CAPTIVATING SOFTWARE WITH LOTS OF POTENTIAL
Creating a virtual environment that looks and feels real typically takes a lot of time and energy. The details have to be developed with graphics chips and things of that nature. One of the latest populate video games Red Dead Redemption 2 took a team of 1,000 developers nearly 8 years to create while working 100 hour weeks. This highlights just how much work typically has to go into such a project. However, some day soon that kind of workload will no longer be required because we will be able to utilize AI to do it all in a fraction of a time. GAN technology is utilized to make this a reality.
Software of this magnitude will not only cut down the amount of time it takes to make games but could be used for more practical applications as well. For example, the ability to generate an entire virtual world very quickly could prove useful for crafting simulations in which self driving cars learn how to drive. It could even be used to teach humans all kinds of new skills in half the time by providing them with a “world” to practice on that could be almost indistinguishable from the real world.
This latest technology has been created by a company called Nvidia. They have a reputation for chip making but have taken the plunge into artificial intelligence to explore the many possibilities. Bryan Catanzaro the vice president of applied deep learning at Nvidia says “We can create new sketches that have never been seen before and render those.” He also says “We’re actually teaching the model to draw based on real video.”
HOW DOES IT WORK?
Nvdia’s research team uses a standard machine learning approach to identify different objects in video such as people and cars. They then use what is know as general adversarial network or GAN to train the computer to create the realistic 3D imagery from the sketch provided. Essentially the system can be fed the outline of a scene and the program will fill in all the amazing details in no time at all. They are having some issues with some of the objects still appearing warped or twisted during testing but nevertheless the effect can be stunning.
This is a very easy method to utilize for game development purposes and will revolutionize the industry as a whole. Also as mentioned before synthetic training simulations can be created to assist robots and humans alike in learning various tasks. It can be difficult to simulate all the possible situations in real life for practice but soon this technology will allow us to do it via virtual reality.
WHAT IS A GAN AND WHAT CAN IT DO?
GAN takes advantage of all the possibilities that arise from taking two neural networks and having them work together in a very specific manner. The result is a mixture of fantastic creativity and a huge dark cloud of looming controversy. They are not only responsible for making entire virtual worlds and captivating images but also for developing what is known on the internet as “deepfakes”.
The secret lies in the fact that the two neural networks are simultaneously working with each other and against each other in a sense. Both networks are fed a ton of training data and each one has a different task. The first network is known as the generator and is responsible for outputting handwriting, video, or voices by mimicking the training data. The second network is known as the discriminator and is designed to determine whether the outputs generated by the first network are correct in the sense that they are as close a match to the training data as possible. Each time the discriminator rejects a series of outputs the generator gets back to work and comes up with closer matches each time.
Eventually the discriminator can’t tell the difference between the outputs and the training data. Mimicry becomes reality. However, one might begin to wonder if the good outweighs the bad when one considers the fact that in addition to being extremely useful in areas of photo editing, animation, and medicine that there is much potential for abuse.
Creating fake videos with the power of GAN.
SERIOUS POTENTIAL FOR ABUSE
This wouldn’t be a proper blog post covering some recently created piece of AI software without talking about the dark side of things. After all, the fact that AI is a double edged sword is part of its appeal. An appeal that must be kept in check.
GAN can be used to develop some beautiful things but it can also be used to create some grossly unethical things like celebrity faces on the bodies of porn stars or creating a fake video depicting the president saying anything you want. Photos depicting certain celebrities in this manner along with other atrocities is a part of a cluster of photos circulating the internet that has come to be known as “deepfakes”.
This is obviously reassuring in the short term because it implies that fake pictures and videos can still be ousted as fake. Eventually though the technology may improve so much that the deepfakes become completely indistinguishable from reality. At that point we could be faced with some serious problems. Also, GAN software has been used to try and steal fingerprints and bio metric data from unsuspecting victims.
For the time being GAN still has many computational limitations which limits the extent to which it can cause damage especially since the deepfakes are still obviously fakes that can be noticed by the human eye. But in terms of long term safety we have only just scratched the tip of the iceberg. Like many other technologies powered by AI this will require some specific regulations to diffuse risks. That in turn requires us to have conversations we might not want to have and might not be ready to have.
What do you think of all this? What should be done in order to keep a lid on all the dangers? Leave your comments down below.