The laws of exponential growth are alive and well it seems, at least judging by the advances Nvidia is making in the field of Artificial Intelligence. According to the company, it has now created an AI that is capable of changing the weather as you go, turning night into day and the other way around, and even create big cats out of little cats.
The researchers for the company created a way for the new AI to change the elements of the video content it is provided with. For instance, a rainy day caught on video can be changed into clear-sky weather, or a snowy day can be turned into perfect summer weather, complete with green lawns and trees full of leaves. They can also help morph cats pictures into cougars and backward, creating new species of cats in the process.
The AI they’ve managed to create is extraordinary from the get to, but it’s even more impressive when you consider the implications, and how this tool can help train things like self-driving cars, who are accustomed only to certain situations, depending on the average weather where they’re trained. Google’s cars for instance, have spent more time in places like California, so they might not be as equipped as others to bad weather, how to deal with the way the light shines when it’s snowing, or even how landmarks they use for guidance look like covered in snow.
Nvidia’s research is based on an artificial intelligence method that’s great at generating visual data, called generative adversarial network (GAN). The system works by actually combining two separate neural networks – one that’s there to create the data, and one that’s there to interpret it. In short, the AI can teach itself to create better images over time because it simply rejects samples that don’t look accurate. In a few years, we could see video content that looks real, but it actually isn’t, thanks to this type of technology. Whatever your feelings on AI, one must admit that the thought is a bit unsettling because it means we’ll reach a point in time when we might not be able to trust the things we see online.
The future of AI
GANs are not new whatsoever, but Nvidia has a bit of an advantage over others because it can learn without as much input from the researchers. Normally, this type of AIs require a lot of data to be fed into them so they can learn to accurately create new images. This one manages to do it a lot faster. For instance, in order to shift from day to night, a regular GAN needs a picture taken in the day and one in the night of the same location. Then, after studying the two pictures, it can generate the same effect on other images. Nvidia’s solution works without being fed so much data, finding a way to get the results without labeled datasets.
This type of technology is more than welcomed for training driverless cars, as mentioned above. While it will not be 100% as if driving in the real conditions offered by bad weather, for instance, or on a pitch black street, it’s close enough to reality to create a good baseline.
Artificial Intelligence and its applications will only continue to grow over the coming years at a much rapid pace, especially as hardware catches up in development with technologies that can support more and more powerful AIs.