Self-supervised learning (SSL) learns high-quality representations from large pools of unlabeled training data. As datasets grow larger, it becomes crucial to identify the examples that contribute the most to learning such representations. This enables efficient SSL by reducing the volume of data required for learning high-quality representations. Nevertheless, quantifying the value of examples for SSL has remained an open question. In this work, we address this for the first time, by proving that examples that contribute the most to contrastive SSL are those that have the most similar augmentations to other examples, in expectation. We provide rigorous guarantees for the generalization performance of SSL on such subsets. Empirically, we discover, perhaps surprisingly, the subsets that contribute the most to SSL are those that contribute the least to supervised learning. Through extensive experiments, we show we can safely exclude 20% of examples from CIFAR100 and 40% from STL10 and TinyImageNet, without affecting downstream task performance. We also show that our subsets outperform random subsets by more than 2% on CIFAR10. We also demonstrate that these subsets are effective across contrastive SSL methods (evaluated on SimCLR, MoCo, SimSiam, BYOL).

Highlights

New Problem


Previous work has studied data-efficiency for supervised learning and shown that nearly 30% of examples can be discarded on many datasets without sacrificing accuracy. However, this problem has not been studied for self-supervised learning. Our paper proposes the first empirical and theoretical method for data-efficient contrastive learning. We show that across various datasets (CIFAR100, STL10 and TinyImageNet), our method allows you to discard 20-40% of examples without affecting downstream classification accuracy.

Theoretical Guarantees for Downstream Accuracy


We use [Huang et al. 2023] theoretical characterization of contrastive learning using alignment and divergence of class centers to show that the subset we choose preserves alignment and divergence of class centers and thus preserves downstream accuracy.

BibTeX

@article{joshi2023dataefficient,
    title={Data-Efficient Contrastive Self-supervised Learning: Most Beneficial Examples for Supervised Learning Contribute the Least},
    author={Joshi, Siddharth and Mirzasoleiman, Baharan},
    journal={International Conference on Machine Learning (ICML)},
    year={2023}
}

Code