Generative Adversarial Nets

Inside the paper

The problem

Deep generative models often required difficult likelihood calculations or sampling procedures during training. The authors sought a way to learn to generate examples using ordinary neural network optimization.

The idea

Two networks play opposing roles: one makes samples and the other tries to tell them from training data. The generator learns from the discriminator’s feedback instead of evaluating an explicit probability for every generated example.

How it works

• The generator transforms random input noise into a candidate example, while the discriminator outputs an estimate of whether an example came from the data.
• Training alternates discriminator updates on real and generated examples with generator updates that try to make generated examples harder to reject.
• The authors use a modified generator objective in practice to give stronger gradients when the discriminator initially succeeds too easily.
• To generate after training, draw fresh noise and pass it through the generator; no Markov chain is needed for that sampling step.

What it showed

The theory identifies the matching data distribution as the optimum under idealized assumptions, but does not guarantee that finite neural networks reach it. Experiments show generated digits, faces and small natural images, and compare approximate likelihood estimates with other models.

Read it for

Begin with Figure 1, Section 3 and Algorithm 1 to understand the game, then read Section 4 for the assumptions behind the proof. Section 5 and Figure 2 show what was actually tested; the related-work survey can be skimmed initially.

Written in our own words from the paper itself. A draft the editors have not yet reviewed.