Jigsaw Clustering for Unsupervised Visual Representation Learning
This work addresses the high computational cost of unsupervised learning methods for computer vision researchers, offering a more efficient alternative that is competitive with existing approaches.
The paper tackles the computational inefficiency of contrastive learning in unsupervised visual representation learning by proposing a jigsaw clustering pretext task that processes each training batch only once, reducing training cost. It achieves state-of-the-art results on ImageNet with a 2.6% improvement over previous single-batch methods and outperforms supervised models on CIFAR datasets by up to 4.1%.
Unsupervised representation learning with contrastive learning achieved great success. This line of methods duplicate each training batch to construct contrastive pairs, making each training batch and its augmented version forwarded simultaneously and leading to additional computation. We propose a new jigsaw clustering pretext task in this paper, which only needs to forward each training batch itself, and reduces the training cost. Our method makes use of information from both intra- and inter-images, and outperforms previous single-batch based ones by a large margin. It is even comparable to the contrastive learning methods when only half of training batches are used. Our method indicates that multiple batches during training are not necessary, and opens the door for future research of single-batch unsupervised methods. Our models trained on ImageNet datasets achieve state-of-the-art results with linear classification, outperforming previous single-batch methods by 2.6%. Models transferred to COCO datasets outperform MoCo v2 by 0.4% with only half of the training batches. Our pretrained models outperform supervised ImageNet pretrained models on CIFAR-10 and CIFAR-100 datasets by 0.9% and 4.1% respectively. Code is available at https://github.com/Jia-Research-Lab/JigsawClustering