#non-saturating

Question 1: Single choice

mediumAI-checkedGenerative Adversarial Networks

Early in GAN training the discriminator $D_\theta$ rejects the generator's samples with high confidence, $D_\theta(G_\phi(z)) \approx 0$ . Which generator update still gives a strong gradient in this situation?

Ascend $\nabla_\phi \frac{1}{m}\sum_i \log D_\theta(G_\phi(z^{(i)}))$ 1
Descend $\nabla_\phi \frac{1}{m}\sum_i \log\left(1 - D_\theta(G_\phi(z^{(i)}))\right)$ 2
Ascend $\nabla_\phi \frac{1}{m}\sum_i \left[\log D_\theta(x^{(i)}) + \log\left(1 - D_\theta(G_\phi(z^{(i)}))\right)\right]$ 3
Ascend $\nabla_\phi \frac{1}{m}\sum_i \log G_\phi(D_\theta(z^{(i)}))$ 4

Timed test