GANs Unplugged

Authors

  • Patrick Virtue Carnegie Mellon University

Keywords:

Generative Adversarial Networks, AI Education, K-12 AI Education, STEM Education

Abstract

With the influx of deepfake and style transfer technology in today's news and social media, everyone is told that these applications are powered by artificial intelligence and deep learning, but too often the explanation of how it works goes no further. Rather than waiting until second semester of computer science grad school to learn about generative adversarial networks (GANs), we propose a classroom activity to introduce GANs to secondary school students. Our GANs Unplugged activity steps outside a traditional classroom environment and forms groups of students that physically act as the various components of a GAN. Students in the Real and Generator (i.e. Fake) groups both draw sketches of animals on index cards, but only the Real group has the secret, specific instructions on how to draw their animals. A third Discriminator group tries to determine if each card it receives came from the Real group or the Fake group. The Discriminator group also makes a single feedback mark on the drawing to suggest how that drawing could be more "real" before passing the card back to the group that drew it. As this game progresses, the Discriminators get better at determining real and fake, that is until the Generators adapt and learn to draw more and more realistic "fake" images. We describe the details of GANs Unplugged and present our experience running this activity at three different artificial intelligence summer camps for high school students.

Downloads

Published

2021-05-18

How to Cite

Virtue, P. (2021). GANs Unplugged. Proceedings of the AAAI Conference on Artificial Intelligence, 35(17), 15664-15668. Retrieved from https://ojs.aaai.org/index.php/AAAI/article/view/17845