---

toc: true title: BYOL Loss

tags: [‘temp’]

---

BYOL Loss

• Similarity loss between $q_{\theta }(z_{\theta })$ and $sg(z_{\xi }^{{}^{\prime }})$
• $\theta$ is trained weights
• $\xi$ is exponentially moving average of $\theta$ and sg is stop gradient
• $f_{\theta }$ is discarded, $y_{\theta }$ is used as image representation
•