--- 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]]