Current research interests and projects
- As a member of PEPR IA Holigrail, I am working on hardware-aware quantization and training of deep neural networks. With F. de Dinechin and our students we are working on new custom data formats for machine learning and associated arithmetic operators. The goal is to provide efficient quantisation-aware training tools that bridge the hardware gap between simulation and inference deployment. We target convolutional networks deployed on FPGA.
- For larger-scale LM models, such as BERT, mixed-precision approaches using standard integer or low-precision floating-point formats are more suitable. With my postdoc Cédric and PhD student Xavier, we are looking into quantization and pruning of large language models, both during training and fine-tuning to domain-specific data sets. Furthermore, with Richard Dufour we look into the impact of quantization onto language capabilities of those heavily reduced models.
- From mathematical optimization point of view, I am interested in modelling hardware circuit design as a combinatorial optimization problem, and then solving it with different exact or heuristic approaches. We are currently working on new constraint programming-based models with Christine Solnon and our student Theo.
- On signal processing side, I am currently collaborating with Romain Michon, through supervision of a master student, on source separation models for hearing aid improvement, based on Convolutional Neural Networks. We are interested in making a connection between the arsenal of quantization techniques I have for DNNs and an actual deployment of a complex CNN for audio signal processing for real-time inference.
PhD students and postdocs
- 2025-now Théo Cantaloube is working on constraint programming for digital circuit design, co-supervision with C. Solnon
- 2025 Cédric Gernigon worked on mixed-precision quantization of language models during his postdoc with me
- 2024-now Romain Bouarah is working on custom data formats and arithmetic operators for DNN inference, co-supervision with F. de Dinechin
- 2024-now Bastien Barbe is working on hardware architectures for efficient DNN inference, co-supervision with F. de Dinechin
- 2023 – now Xavier Pillet is working on Natural Language Processing acceleration, CIFRE thesis with Valeuriad, co-supervision with N. Greffard and R. Dufour at Nantes university
- 2021-2022 Wassim Seifeddine, worked on dynamic-precision training of DNN, supervisor 50%, with C. Jermann et S. Filip. Wassim quit the thesis after 12 months due to moving out of Nantes.
- 2020-2023 Rémi Garcia, defended his thesis on optimisation of multiple constant multiplication operators https://theses.hal.science/tel-04606847
Professional experience
Oct 2023 –now Researcher (tenured), CITI laboratory, Inria, Lyon, France
Sep 2019 – Sep 2023 Associate Professor (tenured), Nantes Université, Computer Science department and LS2N research lab, team OGRE, Nantes, France
Apr – Aug 2019 Research fellow, Intel Corporation – AI Research Group, San Diego, USA
Feb – Apr 2019 Invited Researcher, Max Planck Institute for Software Systems, Kaiserslautern, Germany
Oct 2017 – Feb 2019 Postdoctoral researcher, Inria – ENS Lyon – LIP, Lyon, France
Sep 2014 – Sep 2017 PhD Candidate & Teaching Assistant, Sorbonne Universités – UPMC – LIP6, Paris, France
Sep 2012 – Aug 2013 Software engineer, IT department of Odessa National University, Odessa, Ukraine
Education
2024-2017 PhD in computer Science at Sorbonne University, University Paris 6 and LIP6 laboratory, Paris, France
2008-2014 Bachelor and Master in Applied Mathematics at Odessa National I. I. Metchnikov University, Ukraine