https://piazza.com/class_profile/get_resource/lr49jt48q953s2/lt4wcfg6cjtl1

$$ \min_G \max_D \mathbb{E}{x \sim p{\text{data}}(x)}[\log D(x)] + \mathbb{E}_{z \sim p_z(z)}[\log(1 - D(G(z)))] $$

where G represents the generator and D is the discriminator (hence adverserial) and E is the expectation

Generator is trying to fool discriminator

• does not train well because when it generates bad examples, it has least preferable gradients

Solutions:

• gradient ascent on the discriminator output

• “Smooth” real-valued discriminators: LSGAN, WGAN, WGAN-GP, spectral norm

• instance noise

• backprop from the discriminator into the generator

  

The Gan Game Summarized

• Trying to match the data distribution (ie the pixel distribution)
• Thus goal: maximize P(image = True) / minimize P(image = False)

Question: How are images labeled True/ False in GANS?

Summary