Unsupervised Learning of Visual Features by Contrasting Cluster Assignments

TLDR

Unsupervised image representations have narrowed the gap with supervised pretraining, especially through contrastive learning, yet these methods depend on many pairwise comparisons that are computationally demanding. This work introduces SwAV, an online algorithm that exploits contrastive ideas without performing pairwise feature comparisons. SwAV simultaneously clusters data and enforces consistency between cluster assignments of different augmentations via a swapped prediction mechanism, and it employs a multi‑crop augmentation strategy to reduce memory and compute costs. The method trains efficiently with both large and small batches, scales to unlimited data, achieves 75.3 % top‑1 accuracy on ImageNet with ResNet‑50, and surpasses supervised pretraining on all considered transfer tasks while being more memory‑efficient.

Abstract

Unsupervised image representations have significantly reduced the gap with supervised pretraining, notably with the recent achievements of contrastive learning methods. These contrastive methods typically work online and rely on a large number of explicit pairwise feature comparisons, which is computationally challenging. In this paper, we propose an online algorithm, SwAV, that takes advantage of contrastive methods without requiring to compute pairwise comparisons. Specifically, our method simultaneously clusters the data while enforcing consistency between cluster assignments produced for different augmentations (or views) of the same image, instead of comparing features directly as in contrastive learning. Simply put, we use a swapped prediction mechanism where we predict the cluster assignment of a view from the representation of another view. Our method can be trained with large and small batches and can scale to unlimited amounts of data. Compared to previous contrastive methods, our method is more memory efficient since it does not require a large memory bank or a special momentum network. In addition, we also propose a new data augmentation strategy, multi-crop, that uses a mix of views with different resolutions in place of two full-resolution views, without increasing the memory or compute requirements much. We validate our findings by achieving 75.3% top-1 accuracy on ImageNet with ResNet-50, as well as surpassing supervised pretraining on all the considered transfer tasks.