Even a single simple augmentation with Self-Supervised Learning can be helpful for the downstream tasks
Keywords: Self-supervised learning, Computer Vision
TL;DR: The focus of the paper is to explore the possibility of self-supervised learning to improve model performance on specific data domains using various simple augmentations.
Abstract: This paper explores some unexpected capabilities of Self-Supervised Learning (SSL) and shows that even a single cutout augmentation for SSL can achieve better results in downstream tasks compared to traditional supervised approaches. These unusual properties of SSL can be used for further research in this area.
Submission Number: 2
Loading