Hoang Trieu Vy LE

I am a Senior Research Scientist at Beink Dream, where I lead AI initiatives from fundamental research to production deployment—designing and scaling AI systems for 2D and 3D applications. Previously, I worked at CEA (France) and earned my Ph.D. from École Normale Supérieure under the supervision of Prof. Nelly Pustelnik.

My work focuses on bridging physics and AI, with a particular emphasis on efficient model design. I am especially driven by the challenge of learning compact, physically grounded representations from continuous high-dimensional signals, with strong connections to modern physics research.

Email  /  CV  /  Scholar  /  Github /  Linkedin

Ph.D Thesis

Supervisors: Dr. Nelly Pustelnik, Dr. Marion Foare

Reviewers: Prof. Caroline Chaux, Assoc. Prof. Matthieu Kowalski

Examiners: Prof. Jean-François Aujol, Prof. Elie Bretin, Prof. Luca Calatroni, Prof. Marc Sebban

Publications

My research spans unfolded neural networks, image processing, computer vision, and generative AI.

We propose an unfolded neural network approach for simultaneous image denoising and contour detection, bridging variational procedures designed to perform the combined denoising/edge detection task and black-box neural networks designed for edge detection.

In this paper, we present a new edge detection model based on proximal unfolded neural networks. The architecture relies on unfolding proximal Blake-Zisserman iterations, leading to a composition of two blocks: a smoothing block and an edge detection block.

@ARTICLE{10630640,
  author={Le, Hoang Trieu Vy and Repetti, Audrey and Pustelnik, Nelly},
  journal={IEEE Transactions on Image Processing}, 
  title={Unfolded Proximal Neural Networks for Robust Image Gaussian Denoising}, 
  year={2024},
  volume={33},
  number={},
  pages={4475-4487},
  doi={10.1109/TIP.2024.3437219}}

We propose different learning strategies for our PNN framework, and investigate their robustness (Lipschitz property) and denoising efficiency. Finally, we assess the robustness of our PNNs when plugged in a forward-backward algorithm for an image deblurring problem

@ARTICLE{9723590,
  author={Le, Hoang Trieu Vy and Foare, Marion and Pustelnik, Nelly},
  journal={IEEE Signal Processing Letters}, 
  title={Proximal Based Strategies for Solving Discrete Mumford-Shah 
        With Ambrosio-Tortorelli Penalization on Edges}, 
  year={2022},
  volume={29},
  number={},
  pages={952-956},
  doi={10.1109/LSP.2022.3155307}}

This work is dedicated to joint image restoration and contour detection considering the Ambrosio-Tortorelli functional. Two proximal alternating minimization schemes with convergence guarantees are provided, PALM-AT and SL-PAM-AT, as well as closed-form expressions of the involved proximity operators.

@INPROCEEDINGS{9909592,
  author={Le, Hoang Trieu Vy and Pustelnik, Nelly and Foare, Marion},
  booktitle={2022 30th European Signal Processing Conference (EUSIPCO)}, 
  title={The faster proximal algorithm, the better unfolded deep learning architecture ? 
    The study case of image denoising}, 
  year={2022},
  pages={947-951},
  doi={10.23919/EUSIPCO55093.2022.9909592}}

In this work, we proposed two deep unfolded networks for gaussian denoising that can be activated from standard or accelerated schemes to illustrate the benefit to unroll accelerated schemes when possible.

@inproceedings{HoangTVL2022AlgorithmesPR,
  title={Fast Proximal Unrolled Algorithms for the Analysis of Piecewise Homogeneous Fractal Images},
  author={ Hoang T.V Le, Barbara Pascal, Nelly Pustelnik, Marion Foare and Patrice Abry},
  year={2022},
  url={https://api.semanticscholar.org/CorpusID:253115164}
}

We propose two unrolled deep network architectures built from the proximal FISTA and Chambolle-Pock algorithms to estimate local regularity in piecewise homogeneous fractal images.

Talks

Teaching