Deep Learning
What is the function of the generator in a GAN?
It creates new examples.
Card 1 of 259. Answer: It creates new examples.
It creates new examples.
Card 1 of 259. Answer: It creates new examples.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
It creates new, synthetic examples. The generator is a neural network G that transforms a random noise vector z, usually Gaussian or uniform, into a sample G(z), such as an image. It never sees the training data; it learns only through the discriminator's gradients.
In DCGAN (2015), z is 100 random numbers, projected into a 4 × 4 stack of feature maps and upsampled by transposed convolutions into a 64 × 64 color image, with a tanh output for pixels scaled to [−1, 1]. The original objective, minimizing log(1 − D(G(z))), stalls when D easily rejects fakes: at D(G(z)) = 0.01 its gradient with respect to D's logit is only 0.01, against 0.99 for maximizing log D(G(z)), the non-saturating loss the original paper itself recommends.
After training, the discriminator is usually discarded and the generator is the product: each new z gives a new sample. Follow-ups: mode collapse, where many z map to near-identical outputs, and conditional GANs, which also feed G a class label.