Postdoc: Energy-Efficient AI Diffusion Research

Centrale Lyon; Écully; Sur place; EUR 32 000 - 42 000; Plein temps

Résumé du poste

CENTRALE LYON invites a post-doctoral researcher in mathematics of energy-efficient AI, focusing on diffusion models on emerging hardware. The work aims to quantify score errors and link device physics to theory, advancing energy efficiency in diffusion-based inference.

You will contribute to a co-design with INL and develop a measurable energy–accuracy frontier. The 12-month contract starts 1 December 2026.

Qualifications

• PhD in applied mathematics, statistics, machine learning, electronic engineering, computer science or closely related field; defence before start.
• Strong foundation in probability and stochastic analysis, numerical analysis, optimisation, or machine learning: SDEs, Girsanov’s theorem, time reversal of diffusions.
• Familiarity with convergence theory of sampling algorithms or ability to acquire quickly.
• Solid scientific computing (PyTorch or JAX) and ability to write runnable code across modelling and hardware-aware AI.
• Working English; French is not required.

Responsabilités

• Make the score error measurable starting from Gaussian targets with closed-form error decomposition.
• Map device physics onto theory and propagate error models into a score-error budget.
• Extend guarantees to mixed write-time and per-call errors in score evaluation.
• Co-design with INL to deliver a proof of concept, vary sampler steps and precision to find energy–accuracy trade-offs.

Connaissances

PhD in relevant field

Stochastic analysis

Convergence theory of sampling

PyTorch/JAX

English proficiency

Formation

PhD in a related field

Outils

PyTorch

JAX

Description du poste

CENTRALE LYON invites a post-doctoral researcher in mathematics of energy-efficient AI, focusing on diffusion models on emerging hardware. The work aims to quantify score errors and link device physics to theory, advancing energy efficiency in diffusion-based inference.

You will contribute to a co-design with INL and develop a measurable energy–accuracy frontier. The 12-month contract starts 1 December 2026.