsopenaccess.thecvf.com
30. iGibson 1.0: a Simulation Environment for Interactive Tasks in Large Realistic Scenesarxiv.org
31. iGibson 2.0: Object-Centric Simulation for Robot Learning of Everyday Household Tasksarxiv.org
32. Computer Visionmed.stanford.edu
33. Illuminating the dark spaces of healthcare with ambient intelligencenature.com
34. Vision-Based Gait Analysis for Senior Carearxiv.org
35. Ethical issues in using ambient intelligence in health-care settingsrcastoragev2.blob.core.windows.net
36. Ethical issues in using ambient intelligence in health-care settings — PubMedpubmed.ncbi.nlm.nih.gov
37. Google Cloud AI: Andrew Moore joining Google Cloud; Fei-Fei Li becoming advisorcloud.google.com
38. Empowering businesses and developers to do more with AIcloud.google.com
39. Introducing World Labsradical.vc
40. About | World Labsworldlabs.ai
41. Generating Worlds | World Labsworldlabs.ai
42. Marble: A Multimodal World Model | World Labsworldlabs.ai
43. Opening the Gatehai.stanford.edu
44. Introducing Stanford's Human-Centered AI Initiativehai.stanford.edu
45. HAI Co-Director Fei-Fei Li named to National Artificial Intelligence Research Resource Task Forcehai.stanford.edu
46. Governing AI Through Acquisition and Procurementhsgac.senate.gov
47. Governing AI Through Acquisition and Procurementhai.stanford.edu
48. Fei Fei Li's Testimony before the Senate Committee on Homeland Security and Governmental Affairshai.stanford.edu
49. UN Secretary-General Creates Scientific Advisory Board for Independent Advice on Breakthroughs in Science and Technologyindia.un.org
50. Artificial intelligence camp bridges STEM gender gapspcs.stanford.edu
51. Towards More Gender Diversity in CS through an Artificial Intelligence Summer Program for High School Girlscs.stanford.edu
52. Our Story - AI4ALLai-4-all.org
53. Tess Posner - AI4ALLai-4-all.org
54. Singing A New Song - AI4ALLai-4-all.org
55. 2019 ANNUAL REPORTai-4-all.org
56. Stanford Computer Vision Lab: Publicationscs.stanford.edu
57. ImageNet: A Large-Scale Hierarchical Image Databasecs.stanford.edu
58. Deep Visual-Semantic Alignments for Generating Image Descriptionscs.stanford.edu
59. Image Generation from Scene Graphsarxiv.org
60. National Academy of Engineering: Class of 2020 Membersnae.edu
61. Five professors elected to National Academy of Medicinecontent.medweb.stanford.edu
62. Modern Machine Learning | Queen Elizabeth Prize for Engineeringqeprize.org