Fei-Fei Li

From physics to the study of seeing

Li was born in Beijing and grew up in Chengdu before moving to the United States as a teenager. At Princeton, she studied physics and graduated in 1999. Returning to the university in 2024, she spoke with incoming students about a life that had included helping in her family’s dry-cleaning business while pursuing her studies. Her memoir, The Worlds I See, brings that personal history into conversation with her scientific work. Li’s biographical account · Princeton conversation, September 2024.

Her doctoral work at Caltech brought her to electrical engineering and visual recognition. She completed her PhD in 2005, working with Pietro Perona and Christof Koch. After appointments at Illinois and Princeton, she joined Stanford in 2009. She later directed the Stanford Artificial Intelligence Laboratory from 2013 to 2018 and worked at Google Cloud during a sabbatical in 2017–2018. These settings placed her research in contact with both academic questions and the practical demands of large computing organisations. Doctoral background · Stanford profile.

Making data into a scientific resource

ImageNet addressed a demanding problem: how to assemble enough carefully organised examples for computers to learn visual categories. Its 2009 paper described a database structured around WordNet concepts, with human annotation helping to check the images. The paper was written by Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li and Fei-Fei Li. The contribution was a shared resource and a method for building it, involving researchers and many people who supplied judgements about images. ImageNet paper, CVPR 2009 (PDF).

The ImageNet Large Scale Visual Recognition Challenge began in 2010, giving research groups a common framework for comparison. Its later overview documents several tasks and the collective work behind their evaluation. A benchmark makes competing methods easier to examine, but its categories and tests also define what success means within that experiment. Challenge overview.

The celebrated 2012 neural-network result associated with ImageNet came from Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton. Their paper and the database paper describe different contributions. Li helped establish the data infrastructure on which that result was assessed; she was not an author of the network now widely known as AlexNet. The 2012 neural-network paper.

Looking critically at the labels

Large datasets also preserve choices about how people and objects are described. In work released in 2019 and published for the 2020 fairness conference, Li joined Kaiyu Yang, Klint Qinami, Jia Deng and Olga Russakovsky in examining the person categories of the broader ImageNet hierarchy. They studied problematic vocabulary, concepts that cannot reliably be inferred from appearance, and uneven representation. They proposed filtering and balancing measures, while discussing limitations. This investigation concerned the person subtree, not simply the thousand categories of the familiar recognition challenge. Towards Fairer Datasets, 2019–2020.

That work gives a more demanding account of progress than a rising recognition score alone. The questions include whether a category is appropriate, whether an image supports the label attached to it, and who is represented. Improving a dataset is a continuing research problem; the proposed measures do not establish that every form of bias has disappeared.

Opening the field to more people

Educational work developed alongside the research. In 2015, Li, Olga Russakovsky and Rick Sommer established a Stanford summer programme introducing high-school girls to artificial intelligence. AI4ALL was founded as a nonprofit two years later, with Tess Posner as its founding chief executive. The organisation turned access to learning and mentorship into a continuing institutional project. AI4ALL’s history.

Li’s role has changed over time. In November 2024, AI4ALL announced that Russakovsky would succeed her as board chair. At Stanford, the May 2026 restructuring of HAI named James Landay as director, with Li serving as a senior adviser on AI to the university president and co-chairing the institute’s advisory council with John Hennessy. Her status as a founding director should therefore be distinguished from its current operational leadership. AI4ALL leadership announcement · Stanford’s May 2026 announcement.

Beyond the photograph

World Labs, founded in 2024, extends this trajectory towards spatial intelligence. Li co-founded the company with Justin Johnson, Ben Mildenhall and Christoph Lassner. Its research concerns representations of environments that can be explored from different viewpoints, rather than only recognised in a single image. World Labs and its founders.

The company presented Atlas in September 2026 and listed its next-generation model release on 12 November. It describes tools for generating and reconstructing environments from inputs including images and video. Those are the developer’s accounts of its system, not an independent finding of general physical understanding. Its technical explanation also acknowledges that unseen parts of a scene may be imagined: a plausible new view is not necessarily a faithful record of an actual place. Atlas technical introduction · Release chronology.

A career in seven moments

1. 1999 · PhysicsGraduates from Princeton.
2. 2005 · Visual recognitionCompletes her doctorate at Caltech.
3. 2009 · A shared image resourceThe ImageNet paper appears; Li joins Stanford.
4. 2017 · Access to AIAI4ALL is established as a nonprofit.
5. 2019 · Institutions and scrutinyStanford HAI begins; research examining ImageNet’s person categories is released.
6. 2024 · Spatial intelligenceWorld Labs is founded by Li and three research colleagues.
7. 2026 · A changing landscapeStanford announces new HAI leadership; World Labs presents Atlas and agrees to a proposed acquisition by AMD.

Li’s career shows how research depends on more than a model: it needs examples, categories, institutions and people able to question their assumptions. Her continuing move from image recognition towards spatial representation makes that history relevant to a new generation of systems. The central challenge remains recognisable: deciding what a machine has learned about the world, and what evidence supports that judgement.