Milad Sefidgaran
Research scientist working on the foundations of machine learning algorithms.
My research is about generalization and its relation to the geometry of the representations: why machine learning models work on data they have never seen during training, and how to turn that understanding into better algorithms. I mostly use ideas from statistical learning theory and information theory. Over the past few years I have been bringing that perspective to structured representation learning for vision, large language models, and reinforcement learning: how these models form their internal representations; how the structure learned by an attention block differs from that of a CNN block; and how to put them to use in designing more efficient learning algorithms, by discovering such structures and learning how to shape them.
I am a Principal Research Scientist at Huawei’s Paris Research Center, where I lead a small research group focused on the foundations of AI. Our research spans structured representation learning, generalization, and zero-shot reinforcement learning. We also work on LLM capability assessment and post-training methods, including supervised fine-tuning and reinforcement learning-based approaches.
For more on my research interests, see the research page. You can find my publications on Google Scholar and on the publications page, or reach me by email.
News
| Apr 01, 2026 | Two papers accepted at ISIT 2026: on the generalization of next-token prediction, and the effect of data heterogeneity in distributed learning. |
|---|---|
| Mar 18, 2026 | Our paper On the Effect of Clients-Server Communication on the Generalization Error of Federated Learning was accepted for publication in the IEEE Transactions on Information Theory. |
| Sep 18, 2025 | Tighter CMI-based generalization bounds via projection and quantization was selected for an Oral at NeurIPS 2025 (top 0.36% of submissions). |
| Sep 01, 2025 | Selected as team lead for the foundations-of-AI research group at Huawei Paris, spanning generalization, LLM inference, and reinforcement learning. |
| Jan 22, 2025 | Generalization guarantees for representation learning via data-dependent Gaussian mixture priors was accepted as a Spotlight at ICLR 2025 (top 1.6% of submissions). |
Selected Publications
- Tighter CMI-based generalization bounds via projection and quantizationIn Neural Information Processing Systems (NeurIPS), 2025
- Generalization guarantees for representation learning via data-dependent Gaussian mixture priorsIn International Conference on Learning Representations (ICLR), 2025
- Lessons from generalization error analysis of federated learning: you may communicate less often!In International Conference on Machine Learning (ICML), 2024
- Minimum description length and generalization guarantees for representation learningIn Neural Information Processing Systems (NeurIPS), 2023
- Rate-distortion theoretic bounds on generalization error for distributed learningIn Neural Information Processing Systems (NeurIPS), 2022
- Rate-distortion theoretic generalization bounds for stochastic learning algorithmsIn Conference on Learning Theory (COLT), 2022
- Heavy tails in SGD and compressibility of overparametrized neural networksIn Neural Information Processing Systems (NeurIPS), 2021The first two authors contributed equally