News on Thomas Bouvier

I started a new position at CEA (Maison de la Simulation)

Mon, 29 Dec 2025 00:00:00 +0000

I started working at CEA (Maison de la Simulation) as a research engineer. My role is part of the NumPEx project (http://www.numpex.fr) which aims to build a software stack for Exascale supercomputers including Alice Recoque, scheduled for deployment in 2027. This machine will be among the most powerful in the world (Top500 ranking), used for both traditional scientific applications and artificial intelligence workloads.

My responsibilities include:

Help designing and implementing the packaging and continuous integration strategy for the NumPEx project.
Participating in the deployment and testing of the infrastructure.
Providing user support and training on packaging, deployment, and testing tools and workflows.

The project’s packaging strategy is based on cutting-edge open-source tools such as:

We are launching the OpenFresque digital commons

Tue, 17 Jun 2025 00:00:00 +0000

The Fresk-type awareness workshops enable the engagement of various sectors of activity (industries, transportation, food, buildings, etc.) in their necessary transitions, both in France and around the world.

Currently, tools to manage these events (ticketing, opportunity sharing, pathways, resources) are provided by workshop developers, but this is done in a heterogeneous and non-shared manner.

This poses several problems:

Workshop developers invest in parallel to develop or finance external solutions without pooling these resources.

I will be defending my PhD on Monday, November 4th, 2024

Thu, 31 Oct 2024 00:00:00 +0000

My PhD defense is scheduled for Monday, November 4th, 2024, at 1:30 pm. The event will take place at IRISA/Inria Rennes (Markov room), 263 Avenue du Général Leclerc, 35042 Rennes (maps link). I can’t wait to share this moment with my jury, family, friends and colleagues.

It will also be broadcast live, please get in touch to get the link!

The reviewed dissertation of my thesis can be found here: link to dissertation. It is entitled: “Distributed Rehearsal Buffers for Continual Learning at Scale”. The abstract can be found at the bottom of this page. The presentation will be in English.

I will give a talk at JLESC 16 @ Kobe

Tue, 23 Apr 2024 00:00:00 +0000

I will present our ongoing work entitled “Efficient Distributed Continual Learning for Steering Experiments in Real-Time” at JLESC 16. This presentation is an update on the progress of the JLESC project entitled “Towards Continual Learning at Scale”, which has been running since 2022. You can find the full program here.

Efficient Distributed Continual Learning for Steering Experiments in Real-Time - Project Update

Deep learning has emerged as a powerful method for extracting valuable information from large volumes of data. However, when new training data arrives continuously (i.e., is not fully available from the beginning), incremental training suffers from catastrophic forgetting (i.e., new patterns are reinforced at the expense of previously acquired knowledge). Training from scratch each time new training data becomes available would result in extremely long training times and massive data accumulation. Rehearsal-based continual learning has shown promise for addressing the catastrophic forgetting challenge, but research to date has not addressed performance and scalability. To fill this gap, we propose an approach based on a distributed rehearsal buffer that efficiently complements data-parallel training on multiple GPUs to achieve high accuracy, short runtime, and scalability. It leverages a set of buffers (local to each GPU) and uses several asynchronous techniques for updating these local buffers in an embarrassingly parallel fashion, all while handling the communication overheads necessary to augment input mini-batches (groups of training samples fed to the model) using unbiased, global sampling. After evaluating our approach on classification problems, we further propose a generalization of rehearsal buffers to support generative learning tasks, as well as more advanced rehearsal strategies (notably dark experience replay, leveraging knowledge distillation). We illustrate these extensions with a real-life HPC streaming application from the domain of ptychographic image reconstruction, in which experiments need to be steered in real-time.

One paper has been accepted at CCGrid 2024

Mon, 12 Feb 2024 00:00:00 +0000

Our paper entitled “Efficient Data-Parallel Continual Learning with Asynchronous Distributed Rehearsal Buffers” has been accepted at CCGrid 2024. I will be presenting this work in the “ML for Systems and Systems for ML” track on May 7 at 2:30 PM in Philadelphia. You can find the full program here

Efficient Data-Parallel Continual Learning with Asynchronous Distributed Rehearsal Buffers

Deep learning has emerged as a powerful method for extracting valuable information from large volumes of data. However, when new training data arrives continuously (i.e., is not fully available from the beginning), incremental training suffers from catastrophic forgetting (i.e., new patterns are reinforced at the expense of previously acquired knowledge). Training from scratch each time new training data becomes available would result in extremely long training times and massive data accumulation. Rehearsal-based continual learning has shown promise for addressing the catastrophic forgetting challenge, but research to date has not addressed performance and scalability. To fill this gap, we propose an approach based on a distributed rehearsal buffer that efficiently complements data-parallel training on multiple GPUs, allowing us to achieve short runtime and scalability while retaining high accuracy. It leverages a set of buffers (local to each GPU) and uses several asynchronous techniques for updating these local buffers in an embarrassingly parallel fashion, all while handling the communication overheads necessary to augment input mini-batches (groups of training samples fed to the model) using unbiased, global sampling. In this paper we explore the benefits of this approach for classification models. We run extensive experiments on up to 128 GPUs of the ThetaGPU supercomputer to compare our approach with baselines representative of training-from-scratch (the upper bound in terms of accuracy) and incremental training (the lower bound). Results show that rehearsal-based continual learning achieves a top-5 classification accuracy close to the upper bound, while simultaneously exhibiting a runtime close to the lower bound.

I will give a talk at JLESC 15 @ Bordeaux

Sun, 12 Feb 2023 00:00:00 +0000

I will present our ongoing work entitled “Leveraging Rehearsal Buffers to Enable Efficient Data-Parallel Continual Learning” at JLESC 15. This presentation is an update on the progress of the JLESC project entitled “Towards Continual Learning at Scale”, which has been running since 2022. You can find the full program here.

Leveraging Rehearsal Buffers to Enable Efficient Data-Parallel Continual Learning - Project Update

Deep Learning (DL) emerged as a way to extract valuable information from ever-growing volumes of data. However, when trained on sequential tasks ie. without full access to the dataset at the beginning of the training, typical Deep Neural Networks (DNNs) suffer from catastrophic forgetting, a phenomenon causing them to reinforce new patterns at the expense of previously acquired knowledge. This limitation prevents updating models incrementally, which is problematic in many real-life scenarios where the aforementioned datasets are replaced by data streams generated over time by distributed devices. It is unfeasible to train models from scratch every time new samples are being ingested either, as this would lead to prohibitive time and/or resource constraints.

I will give a talk at JLESC 14 @ Urbana-Champaign

Wed, 21 Sep 2022 00:00:00 +0000

I will present our new JLESC project entitled “Towards Continual Learning at Scale”, as well as preliminary results, at JLESC 14. You can find the full program here. I’m really excited to be giving my first public talk in the US, which will take place at the University of Illinois Urbana-Champaign. 😊

Towards Continual Learning at Scale - Project Kick-off

During the past decade, Deep learning (DL) supported the shift from rule-based systems towards statistical models. Deep Neural Networks (DNNs) are achieving high accuracy on various benchmarks by extracting patterns from complex datasets. Although presenting promising results, most existing supervised learning algorithms operate under the assumptions that the data is (i) i.i.d.; (ii) static; and (iii) available before the training process. These constraints limit their use in real-life scenarios where the aforementioned datasets are replaced by high volume, high velocity data streams generated over time by distributed devices. It is unfeasible to keep training models in an offline fashion from scratch every time new data arrives, as this would lead to prohibitive time and/or resource constraints. At the same time, it is not possible to train learning models incrementally either, due to catastrophic forgetting, a phenomenon causing typical DNNs to reinforce new patterns at the expense of previously acquired knowledge i.e. inducing biases.

I will be spending 4 months working at Argonne National Lab @ Chicago

Wed, 13 Apr 2022 00:00:00 +0000

This summer, I will be working as a visiting PhD student at Argonne National Lab, in the context of the UNIFY Associate Team. I will be mentored by Bogdan Nicolae.

My work is part of the ongoing “Towards Continual Learning at Scale” JLESC project. The project aims to achieve two main objectives: the (1) design and implementation of a distributed replay buffer leveraging distributed systems effectively and the (2) study of trade-offs introduced by large-scale CL in terms of training time, accuracy and memory usage. All experiments will be conducted on the ALCF Polaris supercomputer.