Victor Letzelter

1/ 📢 New preprint: Test-Time Conditioning with Representation-Aligned Visual Features Introducing REPA-G — a framework for controllable image generation at inference time using aligned visual features. 📄 arxiv.org/pdf/2602.03753 💻 github.com/valeoai/REPA-G 🧵👇

Let’s catch up if you’re at @icmlconf and want to chat!

🚗 Ever wondered if an AI model could learn to drive just by watching YouTube? 🎥👀 We trained a 1.2B parameter model on 1,800+ hours of raw driving videos. No labels. No maps. Just pure observation. And it works! 🤯 🧵👇 [1/10]

Inferring 3D human poses from video is highly ill-posed because of depth ambiguity. Our work accepted to #NeurIPS2024, ManiPose, gets one step closer to solving this, by leveraging prior knowledge about poses topology and cool multiple-choice learning techniques.

🌟 Calling all MSc students passionate about computer vision and ML! We’re offering research internships about diffusion models, multi-modal transformers, continual learning, & more. 4 exciting openings await! 🔗 Learn more: valeoai.github.io/interns/ RT to spread the word! 🙌

Check out the paper! Arxiv: arxiv.org/abs/2407.15580 Project page: valeoai.github.io/publications/a… GitHub: github.com/Victorletzelte… A joint work with David Perera, Théo Mariotte, Adrien Cortes, @Mickael_Chen , Slim Essid and Gaël Richard at @valeoai and Télécom Paris.

Key insights: * Annealing enhances exploration compared to greedy convergence. * Inspired by statistical physics and information theory, we describe the training trajectory. * Experiments on synthetic datasets, UCI benchmarks, and speech separation show highly promising results.

Working on ill-posed machine learning tasks, interested in multi-heads neural networks and data #uncertainty quantification ? Sharing here our latest research, which will be presented at @NeurIPSConf in December.

You have an already trained sem. segmentation model ? You want to apply it to data with a domain shift ? You are afraid of degradation during the adaptation ? Then you might want to check out our work TTYD at @eccvconf 2024 in Milan. 👉 Poster # 73: Tuesday, 16:30 #ECCV2024

Glad to announce that Stem-JEPA has been accepted to @ISMIRConf ! In this work, we tackle the task of musical stem compatibility estimation (what “fits” together) as a representation learning problem. (1/7) Paper: arxiv.org/abs/2408.02514 Code: github.com/SonyCSLParis/S…

.@VLetzelter will be at #ICML2024 to present his work on leveraging the geometric properties of the Winner-takes-all learners for conditional density estimation & uncertainty prediction, w/o modifying its original training scheme. Find a tl;dr below & come say hi at the posters

Check out our paper at arxiv.org/abs/2406.04706 A joint work with D.Perera, @ccrommel, @MathieuFontai19, S.Essid, G.Richard and @ptrkprz Multi-head #neuralnetworks #machinelearning #Uncertainty #DensityEstimation

The attached figure illustrates the predictions (shaded blue points) made by WTA-based models compared to other baselines, for the task of estimating conditional distributions on a synthetic datasets (represented by green points). [4/n]

Interested in ill-posed learning tasks, uncertainty prediction, conditional density estimation or multi-head deep neural networks ? In our new paper, accepted at #ICML24, we tackle these challenges by exploring the Winner-Takes-All (WTA) training scheme. [1/n]

Interested in unsupervised domain adaptation for LiDAR data ? Then, I am happy to see you at the spotlight presentation of our paper “SALUDA” this afternoon at 15:00 or at the poster session @3DVconf. Paper: arxiv.org/abs/2304.03251 Code: github.com/valeoai/SALUDA