Emerging Properties in Self-Supervised Vision Transformers

TL;DR

This paper investigates unique properties of self-supervised Vision Transformers, revealing their superior semantic segmentation features and classification performance, and introduces DINO, a simple self-supervised method that enhances ViT performance.

Contribution

The paper uncovers that self-supervised ViTs encode explicit semantic information and introduces DINO, a novel self-distillation approach that improves ViT performance on ImageNet.

Findings

Self-supervised ViTs contain explicit semantic segmentation information.

Self-supervised ViTs achieve 78.3% top-1 accuracy with k-NN on ImageNet.

DINO improves ViT-Base linear evaluation to 80.1% top-1 accuracy.

Abstract

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.