Emerging Properties in Self-Supervised Vision Transformers

Mathilde Caron (1, 2) , Hugo Touvron (1, 3) , Ishan Misra (1) , Hervé Jegou (1) , Julien Mairal (2) , Piotr Bojanowski (1) , Armand Joulin (1)

1 FAIR - Facebook AI Research [Paris]
2 Thoth - Apprentissage de modèles à partir de données massives
3 SU - Sorbonne Université

In this paper, we question if self-supervised learning provides new properties to Vision Transformer (ViT) that stand out compared to convolutional networks (convnets). Beyond the fact that adapting self-supervised methods to this architecture works particularly well, we make the following observations: first, self-supervised ViT features contain explicit information about the semantic segmentation of an image, which does not emerge as clearly with supervised ViTs, nor with convnets. Second, these features are also excellent k-NN classifiers, reaching 78.3% top-1 on ImageNet with a small ViT. Our study also underlines the importance of momentum encoder, multi-crop training, and the use of small patches with ViTs. We implement our findings into a simple self-supervised method, called DINO, which we interpret as a form of self-distillation with no labels. We show the synergy between DINO and ViTs by achieving 80.1% top-1 on ImageNet in linear evaluation with ViT-Base.

Domaines: Informatique [cs]

Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jegou, Julien Mairal, et al.. Emerging Properties in Self-Supervised Vision Transformers. ICCV 2021 - International Conference on Computer Vision, Oct 2021, Virtual, France. pp.1-21. ⟨hal-03323359⟩