Keywords: Optimal Transport, Sinkhorn algorithm, Online Learning, Machine Learning, Signal Processing.
Description:
Optimal transport (OT) is a powerful framework to define and compute distances between distributions (a.k.a. Wasserstein or earth mover’s distance), with a tractable computation thanks to the Sinkhorn algorithm [1]. Entropic regularization, which enables fast iterative scaling and GPU-friendly computation of the OT, providing the backbone of modern scalable OT pipelines.
Major challenges arise in exploring OT for domain adaptation on streaming data. While an online Sinkhorn algorithm was introduced in [2] to address the OT distances computation from sample streams, it does not operate in an online manner. The major issue is that the functions are evaluated on an increasing amount of newly available samples. which yields a memory complexity of O(n) and a time complexity of O(n^2). In order to fully operate in an online manner, the evaluations should not rely on all the previously available samples. Some attempts were provided to mitigate this major issue, such as by performing measure compression techniques (which are computationally expensive) [3] and by combining streaming quantile approximation with sliced OT [4].
This post-doc fellowship aims to provide theoretical foundations and algorithmic developments for OT on streaming data, mainly time series. For this purpose, the post-doc fellow will leverage earlier research results and take full advantage of the literature of adaptive signal processing and representation learning with deep learning.
This post-doc is an integral part of the global project OOD (Online Deep anomaly Detection), bringing together 4 PhD students and several permanent researchers from the Machine Learning group of the LITIS Lab.
- [1] G. Peyré and M. Cuturi, “Computational optimal transport: With applications to data science,” Foundations and Trends® in Machine Learning, vol. 11, no. 5-6, pp. 355–607, 2019.
- [2] A. Mensch and G. Peyré, “Online sinkhorn: Optimal transport distances from sample streams,” in NeurIPS, vol. 33, pp. 1657–1667, 2020.
- [3] F. Wang, C. Poon and T. Shardlow, « Compressed online Sinkhorn,” arXiv preprint arXiv:2310.05019, 2023.
- [4] K. Nguyen, « Streaming Sliced Optimal Transport, » arXiv preprint arXiv:2505.06835, 2025.
Duration : 18 months
Start of contract : when available
Host lab : LITIS Lab (Rouen metropolitan area)
Requirements:
- PhD in applied mathematics, machine learning, advanced statistics, computer science, signal processing or related.
- Strong background in advanced optimization and machine learning.
- Proficiency in Python.
If interested, please send CV in a motivational email to paul.honeine@univ-rouen.fr, gilles.gasso@insa-rouen.fr, maxime.berar@univ-rouen.fr and fannia.pacheco@univ-rouen.fr.
