Antoine Dedieu

About me

I am a Senior Research Scientist at Google DeepMind. I am currently interested in building agents that can efficiently learn new tasks in new environments. I believe that such an agent must learn a latent world model, which pairs (a) a representation model mapping environment observations to a rich compact latent space, with (b) a generative world model describing the representation dynamics. I also believe that LLMs / VLMs can be leveraged as rich priors for solving these new tasks. My research then lies at the intersection of representation learning, model-based reinforcement learning and in-context learning.

Prior to joining Google DeepMind, I worked 3.5 years as a Research Scientist then as a Senior Research Scientist at Vicarious (acquired by Alphabet) where I was building an AI layer for robots. My research there was focus on building novel generative probabilistic graphical models (PGMs) to solve challenging (a) object-centric vision problems, and (b) navigation problems, and on deriving new methods for learning and inference in complex PGMs.

Prior to joining Vicarious, I graduated from MIT Operations Research Center Master of Science, where I was advised by Prof. Rahul Mazumder. My reseach was focused on building new algorithms to compute interpretable estimators, and studying their statistical performance. Before MIT, I earned an MS in Applied Mathematics from Ecole Polytechnique.

I am a movie-fan, a sports addict, and a world-traveller. As an amateur photograph, I share my journeys around on my website.

---

Conference articles

• Schema-learning and rebinding as mechanisms of in-context learning and emergence. [Preprint] [Code]
  Neurips 2024, to appear. Antoine Dedieu*, Joseph Ortiz*, Wolfgang Lehrach, Swaroop Guntupalli, Carter Wendelken, Ahmad Humayun, Sivaramakrishnan Swaminathan, Guangyao Zhou, Miguel Lázaro-Gredilla, Kevin Murphy.
• Learning Cognitive Maps from Transformer Representations for Efficient Planning in Partially Observed Environments. [Preprint]
  ICML 2023. Antoine Dedieu, Wolfgang Lehrach, Guangyao Zhou, Dileep George, Miguel Lázaro-Gredilla.
• Schema-learning and rebinding as mechanisms of in-context learning and emergence. [Preprint]
  Neurips 2023, Spotlight. Sivaramakrishnan Swaminathan, Antoine Dedieu, Rajkumar Vasudeva Raju, Murray Shanahan, Miguel Lázaro-Gredilla, Dileep George.
• Learning noisy-OR Bayesian Networks with Max-Product Belief Propagation. [Preprint] [Code]
  ICML 2023. Antoine Dedieu, Guangyao Zhou, Dileep George, Miguel Lázaro-Gredilla.
• Graphical Models with Attention for Context-Specific Independence and an Application to Perceptual Grouping. [Preprint] [Code]
  Guangyao Zhou, Wolfgang Lehrach, Antoine Dedieu, Miguel Lázaro-Gredilla, Dileep George.
• Perturb-and-max-product: Sampling and learning in discrete energy-based models. [Preprint] [Code]
  Neurips 2021. Miguel Lázaro-Gredilla, Antoine Dedieu, Dileep George.
• Sample-Efficient L0-L2 Constrained Structure Learning of Sparse Ising Models. [Preprint] [Code]
  AAAI 2021. Antoine Dedieu, Miguel Lázaro-Gredilla, Dileep George.
• Query Training: Learning a Worse Model to Infer Better Marginals in Undirected Graphical Models with Hidden Variables. [Preprint] [Code]
  AAAI 2021. Miguel Lázaro-Gredilla, Wolfgang Lehrach, Nishad Gothoskar, Guangyao Zhou, Antoine Dedieu, Dileep George.
• Improved error rates for sparse (group) learning with Lipschitz loss functions. [Preprint]
  Antoine Dedieu.
• An error bound for Lasso and Group Lasso in high dimensions. [Preprint]
  Antoine Dedieu.
• Learning higher-order sequential structure with cloned HMMs. [Preprint]
  Antoine Dedieu, Nishad Gothoskar, Scott Swingle, Wolfgang Lehrach, Miguel Lázaro-Gredilla, Dileep George.
• Error bounds for sparse classifiers in high-dimensions. [Preprint]
  AiStats 2019. Antoine Dedieu.
• Hierarchical Modeling and Shrinkage for User Session Length Prediction in Media Streaming. [Preprint] [Code]
  CIKM 2018. Antoine Dedieu, Rahul Mazumder, Zhen Zhu, Hossein Vahabi.

Journal articles

• PGMax: Factor Graphs for Discrete Probabilistic Graphical Models and Loopy Belief Propagation in JAX. [Preprint] [Code]
  Journal of Machine Learning Research (accepted with minor revisions), 2023. Guangyao Zhou, Antoine Dedieu, Nishanth Kumar, Miguel Lázaro-Gredilla, Shrinu Kushagra, Dileep George.
• Learning attention-controllable border-ownership for objectness inference and binding. [Preprint]
  Antoine Dedieu, Rajeev V. Rikhye, Miguel Lázaro-Gredilla, Dileep George.
• A detailed mathematical theory of thalamic and cortical microcircuits based on inference in a generative vision model. [Preprint]
  Dileep George, Miguel Lázaro-Gredilla, Wolfgang Lehrach, Antoine Dedieu, Guangyao Zhou.
• Solving L1-regularized SVMs and related linear programs: Revisiting the effectiveness of Column and Constraint Generation. [Preprint] [Code]
  Journal of Machine Learning Research, 2022. Antoine Dedieu, Rahul Mazumder, Haoyue Wang
• Subset Selection with Shrinkage: Sparse Linear Modeling when the SNR is low. [Preprint] [Code]
  Operations Research, 2022. Rahul Mazumder, Peter Radchenko, Antoine Dedieu.
• Learning Sparse Classifiers: Continuous and Mixed Integer Optimization Perspectives. [Preprint] [Code]
  Journal of Machine Learning Research, 2021. Antoine Dedieu, Hussein Hazimeh, Rahul Mazumder.
• Clone-structured graph representations enable flexible learning and vicarious evaluation of cognitive maps. [Preprint]
  Nature Communications, 2021. Dileep George, Rajeev V Rikhye, Nishad Gothoskar, J Swaroop Guntupalli, Antoine Dedieu, Miguel Lázaro-Gredilla.

Thesis

• Sparse learning: statistical and optimization perspectives.
  MIT Libraries. Antoine Dedieu.