Notes.md

[Title]

CR-GAN

---

Consistency regularization for generative adversarial networks
[ICLR 2020] (Google) [Code]
Tero Karras, Samuli Laine, Timo Aila

Click to expand

Summary

They propose a training stabilizer based on consistency regularization. In particular, they augment data passing into the GAN discriminator and penalize the sensitivity of the discriminator to these augmentations.

Details

$T(x)$ donates a stochastic data augmentation function. $D(x)$ donates the last layer before the activation function. The proposed regularization is given by:

Latex $$ \operatorname{argmin}{\theta} \mathcal{L}(\theta)=\mathbb{E}{\mathbf{z}, \mathbf{y}, \alpha}\left[\left(A\left(G\left(T_{\theta}(\mathbf{z}, \alpha), \mathbf{y}\right)\right)-(A(G(\mathbf{z}, \mathbf{y}))+\alpha)\right)^{2}\right] $$

---