MaskCon: Masked Contrastive Learning for Coarse-Labelled Dataset

Deep learning has achieved great success in recent years with the aid ofadvanced neural network structures and large-scale human-annotated datasets.However, it is often costly and difficult to accurately and efficientlyannotate large-scale datasets, especially for some specialized domains wherefine-grained labels are required. In this setting, coarse labels are mucheasier to acquire as they do not require expert knowledge. In this work, wepropose a contrastive learning method, called $\textbf{Mask}$ed$\textbf{Con}$trastive learning~($\textbf{MaskCon}$) to address theunder-explored problem setting, where we learn with a coarse-labelled datasetin order to address a finer labelling problem. More specifically, within thecontrastive learning framework, for each sample our method generatessoft-labels with the aid of coarse labels against other samples and anotheraugmented view of the sample in question. By contrast to self-supervisedcontrastive learning where only the sample's augmentations are considered hardpositives, and in supervised contrastive learning where only samples with thesame coarse labels are considered hard positives, we propose soft labels basedon sample distances, that are masked by the coarse labels. This allows us toutilize both inter-sample relations and coarse labels. We demonstrate that ourmethod can obtain as special cases many existing state-of-the-art works andthat it provides tighter bounds on the generalization error. Experimentally,our method achieves significant improvement over the current state-of-the-artin various datasets, including CIFAR10, CIFAR100, ImageNet-1K, Standford OnlineProducts and Stanford Cars196 datasets. Code and annotations are available athttps://github.com/MrChenFeng/MaskCon_CVPR2023.