Rob Fergus
Last reviewed
Jun 5, 2026
Sources
21 citations
Review status
Source-backed
Revision
v2 ยท 2,812 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 5, 2026
Sources
21 citations
Review status
Source-backed
Revision
v2 ยท 2,812 words
Add missing citations, update stale details, or suggest a clearer explanation.
Rob Fergus is a British computer scientist known for his work in computer vision and deep learning, and a professor of computer science at New York University. [1][2] He is widely recognized for the 2014 paper "Visualizing and Understanding Convolutional Networks," co-authored with his doctoral student Matthew Zeiler, which introduced influential techniques for inspecting the internal representations of convolutional neural networks and produced the architecture commonly called ZFNet. [3][4] Fergus has a long association with Meta's Fundamental AI Research lab, which he co-founded, and in 2025 he returned from Google DeepMind to become head of the lab. [5][6]
Fergus works primarily on machine learning, deep learning, representation learning, and generative models, with much of his research grounded in computer vision. [2] He is a professor of computer science at the Courant Institute of Mathematical Sciences at New York University, where he has taught since the late 2000s, and he co-founded the university's CILVR lab with Yann LeCun in 2009. [2] Alongside his academic career he has held research roles in industry, first as a research scientist at what was then Facebook AI Research, later as a research director at Google DeepMind, and from 2025 as the head of Meta's research lab. [5][6] His Google Scholar profile lists more than 160,000 citations and an h-index above 90. [7]
Fergus is originally from the United Kingdom. [1] He completed his undergraduate degree in electrical engineering at the University of Cambridge, then earned a master's degree in electrical engineering at the California Institute of Technology, where he worked with Pietro Perona. [1] He received his doctorate from the University of Oxford under the supervision of Andrew Zisserman. [1] His thesis was recognized in 2005 with the prize for the best computer science doctoral thesis in the United Kingdom. [1] After completing his doctorate he spent two years as a postdoctoral researcher in the Computer Science and Artificial Intelligence Laboratory at the Massachusetts Institute of Technology, working with William T. Freeman. [1]
Fergus studied for his bachelor's and master of engineering degree in electrical and information engineering at Pembroke College, Cambridge, from 1996 to 2000. [11] He then moved to the United States, where he completed a Master of Science in electrical engineering at Caltech between 2000 and 2002 with Perona as his advisor, before returning to the United Kingdom for his doctorate. [11] He carried out his doctoral research in the Visual Geometry Group of Oxford's robotics research group from 2002 to 2005, and the resulting D.Phil. in electrical engineering, titled "Visual Object Category Recognition," was completed in October 2005. [11] He was a postdoctoral research associate at MIT CSAIL from 2005 to 2007. [11] His doctoral thesis was named the best computer science Ph.D. thesis in the United Kingdom by the British Computer Society and the best computer vision Ph.D. thesis by the British Machine Vision Association, both in 2006. [11]
Fergus joined New York University as an assistant professor of computer science at the Courant Institute of Mathematical Sciences and was subsequently promoted to full professor. [1][2] In 2009 he and Yann LeCun founded the Computational Intelligence, Learning, Vision, and Robotics (CILVR) lab at the university. [2] He teaches and supervises research in computer vision and machine learning, and several of his doctoral students went on to prominent positions in the field, including Matthew Zeiler, who founded the company Clarifai, and Wojciech Zaremba, a co-founder of OpenAI. [2]
Fergus took up his post at the Courant Institute in 2007, joining its Vision, Learning and Graphics group. [11][12] He is also a member of the NYU Center for Data Science, where his listed research areas include computer vision, large-scale object recognition, deep learning, statistical methods in astronomy, and computational photography. [12] Beyond Zeiler and Zaremba, his former students and advisees include Denis Yarats, later a co-founder and chief technology officer of Perplexity AI, and Alexander Rives, who led the Evolutionary Scale Modeling (ESM) protein language model effort at Facebook AI Research and went on to co-found EvolutionaryScale. [2][13] His teaching at NYU has included graduate courses on computer vision and on deep learning. [2]
Before his deep learning work, Fergus was known for probabilistic approaches to object category recognition. With his doctoral advisors Pietro Perona and Andrew Zisserman he developed the constellation model, presented in "Object Class Recognition by Unsupervised Scale-Invariant Learning" at the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) in 2003. [14] The method represents an object as a flexible constellation of parts, modeling the appearance, shape, relative scale, and possible occlusion of each part, and learns these categories from cluttered images without manual annotation using an expectation-maximization procedure. [14] The paper received the CVPR best paper prize, awarded out of roughly one thousand submissions. [11][14]
During the same period Fergus worked with Fei-Fei Li and Perona on learning visual categories from very few examples. Their paper "Learning Generative Visual Models from Few Training Examples: An Incremental Bayesian Approach Tested on 101 Object Categories" (2004) was an early demonstration of one-shot and few-shot learning of object categories, building on the Caltech 101 image dataset and extended in the journal article "One-Shot Learning of Object Categories" (2006). [15] He also contributed to large-scale image retrieval: "Spectral Hashing," with Yair Weiss and Antonio Torralba (NeurIPS 2008), showed how to learn compact binary codes for images by relating the problem to graph partitioning and thresholding eigenvectors of a graph Laplacian, an influential early result in learning-to-hash for nearest-neighbor search. [16]
Fergus's best-known contribution is the paper "Visualizing and Understanding Convolutional Networks," written with Matthew Zeiler and first posted to arXiv in November 2013, then published at the European Conference on Computer Vision (ECCV) in 2014. [3][4] The paper addressed why convolutional networks performed so well on large-scale image recognition and how they might be improved. It introduced a visualization technique, built on a deconvolutional network, that projects the activations of intermediate feature layers back into pixel space, giving insight into what individual units in the network respond to. [3] Using these visualizations as a diagnostic tool, the authors refined the architecture of the ImageNet-winning network of Krizhevsky and colleagues, known as AlexNet, and produced an improved model that is now commonly referred to as ZFNet, after the authors' initials. [3][4] The work is among the most cited papers in computer vision, with more than 26,000 citations recorded on Google Scholar. [7]
The visualization method in the ZFNet paper grew out of earlier research by Fergus and Zeiler on deconvolutional networks. In "Deconvolutional Networks," presented at CVPR in 2010 with Dilip Krishnan and Graham Taylor, they proposed a framework for learning image features by decomposing an image into a set of feature maps and learned filters. [8] A follow-up, "Adaptive Deconvolutional Networks for Mid and High Level Feature Learning," appeared at the International Conference on Computer Vision in 2011. [9]
Fergus has contributed to a range of widely used results in vision and learning. He was a co-author of "Intriguing Properties of Neural Networks" (2013), the study that drew broad attention to adversarial examples in deep networks, and of "OverFeat," an early integrated framework for recognition, localization, and detection using convolutional networks. [7] His group released the NYU Depth dataset and the associated work "Indoor Segmentation and Support Inference from RGBD Images" (2012), a benchmark for indoor semantic segmentation and scene understanding. [7] He also co-authored "Deep Generative Image Models Using a Laplacian Pyramid of Adversarial Networks" (2015), an early approach to image synthesis built on generative adversarial networks. [7] More recently his stated research interests have expanded to include protein design. [7]
"Intriguing Properties of Neural Networks," led by Christian Szegedy with Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Fergus, is one of his most cited works, with more than 22,000 citations; it documented that imperceptible, deliberately chosen perturbations could cause networks to misclassify inputs, helping to launch the study of adversarial examples. [7] Another widely used result is "Learning Spatiotemporal Features with 3D Convolutional Networks" (2015), known as C3D, with Du Tran, Lubomir Bourdev, Lorenzo Torresani, and Manohar Paluri, which popularized three-dimensional convolutions for video understanding and has accumulated more than 12,000 citations. [7] His expansion into computational biology is reflected in "Biological Structure and Function Emerge from Scaling Unsupervised Learning to 250 Million Protein Sequences," co-authored with Alexander Rives and colleagues, which trained transformer language models on protein sequences and helped establish protein language modeling as a research direction. [13]
The following table lists selected papers, several of which have received test-of-time recognition.
| Year | Paper | Co-authors (selected) | Venue |
|---|---|---|---|
| 2003 | Object Class Recognition by Unsupervised Scale-Invariant Learning | P. Perona, A. Zisserman | CVPR (Best Paper) |
| 2004 | Learning Generative Visual Models from Few Training Examples | L. Fei-Fei, P. Perona | CVPR Workshop |
| 2008 | Spectral Hashing | Y. Weiss, A. Torralba | NeurIPS |
| 2010 | Deconvolutional Networks | M. Zeiler, D. Krishnan, G. Taylor | CVPR |
| 2011 | Adaptive Deconvolutional Networks for Mid and High Level Feature Learning | M. Zeiler, G. Taylor | ICCV |
| 2012 | Indoor Segmentation and Support Inference from RGBD Images | N. Silberman, D. Hoiem, P. Kohli | ECCV |
| 2013 | Intriguing Properties of Neural Networks | C. Szegedy et al. | arXiv / ICLR |
| 2013 | OverFeat: Integrated Recognition, Localization and Detection | P. Sermanet et al. | arXiv / ICLR |
| 2014 | Visualizing and Understanding Convolutional Networks | M. Zeiler | ECCV |
| 2015 | Learning Spatiotemporal Features with 3D Convolutional Networks (C3D) | D. Tran, L. Bourdev, L. Torresani, M. Paluri | ICCV |
| 2015 | Deep Generative Image Models Using a Laplacian Pyramid of Adversarial Networks | E. Denton, S. Chintala, A. Szlam | NeurIPS |
Fergus co-founded Facebook AI Research, the lab now known as Meta AI, with Yann LeCun in 2013, joining as a research scientist while continuing his academic work. [2][5] He later left to spend roughly five years as a research director at Google DeepMind. [5][6] His career has therefore moved repeatedly between academia and industry, and between Meta and DeepMind, two of the leading organizations in deep learning research.
Fergus worked as a research scientist at Facebook AI Research from around 2014, contributing to the lab during its early years alongside LeCun and others. [5][17] In June 2020 he left Facebook to join Google DeepMind, where he helped build up the company's research team in New York and served as a research director for about five years. [5][17] During this period DeepMind expanded its New York presence, drawing on the city's academic machine learning community. [17]
In May 2025 Meta named Fergus the new head of its Fundamental AI Research lab, returning him to the organization he had helped start. [5][6] He took over from Joelle Pineau, who had announced her departure in April 2025. [5] Announcing the appointment, Yann LeCun, Meta's chief AI scientist, wrote that "Rob Fergus is the new head of Meta-FAIR" and that the lab was "refocusing on Advanced Machine Intelligence: what others would call human-level AI or AGI." [10] Press coverage described his return as part of an effort to strengthen FAIR's research leadership during a period in which the lab had seen significant staff turnover and competition from Meta's product-focused generative AI group. [5][6]
The announcement of Fergus's appointment was reported by outlets including Bloomberg and TechCrunch on May 8, 2025. [5][18] Coverage noted that FAIR had been responsible for early Meta foundation models such as Llama and Llama 2, and that the lab's research spans robotics, audio generation, image understanding, and longer-horizon work aimed at advancing the capabilities of AI systems. [5][6] Several reports framed the move as Meta bringing back a co-founder of the lab to refocus its fundamental research after a stretch of departures to startups and to Meta's own generative AI organization. [5][18]
Soon after, in June 2025, Meta reorganized its artificial intelligence efforts into a new division called Meta Superintelligence Labs (MSL), established through an internal memo from chief executive Mark Zuckerberg. [19] Within that structure FAIR, under Fergus, became one of several teams, sitting alongside a model-training group sometimes referred to as TBD Lab led by chief AI officer Alexandr Wang, a products and applied research group led by former GitHub chief executive Nat Friedman, and an infrastructure team. [19] FAIR retained its identity as the company's longer-term fundamental research lab focused on advanced machine intelligence. [19]
As of 2026 Fergus is vice president of AI research and head of FAIR at Meta, in addition to his professorship at New York University. [5][6] In that role he leads the lab's longer-horizon research agenda, oriented toward artificial general intelligence, or what LeCun and Fergus describe as advanced machine intelligence, spanning areas such as machine reasoning, robotics, perception, and generative modeling. [6][10] FAIR is also the group historically associated with Meta's open foundation model work, including the early Llama models. [6]
A central theme of FAIR's agenda under Fergus and LeCun is the development of world models, AI systems intended to capture the physical dynamics of the real world and to predict how objects and interactions unfold over time. [19][20] This line of work is associated with Meta's Joint Embedding Predictive Architecture, and in June 2025 the lab released V-JEPA 2, a video-based world model trained to learn from raw video, which Meta presented as a step toward agents that can plan and reason about the physical world. [20] Fergus has continued to publish, with recent work on world model learning, agentic reasoning benchmarks, and multimodal pretraining appearing at venues such as the International Conference on Learning Representations. [21]
Fergus has received several honors over his career. His doctoral co-authors and he were awarded the CVPR Best Paper Prize in 2003, and his doctoral thesis won the prize for the best UK computer science thesis in 2005. [1] He has been named a Sloan Research Fellow and received the United States National Science Foundation's CAREER award. [2] Several of his papers have received test-of-time awards at major venues, including ECCV, CVPR, and the International Conference on Learning Representations. [2]
In addition to the 2003 CVPR best paper prize for the constellation model work, Fergus's earlier honors include the best short course prize at the International Conference on Computer Vision in 2005 and a UK graduate research fellowship that funded his doctorate. [11] As of 2026 his Google Scholar profile records more than 165,000 citations, an h-index of 93, and an i10-index above 150, reflecting the broad influence of his work across computer vision and deep learning. [7]