Artificial Intelligence By Example

Generative adversarial networks

Using GANs to generate concept-word images can boost the production CRL datasets and produce millions of concept-word images. No single solution will suffice. CNNs can do the job; GANs can also.

Suppose the Γ gap concept needs to be generalized to thousands of situations. Millions of images with words could be analyzed and automatically stored in a Γ dataset, with tens of thousands of image-word concepts of gaps.

A GAN could be built in a few lines with the following components:

• Input data for the generator
• The generator
• Real data input
• A discriminator (detects fake data from real data)
• An optimizer (training)

The network starts by inputting real data and fake data (with the generator), as shown in the following ...