Hi,

I'm Mickaël Chen,

Currently a Deep Learning Research Scientist at Valeo.ia in Paris, working with Deep Generative and Spatio-Temporal Models.

You can contact me in French or English.

Publications

Unifying GANs and Score-Based Diffusion as Generative Particle Models

By describing the trajectories of GAN outputs during training with particle evolution equations, we propose an unifying framework for GAN and Diffusion Models. We provide a new insights on the role of the generator network, and as proof of concept validating our theories, we propose methods to train a generator with score-based gradient instead of a discriminator, or to use a discriminator's gradient flow to generate instead of training a generator.

Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos*, Thibaut Issenhuth*, Emmanuel de Bézenac*, Mickaël Chen*, Alain Rakotomamonjy*

[Preprint] [Code soon]

Will be presented at NeuRIPS 2023.

Resilient Multiple Choice Learning: A learned scoring scheme with application to audio scene analysis

By casting the problem of Multi-Choice Learning as conditional distribution estimation, we propose a novel algorithm that can be used in regression settings. Our method is equipped with an interpretable scoring scheme and a probabilistic interpretation that we validate on synthetic data, and we demonstrate its suitability for multi-target regression on audio source localization tasks.

Victor Letzelter, Mathieu Fontaine, Patrick Perez, Gaël Richard, Slim Essid, Mickael Chen.

[Preprint soon] [Code soon]

Will be presented at NeuRIPS 2023.

Towards Motion Forecasting with Real-World Perception Inputs: Are End-to-End Approaches Competitive?

We build an evaluation framework for motion forecasting method that assess their performance not in isolation but as part of a realistic deployment pipeline. We are able to show, for the first time, that not only recent end-to-end methods fail when compared to conventional approaches, but also that conventional approaches are far from being able to perform correctly in the real-world. We conduct an in-depth study and make recommandations accordingly.

Yihong Xu, Loïck Chambon, Éloi Zablocki, Mickaël Chen, Alexandre Alahi, Matthieu Cord, Patrick Pérez

[Preprint] [Code soon]

Will be presented at ICCV Workshop ROAD++ 2023.

DiffHPE: Robust, Coherent 3D Human Pose Lifting with Diffusion

We show that diffusion models lead to more robust estimations in the face of occlusions, and improve the time-coherence and the symmetry of predictions.
Moreover, combined with supervised models, we improve can their accuracy further.

Cédric Rommel, Eduardo Valle, Mickaël Chen, Souhaiel Khalfaoui, Renaud Marlet, Matthieu Cord, Patrick Pérez

[Preprint] [Code soon]

Will be presented at ICCV Workshop AMFG 2023.

OCTET: Object-aware Counterfactual Explanations

We propose a generative method for counteractual explanations. Our method is object-aware, allowing for fine-grained, disentangled and per object, explanations.

Mehdi Zemni, Mickaël Chen, Éloi Zablocki, Hédi Ben-Younes, Patrick Pérez, Matthieu Cord

[Preprint] [GitHub]

Presented at CVPR2023.

Multi-Modal 3D GAN for Urban Scenes

We incorporate LiDAR points to condition generative NeRF for 3D driving scene generation.

Loïck Chambon, Mickael Chen, Tuan-Hung Vu, Alexandre Boulch, Andrei Bursuc, Matthieu Cord, Patrick Pérez

[Paper and Video][Poster]

Presented at NeurIPS Workshop ML4AD 2022.

STEEX: Steering Counterfactual Explanations with Semantics

We propose a method to generate counterfactual images that scales up to complex scenes, with application for autonomous driving tasks. To do so, we leverage recent segmentation-to-image networks.

Paul Jacob, Éloi Zablocki, Hédi Ben-Younes, Mickaël Chen, Patrick Pérez, Matthieu Cord

[Preprint] [GitHub]

Presented at ECCV2022.

A Neural Tangent Kernel Perspective of GANs

GAN analyses by modeling the discriminator in its infinite-width limit. By taking into account the impact of the neural architecture and gradient descent training, this modelization explains, among other results, why GANs can be trained even though an optimal arbitrarily powerful discriminator would provide no gradients.

Jean-Yves Franceschi*, Emmanuel de Bézenac*, Ibrahim Ayed*, Mickaël Chen, Sylvain Lamprier, and Patrick Gallinari

[Preprint] [GitHub]

Presented at ICML 2022.

Raising context awareness in motion forecasting

We find that state-of-the-art forecasting methods tend to overly rely on the agent's dynamics and fail to exploit the context cues provided at its input. We fix this by introducing a forecasting model equipped with a training procedure designed to promote the use of semantic contextual information.

Hédi Ben-Younes*, Éloi Zablocki*, Mickaël Chen, Patrick Pérez, Matthieu Cord

[Preprint]

Presented at the CVPR WAD 2022.