6  Why an AutoEncoder in a GAN?

In general the main advantage of AutoEncoders (AE) is the dimensionality reduction of its input. This compact representation then force the network to learn the salient features of the initial data.

However using an autoencoder in combination with a GAN has several additional benefits which in my opinion are more important:

1. Help Generator Training: The generator in a GAN learns to generate realistic samples by fooling the discriminator. The autoencoder acts as a pre-training mechanism for the generator since it won’t need to learn to generate discrete data but only salient features for the decoder.

2. Regularization: Incorporating an autoencoder as part of a GAN architecture can act as a regularization technique since it acts as an additional constraint for the generator by forcing it to generate samples that can be accurately reconstructed. This helps in avoiding mode collapse.

More on this topics later on in the GAN chapter.