GAN(Generative Adversarial Networks)

paper: https://arxiv.org/pdf/1406.2661
cover: いずもねる
I’m so sad :<

Adversarial nets

was defined to learn over data , then presenting a mapping to ; Second mlp outputs a single scalar. means the probability that came from the data rather than . The training, or the game between and is:

Train

In the earlier training, the objective function is to maximize , the result is the same but the gradients are stronger early in learning.

The algorithm

In the paper, the author explained the algorithm to optimize Eq1, but it is a bit deep for me. I may watch some videos explaining this later.

Experiment

Dataset: MNIST, Toronto Face Database and CIFAR-10.

The generator nets used a mixture of rectifier linear activations and sigmoid activations.
The discriminator net used maxout activations.
Dropout was applied in training the discriminator net.
Estimate the probability with Gaussian Parzen window.