Semi-Supervised Learning with Generative Adversarial Networks

TL;DR

The paper tackles semi-supervised image classification by extending GANs to have the discriminator output N+1 labels (N real classes plus FAKE). The SGAN trains G and the discriminator-as-classifier (D/C) jointly, using semi-supervised labels to push real data into the N classes and fake data into the FAKE class. Experiments on MNIST show SGAN produces higher-quality samples than vanilla GANs and improves classifier performance when labeled data is scarce. The work highlights potential improvements via partial weight sharing, conditional labeling, and ladder-network-inspired approaches to leverage unlabeled G samples.

Abstract

We extend Generative Adversarial Networks (GANs) to the semi-supervised context by forcing the discriminator network to output class labels. We train a generative model G and a discriminator D on a dataset with inputs belonging to one of N classes. At training time, D is made to predict which of N+1 classes the input belongs to, where an extra class is added to correspond to the outputs of G. We show that this method can be used to create a more data-efficient classifier and that it allows for generating higher quality samples than a regular GAN.

Figures (1)

• Figure 1: Output samples from SGAN and GAN after 2 MNIST epochs. SGAN is on the left and GAN is on the right.