When Does Contrastive Visual Representation Learning Work?
This work addresses the lack of best practices for replicating self-supervised learning success beyond ImageNet, providing insights for researchers and practitioners in computer vision.
The paper investigates the conditions for successful contrastive self-supervised learning on diverse datasets, finding that benefits from additional data beyond 500k images are modest, cross-domain pretraining does not improve generalization, and contrastive learning lags behind supervised methods on fine-grained tasks.
Recent self-supervised representation learning techniques have largely closed the gap between supervised and unsupervised learning on ImageNet classification. While the particulars of pretraining on ImageNet are now relatively well understood, the field still lacks widely accepted best practices for replicating this success on other datasets. As a first step in this direction, we study contrastive self-supervised learning on four diverse large-scale datasets. By looking through the lenses of data quantity, data domain, data quality, and task granularity, we provide new insights into the necessary conditions for successful self-supervised learning. Our key findings include observations such as: (i) the benefit of additional pretraining data beyond 500k images is modest, (ii) adding pretraining images from another domain does not lead to more general representations, (iii) corrupted pretraining images have a disparate impact on supervised and self-supervised pretraining, and (iv) contrastive learning lags far behind supervised learning on fine-grained visual classification tasks.