# ReSSL: Relational Self-Supervised Learning with Weak Augmentation

**Authors:** Mingkai Zheng, Shan You, Fei Wang, Chen Qian, Changshui Zhang, Xiaogang Wang, Chang Xu  
**arXiv:** [2107.09282](https://arxiv.org/abs/2107.09282)  
**Date:** 2021-07-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2107.09282v1/

## Key points

- Self-supervised learning (SSL) is a popular approach for learning visual representations without data annotations.
- Most SSL methods focus on instance-level information, but lack attention on relationships between different instances.
- A novel SSL paradigm called Relational Self-Supervised Learning (ReSSL) framework has been introduced to address this gap.
- ReSSL learns representations by modeling the relationship between different instances and employs sharpened distribution of pairwise similarities among different instances as a "relation" metric to match feature embeddings of different augmentations.
- The proposed method argues that weak augmentations matter to represent a more reliable relation and leverages momentum strategy for practical efficiency.
- Experimental results show that ReSSL significantly outperforms previous state-of-the-art algorithms in terms of both performance and training efficiency.
- The authors have made their code available at https://github.com/KyleZheng1997/ReSSL.
- The paper titled "ReSSL: Relational Self-Supervised Learning with Weak Augmentation" highlights the importance of modeling relationships between different instances in SSL and proposes an effective solution to improve performance while maintaining training efficiency.

## Summary

Self-supervised learning (SSL) has become a popular approach for learning visual representations without the need for data annotations. However, most SSL methods focus on instance-level information, where different augmented images of the same instance should have the same feature or cluster into the same class. There is a lack of attention on the relationships between different instances. To address this gap, a novel SSL paradigm called Relational Self-Supervised Learning (ReSSL) framework has been introduced in this paper. ReSSL learns representations by modeling the relationship between different instances and employs sharpened distribution of pairwise similarities among different instances as a "relation" metric to match feature embeddings of different augmentations. The proposed method also argues that weak augmentations matter to represent a more reliable relation and leverages momentum strategy for practical efficiency. Experimental results show that ReSSL significantly outperforms previous state-of-the-art algorithms in terms of both performance and training efficiency. The authors have made their code available at https://github.com/KyleZheng1997/ReSSL. The authors of this paper are Mingkai Zheng, Shan You, Fei Wang, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Their paper titled "ReSSL: Relational Self-Supervised Learning with Weak Augmentation" highlights the importance of modeling relationships between different instances in SSL and proposes an effective solution to improve performance while maintaining training efficiency.