In the last two decades, computer scientists have been developing, training, and teaching machines to successfully see the world around them. However, artificial eyes have only recently started to match their biological predecessors. This year has seen improvements in two areas when it comes to AI image processing, including facial-recognition technology in security and commerce as well as image generation in art.
Google’s DeepMind Division
In September earlier this year, a team of experts in the DeepMind division at Google published a paper which outlines the operation of their brand-new Generative Adversarial Network. This image-generation engine, dubbed BigGAN, influences the massive cloud computing power of Google to create incredibly realistic images. The system can also be influenced to generate visual mashups of symbols, objects, and anything else you tell the system to do. The source code has already been released onto the internet by Google, allowing creators to borrow its processing power to utilize the system as they see fit.
“I’ve been quite surprised by the interactive demos on the web that people have started to transform the algorithms into since we released it,” Janelle Shane told researchers who is a neural-network programmer by night and an optics scientist by day. She also stated that in the past, most researchers would usually publish their findings and that was that. If was difficult to find a YouTube video when it comes to the subject.
“However, these days, researchers will publish their code, model, and a kind of web application for others to attempt their own model as well,” Shane continued. This is exactly what a GANbreeder developer has done by the name of Joel Simon. The web app will enable users to remix and generate BigGAN images over several generations to develop and create their own unique creations. “With the web interface from Simon, you can witness what occurs when you not generating symbols or pictures but can generate something that’s a mix between a shark and a comic book for instance.
By offering access to systems like this, people who don’t know how to operate, train, program, or develop complex neural networks will be able to quickly learn how to do it with this type of technology. “You can quickly interact with the algorithms provided by GANbreeder and receive new artistic results, Shane informed reporters. “You will also be able to see where the limitations are which enables the technology to evolve more quickly as well. There’s loads of room to play around and the best part about this is that people with no previous experience with programming will be able to use the technology and computing power to train themselves.”
There are certain drawbacks to using this technology, however, as trolls and scammers can utilize BigGAN to make counterfeit video and images. This makes things a bit more difficult for Shane as they need to develop an eye for spotting an image that has been generated as opposed to an image that is authentically taken.