[Paper Review] Semi-Supervised Learning with Generative Adversarial Networks

Augustus Odena|arXiv (Cornell University)|Jun 5, 2016

Generative Adversarial Networks and Image Synthesis9 references542 citations

TL;DR

Extends GANs to semi-supervised learning by making the discriminator output N+1 classes (N real classes plus a fake class), enabling joint training of a classifier with a generator and yielding data-efficient classification and better sample quality.

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.

Motivation & Objective

• Motivate and develop a method to perform semi-supervised classification while training a generative model.
• Enable the discriminator to output class labels in addition to real/fake, creating a shared representation for G, D, and C.
• Demonstrate data-efficiency gains and improved sample quality from joint G-D-C training on MNIST.

Proposed method

• Augment the GAN discriminator to have N+1 outputs corresponding to N real classes plus a FAKE class.
• Train D/C to minimize negative log likelihood for observed labels on data and FAKE label on generated samples; train G to maximize D/C’s accuracy.
• Use a single network architecture where weights are shared between discrimination and classification tasks (D acts as C).
• Optionally compare SGAN with a baseline classifier trained without updating the generator.
• Evaluate generative quality and classification accuracy on MNIST under varying data availability.

Experimental results

Research questions

• RQ1 Can a GAN discriminator be extended to output explicit class labels for semi-supervised learning?
• RQ2 Does sharing weights between discrimination and classification improve data efficiency for MNIST with limited labeled data?
• RQ3 Does joint G-D-C training yield higher quality samples and faster generator training than a standard GAN?
• RQ4 How does SGAN performance compare when using true class labels versus only REAL/FAKE labels?
• RQ5 What are the practical benefits and limitations of SGAN as a framework for semi-supervised learning?

Key findings

• SGAN outputs are significantly clearer than standard GAN samples on MNIST across initializations.
• SGAN outperforms a baseline classifier trained without generator updates as the labeled data is reduced, indicating improved data efficiency.
• Under restricted training data, SGAN maintains higher classification accuracy than the baseline.
• The classifier accuracy examples show SGAN achieving higher accuracy than the CNN baseline at smaller data sizes.