top of page
The Potential for General Adversarial Networks (GANs)
In a previous newsletter we discussed Nvidia's StyleGAN 2, a generative method of machine-facial synthesis. However, that is just one application of GAN technology. Today, we are going to take a broader look at the possibilities and risks associated with the general adversarial network.
Figure 1, Figure 2, Figure 3, Figure 4
The Potential for General Adversarial Networks (GANs): Text
The Potential for General Adversarial Networks (GANs): Pro Gallery
If you missed the previous newsletter , a general adversarial network is a system of machine learning that pits a generative neural network against a discriminative neural network. The generative network attempts to create items (photos, songs, videos, etc.) that appear authentic while the discriminative network attempts to identify whether the item is real or fake. Slowly, the generative network begins to create items that the discriminator cannot discern from real ones. This technology is already being applied to photo generation, gaming, and a host of other industries. Figures 1 to 3 provide a glimpse into the inner workings of the machine learning process. Each beginning and end photo are uncurated images generated with NVIDIA’s StyleGAN 2. The middle 7 images are StyleGans interpolation of the images that would provide a logical transition from the start to end image.
We have firsthand observed some of the deployment tactics for GAN images. Reddit is a popular social media platform with over 400 million active users. Users can upvote or downvote a post based on the relevancy. Upvotes are tallied, then added to a user's “karma” which is a reddit-wide contribution score.
Figure 5
The Potential for General Adversarial Networks (GANs): Text
The Potential for General Adversarial Networks (GANs): Image
Figure 5 displays a typical reddit profile. However, upon further investigation, we can see that the profile picture is in fact, a GAN generated image. These accounts are usually automated with code to post news articles in order to farm “karma”
Why is this concerning? The account is able to control online sentiment by posting politically targeted articles across various pages of reddit without any human involvement.
Another interesting application of general adversarial networks is to gaming, specifically to game level creation and voice chat. The contribution of GAN technology to gaming is behind that of other industries, largely because of the limited amount of data available to train the neural networks. In a paper published in October of 2019, researchers from New York University and The University of Copenhagen cited this as the main difficulty for GAN’s integration into gaming: “For most games, only a limited amount of content exists. Super Mario Bros has a few dozen levels, Mass Effect probably less than a hundred named characters, Skyrim only tens of non-trivial quests, and Grand Theft Auto V a handful of car models and weapon types.” However, researchers are still finding success in using GANs to generate unique, playable game levels. So far GAN technology has been used to create levels for The Legend of Zelda and also to entirely recreate Pac-Man.
The Potential for General Adversarial Networks (GANs): Text
The Potential for General Adversarial Networks (GANs): Video
Furthermore, GAN technology is being utilized by a company called Modulate to create real time voice skins for in-game voice chat. Modulate has named their product “voice skins”, similar to how one can purchase different “skins” (outfits) for characters in video games. With this technology, a player could select the Morgan Freeman voice skin and everything that player says into their microphone would sound as though Morgan Freeman were saying it.
Not only is the concept of voice skins downright awesome, they also can provide anonymity for those wishing to hide their identities online. Many see this anonymity as a blessing and a curse, for while it adds a new level of privacy to multiplayer gaming, it can also allow for people to hide behind their voice skin to use hateful language. However, Modulate says they are prepared for this as they watermark all of their audio.
General adversarial networks are also promising for the growing open world model in titles such as Grand Theft Auto 5, or infinite world generation, such as Minecraft. Leveraging GAN entity generation can promote a more individualized player experience through generating unique entities that are localised to each player's game.
There is still much to be learned about general adversarial networks, as the technology is very young and has not yet piqued the public’s interest. However, once this style of machine learning becomes more mature and refined, the applications for it are endless. Although the use of GANs prompts concern regarding copyrighting, intellectual property, and other issues associated with artificial intelligence, the people developing these networks are aware of the risks and actively working to mitigate them. Hyperion Group will be tracking the various applications and overall popularity of GAN technology in the coming months/years, and we think everyone should be excited about what the future has in store for this brilliant extension of machine learning.
Copyright © 2020 Hyperion Group Inc, All rights reserved.
This newsletter is for informational purposes only. The newsletter does not constitute any actionable entity or provision of service. These statements are not investment advice. Hyperion Group does not claim ownership of any content provided.
Sources
https://arxiv.org/abs/1712.00679
https://techxplore.com/news/2018-05-adversarial-networks-unleashed-video-games.html
https://arxiv.org/abs/1712.00679
https://medium.com/syncedreview/using-conditional-gans-to-build-zelda-game-levels-6004b30bb753
The Potential for General Adversarial Networks (GANs): Text
bottom of page