Hervé Jegou

Hervé Jegou is a French computer scientist whose research spans large-scale similarity search, vector quantization, image retrieval, and self-supervised vision transformers. He is the originator of the FAISS similarity-search library and a co-author on the Product Quantization paper (2011), the DeiT, DINO, and DINOv2 vision-transformer papers, and the FAISS library paper (2024). As of May 2026, he is a Senior Research Director at Meta AI and a co-founder of Kyutai, the Paris-based nonprofit AI research lab he helped launch in November 2023 alongside Edouard Grave, Patrick Pérez, and three other researchers.

At a glance

• Education: Diploma from the Computer Science and Telecommunications department of ENS Cachan-Bretagne (now ENS Rennes), 2001; MS in computer science, Université de Rennes 1, 2002; PhD in computer science, Université de Rennes 1, 2005, advised by Christine Guillemot at Inria.
• Current roles: Senior Research Director at Meta AI FAIR (since 2025); co-founder of Kyutai (November 2023 to 2024).
• Key contributions: the Product Quantization paper (PAMI 2011); the FAISS similarity-search library; the DeiT (2020), DINO (2021), and DINOv2 (2023) vision-transformer papers; co-author on Moshi (2024).
• X / Twitter: @hjegou
• GitHub: jegou
• Google Scholar: 1lcY2z4AAAAJ
• Personal site: people.rennes.inria.fr/Herve.Jegou

Origins

Jegou was born and raised in France. He entered the Computer Science and Telecommunications department of ENS Cachan-Bretagne (the Brittany branch of École Normale Supérieure de Cachan, which became ENS Rennes in 2014) and graduated in 2001. He then completed a Master of Science in computer science at Université de Rennes 1 in 2002.

He carried out his doctoral research at Inria Rennes under Christine Guillemot on robust source coding for multimedia transmission over noisy mobile channels. He defended in 2005 with a thesis titled "Codes robustes et codes joints source-canal pour transmission multimédia sur canaux mobiles," awarded the national prize for best PhD thesis in image and signal processing by the Club EEA (the French electrical-engineering university association). The early line on robust source coding laid the technical foundation for the large-scale indexing and quantization work that followed.

Career

After his 2005 doctorate, Jegou was hired as a permanent researcher (chargé de recherche) at Inria in 2006. His first three years were with the LEAR project-team at Inria Grenoble Rhône-Alpes, working with Cordelia Schmid and Patrick Pérez on image-search and computer-vision indexing. The Inria Grenoble period produced two of his most-cited early papers: Hamming embedding and weak geometric consistency for large scale image search at ECCV 2008, and the foundational Product Quantization for Nearest Neighbor Search paper at IEEE PAMI 2011 with Matthijs Douze and Schmid.

In 2009, Jegou moved to Inria Rennes Bretagne-Atlantique to join the TEXMEX project-team, then later the LinkMedia team, where he led the large-scale-indexing research line. The Rennes period included the 2014 ERC Starting Grant for the VIAMASS project on automatic discovery of visual links between images at very large scale, and the 2010 Prix Bretagne Jeune Chercheur. The same period produced Aggregating local image descriptors into compact codes at IEEE PAMI 2012 with Douze, Schmid, and Pérez (over 6,100 citations as of mid-2026).

In May 2015, Jegou left Inria to join Facebook AI Research (FAIR) Paris as a research scientist. The FAIR Paris office had been established the year before, and Jegou and Matthijs Douze shipped the first public version of the FAISS similarity-search library in March 2017, formalizing the product-quantization line into a production-grade open-source library used across the AI industry for billion-scale vector search. Jegou rose through FAIR's research-scientist track to become a Director of Research at FAIR Paris, with leadership responsibility for the lab's vision and indexing research lines.

The 2020-to-2023 FAIR period produced the vision-transformer line that has shaped most of his recent citation impact. The DeiT paper (Touvron, Cord, Douze, Massa, Sablayrolles, Jegou, December 2020) introduced data-efficient image transformers and demonstrated that vision transformers could be trained competitively on ImageNet without proprietary JFT-300M data. The DINO paper (Caron, Touvron, Misra, Jegou, Mairal, Bojanowski, Joulin, ICCV 2021) introduced self-distillation with no labels and showed that self-supervised vision transformers learned semantic-segmentation-like features without supervision. The DINOv2 paper (Oquab et al., April 2023) extended the line to larger models and richer pretraining data, with Jegou among the 24-author byline. The three papers carry over 22,800 combined Google Scholar citations as of mid-2026.

In November 2023, Jegou co-founded Kyutai, the French nonprofit AI research lab launched in Paris by Iliad, CMA CGM, and Schmidt Futures with a €300 million ($330 million) funding commitment. He was one of six founding researchers, alongside Patrick Pérez (CEO), Edouard Grave, Laurent Mazaré, Neil Zeghidour, and Alexandre Défossez, and served as Chief Science Officer through 2024. He is a co-author on the Moshi full-duplex spoken-dialogue paper released by Kyutai in 2024.

In 2025, Jegou returned to Meta AI FAIR as a Senior Research Director, the role he holds as of May 2026, while retaining his Kyutai co-founder status (listed in the Kyutai team Alumni section). The same year, he co-authored The FAISS library with Douze and seven other Meta researchers, the first comprehensive technical paper on the FAISS codebase eight years after the library's initial release.

Affiliations

• ENS Cachan-Bretagne (now ENS Rennes): Student, Computer Science and Telecommunications department, graduated 2001.
• Université de Rennes 1 and Inria: PhD candidate in computer science, 2002 to 2005, advised by Christine Guillemot.
• Inria Grenoble Rhône-Alpes (LEAR project-team): Permanent researcher (chargé de recherche), 2006 to 2009.
• Inria Rennes Bretagne-Atlantique (TEXMEX, then LinkMedia teams): Permanent researcher (chargé de recherche), 2009 to 2015.
• Meta AI (FAIR Paris): Research Scientist, 2015 to 2020; Director of Research, 2020 to 2023.
• Kyutai: Co-founder and Chief Science Officer, November 2023 to 2024; Co-founder (alumni), 2024 to present.
• Meta AI (FAIR Paris): Senior Research Director, 2025 to present.

Notable contributions

Jegou's published record runs from robust source-coding work at Inria, through the product-quantization and image-retrieval line, into the FAISS library and the vision-transformer line at FAIR, and on to the Moshi-paper byline at Kyutai. His Google Scholar profile lists approximately 86,000 citations and an h-index of 77 as of mid-2026.

• Product Quantization for Nearest Neighbor Search (IEEE PAMI 2011). With Matthijs Douze and Cordelia Schmid; introduced product quantization as a vector-compression technique enabling nearest-neighbor search over billion-scale collections, with over 5,400 citations and a foundational role in the vector-search line.
• FAISS (Facebook AI Research, March 2017). Open-source similarity-search and clustering library originated by Jegou and Douze; the de facto industry standard for billion-scale vector search and the technical underpinning for many production retrieval systems.
• Hamming Embedding and Weak Geometric Consistency for Large Scale Image Search (ECCV 2008). With Douze and Schmid; an early high-impact image-retrieval paper introducing Hamming-embedded inverted-file indexing.
• Aggregating local image descriptors into compact codes (IEEE PAMI 2012). With Perronnin, Douze, Schmid, and Pérez; over 6,100 citations on the VLAD compact-image-representation line.
• Billion-scale similarity search with GPUs (Johnson, Douze, Jegou, 2017 / IEEE Big Data 2021). Over 6,800 citations on GPU-accelerated FAISS.
• Training data-efficient image transformers and distillation through attention (DeiT, ICML 2021). With Touvron, Cord, Douze, Massa, and Sablayrolles; over 11,700 citations.
• Emerging Properties in Self-Supervised Vision Transformers (DINO, ICCV 2021). With Caron, Touvron, Misra, Mairal, Bojanowski, and Joulin; over 11,100 citations.
• DINOv2: Learning Robust Visual Features without Supervision (April 2023). 24-author paper extending DINO to larger models and richer pretraining; over 10,000 citations.
• Moshi: a speech-text foundation model for real-time dialogue (2024). Co-author on the Kyutai full-duplex speech-text foundation model paper, with Alexandre Défossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Patrick Pérez, Edouard Grave, and Neil Zeghidour.
• The FAISS library (2024 / 2025). With Douze and seven Meta colleagues; the first comprehensive technical paper on the FAISS codebase.

Investments and boards

No public investor activity on record in AI, semiconductors, datacenters, software, or energy as of May 2026.

Network

Jegou's research collaborations trace through three overlapping cohorts. The first is his Inria PhD lineage and the early indexing-and-retrieval period: Christine Guillemot was his doctoral advisor, and the longest-running scientific collaboration through the Inria Grenoble and Rennes years was with Cordelia Schmid, Matthijs Douze, and Pérez (then at Technicolor, later Valeo) on the Hamming-embedding, product-quantization, and VLAD lines.

The second is FAIR Paris. The vision-transformer line ran with Hugo Touvron (Jegou's PhD advisee at FAIR), Matthieu Cord, Matthijs Douze, Francisco Massa, and Alexandre Sablayrolles. The DINO and DINOv2 papers added Mathilde Caron, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Yann LeCun was Meta's chief AI scientist during his FAIR tenure. PhD advisees at FAIR include Pierre Fernandez, Alexandre Sablayrolles, and Pierre Stock.

The third is the Kyutai founding cohort: Pérez, Edouard Grave, Laurent Mazaré, Neil Zeghidour, and Alexandre Défossez. The Moshi paper byline collects all six co-founders. The Kyutai scientific committee includes Yann LeCun, Yejin Choi, and Bernhard Schölkopf.

Position in the field

As of May 2026, Jegou occupies a position at the intersection of three research trajectories. The first is the large-scale similarity-search and vector-quantization line through product quantization, VLAD, and FAISS, which made FAISS one of the most widely deployed open-source libraries in production AI infrastructure. The second is the vision-transformer line through DeiT, DINO, DINOv2, and CaiT, which carried much of the visual-representation-learning research direction at FAIR through 2020 to 2023. The third is the Kyutai founding cohort, where his role moved from full-time Chief Science Officer in 2023-2024 to founding alumnus from 2025.

His public-facing presence is moderate. He posts research updates on the @hjegou X account and has presented at major computer-vision and machine-learning venues including CVPR, ICCV, NeurIPS, and ECCV. The 2025 return to Meta as Senior Research Director, while retaining Kyutai co-founder status, makes him one of the few senior FAIR Paris researchers to operate across the FAIR-and-Kyutai axis simultaneously.

Outlook

Open questions over the next 6 to 18 months:

• Vision-research direction at FAIR. Whether the DINO and DeiT vision-transformer lines continue under Jegou's leadership, or whether the 2025 Meta restructuring under Alexandr Wang and Meta Superintelligence Labs reorients the visual-representation-learning program.
• FAISS evolution. The trajectory of the FAISS library through 2026 to 2027, including newer indexing techniques and positioning against commercial vector-database alternatives.
• Kyutai contribution cadence. Whether the continuing co-founder status produces additional Kyutai paper bylines beyond Moshi, or whether the role becomes more advisory.
• Kyutai-FAIR talent flow. Whether the dual Director / co-founder status produces additional collaboration between the two labs.
• PhD pipeline. Whether prior advisees (Touvron, Sablayrolles, Fernandez, Stock) continue producing high-impact research and whether a next cohort emerges through FAIR Paris.

Sources

• Hervé Jégou's personal site at Inria Rennes. Personal homepage with research overview, publication list, and earlier-career biographical details.
• Hervé Jégou's Google Scholar profile. Citation metrics, h-index, and chronological publication record.
• Hervé Jégou on dblp. Bibliographic record listing publications and affiliations.
• Hervé Jégou on OpenReview. Career and education record covering Inria, FAIR, Kyutai, and Meta positions with dates.
• Hervé Jégou's PhD thesis (2005). The 2005 Université de Rennes 1 thesis on robust codes and joint source-channel codes for multimedia transmission over mobile channels, advised by Christine Guillemot.
• Hervé Jégou: visual recognition on a very large scale. Université de Rennes profile covering the 2014 ERC VIAMASS grant, Inria career, and research direction.
• Hervé Jégou, lauréat du prix Bretagne Jeune Chercheur 2010. 2010 ENS Rennes announcement of the Brittany Young Researcher Prize.
• Product Quantization for Nearest Neighbor Search. The 2011 IEEE PAMI paper introducing product quantization.
• Aggregating local image descriptors into compact codes. The 2012 IEEE PAMI paper on the VLAD compact-image-representation line.
• FAISS library on GitHub. The open-source similarity-search and clustering library originated by Jegou and Douze.
• The FAISS library (paper). The 2024 / 2025 comprehensive technical paper on the FAISS codebase.
• Training data-efficient image transformers and distillation through attention. The 2020 DeiT paper introducing data-efficient vision transformers.
• Emerging Properties in Self-Supervised Vision Transformers. The 2021 DINO paper introducing self-distillation with no labels.
• DINOv2: Learning Robust Visual Features without Supervision. The 2023 DINOv2 paper extending the DINO line.
• Moshi: a speech-text foundation model for real-time dialogue. The 2024 Kyutai full-duplex spoken-dialogue model paper.
• Kyutai launch announcement (TechCrunch). November 2023 coverage of the Kyutai launch and founding research team.
• Kyutai team page. Lists Jegou in the Alumni section as a Co-founder.
• Faiss: A library for efficient similarity search. 2017 Engineering at Meta blog post announcing FAISS.