2016 ICLR | Adversarial Autoencoders

Paper: https://arxiv.org/abs/1511.05644

Adversarial Autoencoders

Adversarial Autoencoder（AAE） Yes, it will GAN And autoencoder Combination , Modelled on the GAN To train the self encoder network . The first half of the project is a simple and typical AE structure , contain input layer,encoder layer, hidden layer, decoder layer , output layer.encoder The real distribution x Mapped to a hidden layer z, decoder then z Decode and restore to x.

AAE The characteristic of is in the hidden layer hidden layer The idea of confrontation is introduced to optimize the hidden layer z, Judging device discriminator Need to be in hidden layer Judge the real data and generator after sampling encoder The resulting false data . therefore discriminator The purpose of is to make q(z | x) Keep turning to p(z) near , Matching the aggregated posterior q(z) to the prior p(z）.

There are three of them q(z | x) The choice of , Including deterministic functions , Gaussian posterior and general approximate posterior , Where deterministic functions are used , its q（z） It is only related to the distribution of real data , The other two q（z） Except that it is related to the real distribution , It is also related to some randomness in these two ways .

advantage

take GAN Medium discriminator And autoencoder Combination , Get a powerful self encoder
You can train without using reparameterization , Better than VAE good
AAE The structure modification of can be used for many tasks , Excellent scalability