Yann LeCun is a French-American computer scientist who is VP and Chief AI Scientist at Meta and a professor at New York University. Together with Geoffrey Hinton and Yoshua Bengio, he shared the 2018 ACM A.M. Turing Award for contributions to deep learning. While his primary contributions are in computer vision, his development of convolutional neural networks and his advocacy for self-supervised learning have had substantial impact on text classification, character recognition, and representation learning for NLP.
Early Life and Education
Born in Paris, France, in 1960, LeCun studied at the Ecole Superieure d'Ingenieurs en Electrotechnique et Electronique (ESIEE) and earned his PhD from the Universite Pierre et Marie Curie in 1987. He worked at Bell Laboratories, where he developed LeNet for handwritten digit recognition, before joining NYU and later Facebook (now Meta) AI Research.
Born in Paris, France
Completed PhD at Universite Pierre et Marie Curie
Published backpropagation applied to handwritten zip code recognition
Published LeNet-5 and gradient-based learning for document recognition
Became director of Facebook AI Research (FAIR)
Received the ACM Turing Award with Hinton and Bengio
Key Contributions
LeCun's convolutional neural networks (CNNs) apply learned filters across input data using shared weights, dramatically reducing the number of parameters compared to fully connected networks. While developed for image recognition, CNNs have been widely adopted for NLP tasks: text classification using 1D convolutions over word embeddings (Kim, 2014), character-level models that learn features directly from characters, and sentence modelling using hierarchical convolutions.
LeCun has been a leading advocate for self-supervised learning, arguing that learning representations from unlabelled data through prediction tasks is the key to general intelligence. This philosophy directly aligns with the pre-training paradigm in NLP, where models like BERT learn from masked language modelling and GPT learns from next-word prediction — both forms of self-supervised learning on unlabelled text.
"Self-supervised learning is the cake, supervised learning is the icing on the cake, and reinforcement learning is the cherry on the cake." — Yann LeCun, on the importance of self-supervised learning
Legacy
LeCun's CNNs are used in text classification systems worldwide and his backpropagation techniques are fundamental to training all neural NLP models. His advocacy for self-supervised learning anticipated and inspired the pre-training revolution in NLP. As head of FAIR, he has overseen research that produced FastText, RoBERTa, and other influential NLP contributions. His emphasis on energy-based models and representation learning continues to influence the direction of both vision and language AI research.