---
toc: true
title: BYOL
tags: ['temp']
---
# BYOL
- [Bootstrap Your Own Latent: a New Approach to Self-supervised Learning](https://arxiv.org/abs/2006.07733)
- [[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.md|Augmentation]] that are specific to vision applications
- [[BYOL Loss]]