Project Overview
MEPHISTO is a research project aiming to develop mathematically founded, resource-efficient optimization methods for large-scale machine learning and signal processing. The project focuses on building and exploiting hierarchical structures— in data, models, algebra, and numerical precision—to achieve a balance between accuracy and parsimony.
Research Axes
- Axis A: Physical and geometrical hierarchies in imaging and PDEs.
- Axis B: Structural hierarchies in neural network architectures and datasets.
- Axis C: Hierarchical algebraic tools (e.g., butterfly factorizations).
- Axis D: Mixed numerical precision and dynamic quantization strategies.
Goals
- Reduce the computational cost of DNN training and inference.
- Enable efficient multiresolution image restoration.
- Design low-resource algorithms with rigorous error analysis.
- Bridge variational optimization and deep learning with hierarchical methods.
Team
Principal Investigator: Elisa Riccietti (MCF, ENS Lyon)
Collaborators: Ockham team, Nelly Pustelnik (ENS Lyon), Nicolas Brisebarre (ENS Lyon), Théo Mary (LIP6–Sorbonne), Silviu Filip (Inria Rennes) and others
Funding: ANR JCJC 2024–2028
Involvement: 1 PhD, 1 Postdoc.