independent-research

Home	Independent Research

Index

Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, and Ian Goodfellow. Adversarial Autoencoders. arXiv preprint arXiv:1511.05644, 2015.

Motivation

The paper proposed a new inference algorithm Adversarial Autoencoder (AAE), which uses GAN framework which could better deal with “applications such as semi-supervised classification, disentangling style and content of images, unsupervised clustering, dimensionality reduction and data visualization”.

Method

The algorithm aims to find a representation for graphs which follows a certain type of distribution. And it consists of two phases: Reconstruction phase and Regularization phase.

Reconstruction phase: Update the encoder and decoder to minimize reconstruction error.

Regularization phase: Update discriminator to distinguish true prior samples from generated samples; update generator to fool the discriminator

Reconstruction phase and regularization phase are referred to the generator and discriminator in GANs.

And the method could be used in semi-supervised learning and unsupervised clustering.

For semi-supervised learning, there is a semi-supervised classification phase besides reconstruction phase and regularization phase. And labeled data would be trained at this stage.

which is a aggregated categorical distribution.

The architecture of unsupervised clustering is similar to semi-supervised learning, the difference is that the semi-supervised classification stage is removed and “thus no longer train the network on any labeled mini-batch”.