CS7643 Quiz 5 - P2 EXAM
questions and verified correct
answers 100% guaranteed pass
latest 2026-2027
2. A discriminator takes p(x) and a series of fake images and attempts to determine if it's real or
fake
Generator: Update weights to improve realism
Discriminator: Update weights to better discriminate
Compete against each other in a minimax game
Trained with minibatch gradient descent.
Gradient Properties of Generator for min-max - CORRECT ANSWER-- Not very good
- High gradient when D(G(z)) is high (discriminator is wrong)
- We want it to improve when samples are bad (discriminator is right)
Generator is minimizing
Discriminator is maximizing
min_G max_D V(D,G) = E[log D(x)] + E[log(1 - D(G(z))]
, Note this does change after fixing the gradient issue
The discriminator is maximizing this objective by pushing D(x) to 1 and D(G(z)) to 0, the
generator is minimizing it by pushing D(G(Z)) to 1.
Gradient Fix for GAN Minimax - CORRECT ANSWER-High gradient when D(G(z)) is high
(discriminator is wrong) and the generator is not efficient in updating the gradient.
Instead maximize E[log(D(G(z)))] (or minimize (1-D(G(z)))
Now when D(G(z)) is close to 0, the gradient will be near infinity, and when D(G(z)) is close to 1,
the gradient will be near 0
Difficulty in GANs - CORRECT ANSWER-- minimax object is difficult to optimize
Advancements:
1. More stable architectures
2. Regularization methods to improve optimization
3. Progressive growing/training and scaling
DCGAN Paper - CORRECT ANSWER-Conclusion: Some architectures are better for training GANS
1. Replace any pooling layers with striped convolutions (discriminator) and fractional-striped
convolutions (generator)
2. Use BatchNorm in both the generator and discriminator
3. Remove fully connected hidden layers for deeper architectures
4. Use ReLU activation in generator for all layers except the output which uses Tanh
questions and verified correct
answers 100% guaranteed pass
latest 2026-2027
2. A discriminator takes p(x) and a series of fake images and attempts to determine if it's real or
fake
Generator: Update weights to improve realism
Discriminator: Update weights to better discriminate
Compete against each other in a minimax game
Trained with minibatch gradient descent.
Gradient Properties of Generator for min-max - CORRECT ANSWER-- Not very good
- High gradient when D(G(z)) is high (discriminator is wrong)
- We want it to improve when samples are bad (discriminator is right)
Generator is minimizing
Discriminator is maximizing
min_G max_D V(D,G) = E[log D(x)] + E[log(1 - D(G(z))]
, Note this does change after fixing the gradient issue
The discriminator is maximizing this objective by pushing D(x) to 1 and D(G(z)) to 0, the
generator is minimizing it by pushing D(G(Z)) to 1.
Gradient Fix for GAN Minimax - CORRECT ANSWER-High gradient when D(G(z)) is high
(discriminator is wrong) and the generator is not efficient in updating the gradient.
Instead maximize E[log(D(G(z)))] (or minimize (1-D(G(z)))
Now when D(G(z)) is close to 0, the gradient will be near infinity, and when D(G(z)) is close to 1,
the gradient will be near 0
Difficulty in GANs - CORRECT ANSWER-- minimax object is difficult to optimize
Advancements:
1. More stable architectures
2. Regularization methods to improve optimization
3. Progressive growing/training and scaling
DCGAN Paper - CORRECT ANSWER-Conclusion: Some architectures are better for training GANS
1. Replace any pooling layers with striped convolutions (discriminator) and fractional-striped
convolutions (generator)
2. Use BatchNorm in both the generator and discriminator
3. Remove fully connected hidden layers for deeper architectures
4. Use ReLU activation in generator for all layers except the output which uses Tanh
