Inria, the French national research institute for the digital sciences.
Offer Description
The position will be in the framework of the ERC Starting Grant DYNASTY (Dynamics‑Aware Theory of Deep Learning).
The position might include traveling to conferences for paper presentation. Travel expenses will be covered within the limits of the scale in force.
The rapid success of diffusion models [1, 2, 3], which have achieved state‑of‑the‑art performance across diverse domains, motivates the need for a theoretical understanding of the mechanisms underpinning their strong capabilities. The unique structure of these models, as well as recent evidence [4, 5] indicating they forgo the advantageous properties of benign overfitting, suggests that diffusion models are a fundamental divergence from traditional deep learning paradigms. This suggests that existing generalisation theories are insufficient and highlights the need for a bespoke, algorithm‑dependent framework to capture the phenomena present in these models.
We are seeking a postdoctoral researcher to develop theoretical frameworks for analysing generalisation and memorisation in diffusion models. The project's central goal is to move beyond algorithm‑independent bounds and develop rigorous theory that unpacks the generalisation properties inherent to the training and sampling processes. A key component of this analysis will be to precisely characterise the mechanisms driving memorisation. This theoretical work will serve as the foundation for developing well‑founded algorithms that target the task of preventing data copying (e.g. [6, 7]) as well as the problem of data attribution (e.g. [8, 9]), allowing for the precise measurement of individual training examples' influence on model outputs.
References

[1] Sohl-Dickstein, Jascha, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. 2015. “Deep Unsupervised Learning Using Nonequilibrium Thermodynamics.” ICML.
[2] Ho, Jonathan, Ajay Jain, and Pieter Abbeel. 2020. “Denoising Diffusion Probabilistic Models.” NeurIPS.
[3] Song, Yang, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2021. “Score-Based Generative Modeling through Stochastic Differential Equations.” ICLR.
[4] Pidstrigach, Jakiw. 2022. “Score-Based Generative Models Detect Manifolds.” NeurIPS.
[5] Dupuis, Benjamin, Dario Shariatian, Maxime Haddouche, Alain Durmus, and Umut Simsekli. 2025. “Algorithm- and Data-Dependent Generalization Bounds for Score-Based Generative Models.” arXiv [Stat.ML].
[6] Vyas, Nikhil, Sham M. Kakade, and Boaz Barak. 2023. “On Provable Copyright Protection for Generative Models.” ICML.
[7] Alberti, Silas, Kenan Hasanaliyev, Manav Shah, and Stefano Ermon. 2025. “Data Unlearning in Diffusion Models.” ICLR.
[8] Mlodozeniec, Bruno Kacper, Runa Eschenhagen, Juhan Bae, Alexander Immer, David Krueger, and Richard E. Turner. 2025. “Influence Functions for Scalable Data Attribution in Diffusion Models.” ICLR.
[9] Zheng, Xiaosen, Tianyu Pang, Chao Du, Jing Jiang, and Min Lin. 2024. “Intriguing Properties of Data Attribution on Diffusion Models.” ICLR.

Main activities

Conduct theoretical research.
Conduct experiments for empirical verification.
Write scientific articles.
Disseminate the scientific work in appropriate venues.

Requirements

High professional/academic English proficiency.
Good coding skills in Python and related deep-learning libraries.
Basic French language skills.

Additional Information

Partial reimbursement of public transport costs.
Leave: 7 weeks of annual leave + 10 extra days off due to RTT + possibility of exceptional leave.
Possibility of teleworking and flexible organization of working hours.
Professional equipment available (videoconferencing, loan of computer equipment, etc.).
Social, cultural and sports events and activities.

#J-18808-Ljbffr