- Keywords: Unsupervised embedding learning, computer vision, anchor neighborhood discovery, image clustering
- TL;DR: We proposed a comprehensive approach for unsupervised embedding learning on the basis of AND algorithm.
- Abstract: Unsupervised embedding learning aims to extract good representations from data without the use of human-annotated labels. Such techniques are apparently in the limelight because of the challenges in collecting massive-scale labels required for supervised learning. This paper proposes a comprehensive approach, called Super-AND, which is based on the Anchor Neighbourhood Discovery model. Multiple losses defined in Super-AND make similar samples gather even within a low-density space and keep features invariant against augmentation. As a result, our model outperforms existing approaches in various benchmark datasets and achieves an accuracy of 89.2% in CIFAR-10 with the Resnet18 backbone network, a 2.9% gain over the state-of-the-art.
- Code: https://github.com/super-AND/super-AND