Conditional Generative Adversarial Nets: Turn Labels Into Pictures

This article discusses Conditional Generative Adversarial Nets (cGANs), a machine learning technique that pairs two programs - a generator and a discriminator - to create images that match a given label or condition.

Details

Conditional Generative Adversarial Nets (cGANs) are a machine learning technique that pairs two simple computer programs - a generator and a discriminator - that learn by competing with each other. The generator tries to create images, while the discriminator tries to identify whether an image is real or generated. By providing both programs with the same conditional hint, like a class name, they learn to create images that match that hint. This allows the generator to draw digits, scenes, or other content that fits the requested condition, while the discriminator helps ensure the results are realistic. cGANs have been used to generate clear examples of handwritten digits like MNIST, and can also suggest useful image tags or even invent new descriptive words. The process feels like directing a digital artist, giving instructions and watching it learn to match your words, even when the instructions are not perfect.