Florian Bordes

Background

About Florian Bordes

I did my PhD at Mila - Quebec artificial intelligence institute under the supervision of Pascal Vincent. I am currently a Research Scientist at Meta. One of my research interest focus on Generative models. In 2017, we presented the paper "Learning to Generate Samples from Noise through Infusion Training", Bordes et al. at ICLR in which we were the first to show that diffusion-based approaches were competitive with GANs. Then, I worked on conditional generative based diffusion models "High fidelity visualization of what your self-supervised representation knows about" Bordes et al, TMLR 2022. In this paper we show that SSL representation are well suited to be used as conditioning for diffusion based models (a similar technique was then used later by DALL.E-2). The second line of research I am interested about is Self-Supervised Learning. I co-authored "Masked Siamese Networks for Label-Efficient Learning", Assran et al. ECCV 2022,"The hidden uniform cluster prior in self-supervised learning", Assran et al. ICLR 2023 and "Guillotine Regularization: Why removing layers is needed to improve generalization in Self-Supervised Learning" Bordes et al. TMLR 2023. I also co-authored "A Cookbook of Self-Supervised Learning", Balestriero et al. 2023, with my colleagues at FAIR and several academic collaborators. I am also interest in privacy and have worked on memorization in SSL "Do SSL Models Have Déjà Vu? A Case of Unintended Memorization in Self-supervised Learning", Meehan et al. Lastly, I am interest more globally on how to better evaluate vision and vision language models. For this purpose, I leveraged video game engine to create datasets that offer fine-grained control over the factors of variations with "PUG: Photorealistic and Semantically Controllable Synthetic Data for Representation Learning" Bordes et al 2023. You can take a look at the project website there: https://pug.metademolab.com/ I also made significant contribution in term of softwares- With FFCV-SSL, I increased by 6 the speed of training SSL models https://github.com/facebookresearch/FFCV-SSL - With RCDM, I significantly improved the interpretability of SSL representations: https://github.com/facebookresearch/RCDM - With PUG, I release the software the enable a webrtc connection between Unreal Engine and a python client for fine-grained controlled over the factor of variation https://github.com/facebookresearch/PUG