Optimizing contrastive learning through large-scale coverage for difficult sample handling

The handling of difficult samples remains a critical challenge in contrastive learning, hindering its broader application and development. To address this issue, we introduce Large-scale Coverage (LSC), a novel plug-and-play method designed to optimize and mitigate the impact of difficult samples. Through theoretical analysis, we demonstrate that adjusting the initial distribution of samples in the feature space can effectively alleviate the problem of difficult samples. The LSC method achieves this by modifying the similarity between sample pairs during the initial training phase, resulting in a more uniform distribution of samples. Experiments conducted on the CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets show that LSC, as an independent module, enhances the linear classification accuracy of SimCLR, MoCo V1, and V2 by 0.4% to 2.2% with minimal computational overhead. Moreover, pre-training on the ImageNet1k dataset reveals that LSC significantly improves performance in downstream tasks such as linear classification, object detection, and instance segmentation. Our method not only optimizes contrastive learning for difficult samples but also demonstrates its broad applicability and effectiveness. Source code link: https://github.com/PPChuan/Large-scaleCoverage.