ADONIS - Asynchronous Decentralized Optimization of MachiNe LearnIng ModelS

Project ANR-21-CE23-0030 ADONIS and EMERG-ADONIS from Alliance SU

Welcome to the homepage of the ADONIS research project

The ADONIS project explores asynchronous decentralized optimization for both convex and non-convex problems. In the convex setting, we can solve well-posed mathematical problems and obtain theoretical guarantees regarding the convergence of optimization algorithms. However, deep learning poses a unique challenge in the non-convex setting, where practical performance often deviates significantly from theoretical predictions. In light of this, the ADONIS project adopts a principled approach to optimization research, guided by practical considerations and real-world applications.

Job offers

We are currently proposing an internship on Adaptive optimization for Decentralized DNNs.

news

Jul 29, 2023 The ADONIS team is giving a talk at the Localized Learning Workshop of ICML 2023.
Jun 26, 2023 Louis Fournier gave a talk on Forward Gradients to the MLIA team.
Jun 15, 2023 Edouard Oyallon is giving a talk at Centre Borelli about the recent advances of the project.
Jun 2, 2023 Welcoming Belilovsky Lab director, Eugene Belilovsky, for a month.
Apr 24, 2023 Two papers accepted at ICML 2023, on accelerated gossip and forward gradient methods.

selected publications

  1. DADAO: Decoupled Accelerated Decentralized Asynchronous Optimization
    Adel Nabli, and Edouard Oyallon
    In Proceedings of the 40th International Conference on Machine Learning, 2023
  2. Can Forward Gradient match Backpropagation?
    Louis Fournier, Stéphane Rivaud, Eugene Belilovsky, and 2 more authors
    In Proceedings of the 40th International Conference on Machine Learning, 2023