Generative Adversarial Networks
by Casandra Laskowski
Generative Adversarial Networks (GANs) gained some notoriety when an AI generated “painting” sold for $432,500. The painting, though interesting, looks computer generated and is unlikely to spark much concern. However, some GANs can produce images that are hard to discern from real images. In just four years, face generating GANs have gone from producing fuzzy black and white faces to photorealistic images. This should spark much more concern, especially as we teach information literacy to our respective patrons.
First created by Ian Goodfellow, Generative Adversarial Networks are made up of deep learning neural networks, a generator and a discriminator, that are pitted against each other to facilitate faster pattern recognition and more realistic results. Both of the networks reference the same training data set though they use it for different purposes. The generator uses the data to attempt to create similar images while the discriminator will compare those generated images with the data set to decide if it is real or fake. The goal is for the generator to create an image that is realistic enough to fool the discriminator. In the case of the painting, that was a very profitable goal.
However, GANs can also be damaging as they also facilitate the creation of deep fakes, manipulated and misleading doctored videos. One popular and somewhat ironic deepfake video has comedian Jordan Peele speaking through an AI manipulated video of Obama warning about this type of technology. If you know what to look for, you can spot the markers that reveal the video as fake, but the technology is improving at a remarkable speed.
As with all technology, GANs have good applications. GANs are being used to improve graphics in older video games. AI researcher David Kale is looking to use GANs to produce a training data set of fake medical records that other healthcare researchers could use to train their AI programs. Another team of researchers is using GANs to create “accurate, reliable synthetic images of abnormal brain MRIs to train an AI system.” GANs have potential for good and ill. For this reason, we have to be aware and we have to make our patrons aware of what is possible and how to leverage information literacy skills to identify questionable content.
Copyright 2018 by Casandra Laskowski.
About the author: Casandra Laskowski is a Reference Librarian and Lecturing Fellow at Duke Law. She received her J.D. from the University of Maryland School of Law, and her M.L.I.S. from the University of Arizona. Prior to pursuing her career as a law librarian, she worked as a geospatial analyst in the United States Army and served a fifteen-month tour of duty in Iraq. Her areas of interest include privacy, censorship, and the intersection of national security and individual liberty.