toc: true title: BYOL
tags: [‘temp’]
BYOL
- Bootstrap Your Own Latent: a New Approach to Self-supervised Learning
- Self Supervised image representation learning
- predicting previous versions of its outputs, without using negative pairs
- two neural networks, referred to as online and target networks
- that interact and learn from each other
- From an augmented view of an image, they train the Online Learning network to predict the target network representation of the same image under a different augmented view
- update the target network with a slow-moving average of the online network
- ImageNet
- Res Net
- dependent on existing sets of Augmentation that are specific to vision applications
- BYOL Loss