Yann LeCun

Yann LeCun is an AI researcher whose contributions include convolutional neural networks, early backpropagation methods, and architectures for learning abstract representations. He shared the 2018 ACM A.M. Turing Award with Yoshua Bengio and Geoffrey Hinton for conceptual and engineering breakthroughs in deep neural networks. His later research agenda proposes self-supervised world models that support prediction, reasoning, and planning.[4][6]

LeCun is a founding-team member of AMI Labs, alongside Alex LeBrun, Laurent Solly, Mike Rabbat, Min Lin, Pascale Fung, and Saining Xie. The lab aims to learn abstract representations of sensor data and predict in representation space while ignoring unpredictable details. Its proposed action-conditioned world models would predict consequences of actions so agents could plan sequences before acting; these are research aims rather than demonstrated general capabilities.[2]

His world-model proposals followed decades of work on trainable visual recognition and modular learning systems. That earlier work supplies context for the representation-learning methods he later pursued.[4][6]

LeCun earned his computer-science PhD at Université Pierre et Marie Curie in 1987. He performed postdoctoral work under Hinton’s supervision and worked with Bengio at Bell Labs beginning in the early 1990s.[3][4]

In the 1980s, LeCun developed convolutional neural networks; in the late 1980s, at the University of Toronto and Bell Labs, he trained convolutional systems on handwritten-digit images. The work reached operational document processing: according to NYU, an AT&T check-reading system based on his convolutional-network research was reading about 20 percent of checks written in the United States by the late 1990s.[4][5]

LeCun also proposed an early version of backpropagation, derived it using variational principles, and described methods to accelerate learning. With Léon Bottou, he proposed learning systems composed of networks of modules, with backpropagation performed through automatic differentiation. Their work also included deep-learning architectures for structured data such as graphs.[4]

Facebook appointed LeCun to direct a new AI and deep-learning laboratory in December 2013, with his tenure scheduled to begin January 1, 2014. The NYU announcement said he would retain part-time research and teaching at Courant and relinquish the directorship of NYU’s Center for Data Science once a replacement was selected. By March 2019, he was Facebook’s VP and Chief AI Scientist as well as an NYU professor. NYU’s faculty directory lists him as Silver Professor of Computer Science.[5][4][3]

In February 2022, LeCun proposed learning world models—internal models of how the world works—through self-supervision, then using them to predict, reason, and plan. World models already had a long history in psychology, control, and robotics; his proposal combined those ideas with self-supervised learning and joint-embedding architectures.[6]

His proposed architecture has six differentiable modules: a configurator that prepares the system for a task, perception that estimates the current state, a world model that fills in missing information and predicts outcomes, a cost module, an actor that proposes action sequences, and short-term memory. Planning would optimize proposed actions against estimated future costs, then execute the first action. As he explained in the 2022 presentation, “Because the cost module is differentiable, the gradient of the cost can be back-propagated through the other modules for planning, reasoning, or learning.” Training the critic, configuring tasks, and implementing memory remained unresolved parts of the proposal.[6]

The Joint Embedding Predictive Architecture (JEPA) encodes two inputs into abstract representations and trains a predictor to estimate the target representation from the context representation. For video, the inputs could be successive segments. The proposed encoder can discard hard-to-predict information, while a latent variable can represent multiple plausible outcomes. The practical distinction is between anticipating nearby cars and predicting every roadside leaf: the former matters for driving, while the latter need not be represented in detail.[6]

LeCun coauthored I-JEPA, first submitted in January 2023, with Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, and Nicolas Ballas. I-JEPA predicts representations of target blocks from a context block in the same image. The team identified large target blocks and an informative, spatially distributed context as central masking choices for learning semantic representations.[7]

V-JEPA, released by Meta in February 2024, extends representation prediction to video. It pretrains on unlabeled visual data, masking regions across both space and time so the task cannot be solved simply by observing adjacent frames. Meta reported adapting the pretrained system to downstream tasks through frozen evaluation: retaining its pretrained components while training a small task-specific layer or network. The model was released under CC BY-NC, a noncommercial license.[8]

LeCun called V-JEPA “a step toward a more grounded understanding of the world so machines can achieve more generalized reasoning and planning” in the release announcement. The work then concerned perception, including fine-grained object interactions over short video clips; planning and sequential decision-making were stated next steps.[8]

In November 2025, Randall Balestriero and LeCun introduced LeJEPA, combining JEPA’s predictive loss with Sketched Isotropic Gaussian Regularization (SIGReg) to encourage isotropic Gaussian embeddings. The authors presented it as a way to reduce training heuristics, including teacher–student arrangements and stop-gradient operations. They reported 79 percent in ImageNet-1k frozen-backbone linear evaluation with a ViT-H/14—a result in image representation learning, rather than a demonstration of general planning.[9]