Rethinking Samples Selection for Contrastive Learning: Mining of Potential Samples

TL;DR

This paper proposes a comprehensive sample mining strategy for contrastive learning that considers both positive and negative samples, improving the quality of sample selection and enhancing model performance.

Contribution

It introduces a novel sample mining approach that combines data augmentation, data mining, and gradient analysis to select more informative positive and negative samples.

Findings

Achieves 88.57% top-1 accuracy on CIFAR10

Achieves 61.10% top-1 accuracy on CIFAR100

Achieves 36.69% top-1 accuracy on TinyImagenet

Abstract

Contrastive learning predicts whether two images belong to the same category by training a model to make their feature representations as close or as far away as possible. In this paper, we rethink how to mine samples in contrastive learning, unlike other methods, our approach is more comprehensive, taking into account both positive and negative samples, and mining potential samples from two aspects: First, for positive samples, we consider both the augmented sample views obtained by data augmentation and the mined sample views through data mining. Then, we weight and combine them using both soft and hard weighting strategies. Second, considering the existence of uninformative negative samples and false negative samples in the negative samples, we analyze the negative samples from the gradient perspective and finally mine negative samples that are neither too hard nor too easy as potential negative samples, i.e., those negative samples that are close to positive samples. The experiments show the obvious advantages of our method compared with some traditional self-supervised methods. Our method achieves 88.57%, 61.10%, and 36.69% top-1 accuracy on CIFAR10, CIFAR100, and TinyImagenet, respectively.