InfoGAN – Extension to Generative Adversarial NNs

InfoGAN learned to distinguish between two subjects and create realistic computer-generated representations of them.  A continuous code captured their rotation separately (top).  Below, 4 different categorical codes captured the subjects separately, and as well as wearing sunglasses, one of several features of the training dataset.

Read our original research paper!

All images shown are completely computer generated

Download Paper PDF

Executive summary:

InfoGAN, without any supervision, can extract categorical information (discrete properties like identifying people, the authors of paintings, features) and continuous information (continual properties like rotation, color gradients, stylistic traits, incremental features) from images.  This is a generative system, so all images shown are computer-generated images from the representation it learns, which has many useful applications.  One could imagine a final evolution of this as like a video-game avatar generator.  Create a photo-realistic version of any person with buttons and sliders for each of their traits; the same is possible for paintings, doodles, etc.

Fantastic extended results from our paper. Demonstrates the useful new knowledge uniquely available in this generative, unsupervised system.

This poster should explain more clearly the purpose and context of our research:

InfoGAN Poster

Please feel free to reach out with questions or comments regarding our work! I love to hear from people interested in this stuff.

This research was conducted as part of Swarthmore Adaptive Robotics (CS 81) course.

%d bloggers like this:
search previous next tag category expand menu location phone mail time cart zoom edit close