UC San Diego

Generative Adversarial Neural Networks are neural networks which participate in a zero- sum game, competing against each other to maximize an objective. One network, the generator, hopes to generate data that lies in a similar distribution to given data. The discriminator aims to separate data generated by the generator and ground truth data. This allows us to generate and replicate data given a dataset.

These networks have been increasingly popular in generating images from text. By leveraging the Introspective Learning framework, we are able to take image classification networks and synthesize images. We show that our results are competitive on many-to-many mappings against Conditional Generative Adversarial Neural Networks.