[Paper Review] Unsupervised Learning of Visual Features by Contrasting Cluster Assignments

Mathilde Caron, Ishan Misra+4 more|arXiv (Cornell University)|Jun 17, 2020

Image Enhancement Techniques61 references1,908 citations

TL;DR

The paper introduces SwAV, an online clustering-based self-supervised method that predicts cluster codes across multiple views of the same image, achieving strong ImageNet results and transfer performance without large memory banks or momentum encoders. It also introduces multi-crop augmentation to increase views efficiently.

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.

Motivation & Objective

• Develop an online, scalable self-supervised method that learns visual representations without labels.
• Eliminate the need for pairwise feature comparisons by contrasting cluster assignments across views.
• Improve data augmentation via a multi-crop strategy to increase view count without extra cost.
• Demonstrate strong transfer to downstream vision tasks and robustness to batch size constraints.

Proposed method

• Compute online cluster assignments by mapping image features to K prototypes and enforcing a swapped prediction between two augmented views of the same image.
• Optimize a joint loss that swaps codes between views: L(z_t, q_s) + L(z_s, q_t) using a softmax over prototype similarities with temperature tau.
• Solve for soft assignments Q via an online Sinkhorn-Knopp optimization with minibatch constraints to ensure equipartition across prototypes.
• Represent features on the unit sphere and learn prototypes C jointly with the encoder f_theta through backpropagation.
• Introduce multi-crop augmentation to generate multiple views, including V low-resolution crops, while computing codes only for full-resolution crops.
• Operate in online fashion with minibatch-based codes and a small feature queue when batch sizes are limited, avoiding large memory banks and momentum encoders.

Experimental results

Research questions

• RQ1 Can online clustering with swapped assignments match or exceed contrastive methods for self-supervised visual representation learning?
• RQ2 Does a multi-crop augmentation strategy improve performance without increasing memory or compute requirements?
• RQ3 How do online prototype-based codes transfer to downstream tasks compared to supervised pretraining?
• RQ4 IsSwAV effective with small batches and without momentum encoders or large memory banks?
• RQ5 What is the impact of training duration and data scale (curated vs uncurated) on unsupervised pretraining quality?

Key findings

• SwAV achieves 75.3% top-1 accuracy on ImageNet with ResNet-50 in linear evaluation.
• SwAV outperforms prior self-supervised methods on transfer tasks, surpassing supervised ImageNet pretraining on several datasets.
• The method remains strong with small batches and does not require a momentum encoder or large memory bank.
• The multi-crop augmentation yields consistent 2–4% gains across multiple self-supervised methods.
• Combining online clustering with multi-crop and larger architectures narrows or surpasses the gap to supervised pretraining on several downstream tasks.
• Pretraining on uncurated data with SwAV yields gains over random initialization and is competitive with SimCLR, demonstrating data- and model-scale robustness.