week 6 - chatgpt Flashcards
What are the two components of a Generative Adversarial Network (GAN)?
The Generator (G) which creates fake data, and the Discriminator (D) which tries to distinguish real from fake data.
What is the main objective of a GAN during training?
To train G to generate data that fools D, while training D to distinguish real data from G’s outputs — forming a minimax game.
What is the typical loss function used for GANs?
D maximizes log(D(x)) + log(1 − D(G(z)); G minimizes log(1 − D(G(z))) or equivalently maximizes log(D(G(z))).
What is mode collapse in GANs?
A failure where the generator produces limited variety, mapping many inputs to the same output.
What is the vanishing gradient problem in GANs?
If D becomes too good, G’s gradient becomes very small, preventing effective learning.
How does DCGAN improve the stability of GAN training?
By using strided convolutions, batch normalization, ReLU in the generator, and Leaky ReLU in the discriminator.
What is experience replay and how does it help GANs?
It stores past generated samples and reuses them during training to stabilize the generator and avoid mode collapse.
What is unrolling in GAN training?
It allows the generator to anticipate the discriminator’s updates, helping prevent mode collapse by smoothing training dynamics.
Why is GAN training often unstable?
Because it involves a two-player minimax game where G and D are competing, not minimizing a single convex loss.
What are some applications of GANs?
Image generation, style transfer, super-resolution, text-to-image generation, video prediction, and data augmentation.
