GAN Training — Generator vs Discriminator
min_G max_D E[log D(x)] + E[log(1-D(G(z)))]
Controls
Real data:
Gaussian
Gaussian
Bimodal
Uniform
Learning rate:
0.05
D steps / G step:
1
Start Training
Stop
Reset
Step:
0
G loss:
—
D loss:
—
D(G(z)):
—
G
maps noise z~N(0,1) → fake x.
D
distinguishes real vs fake.
Nash equilibrium: p_G = p_data, D≡0.5 everywhere.