Interest in artificial neural networks has skyrocketed over the years as companies like Google and Facebook have invested heavily in machines that can think like humans. Today, an AI can recognize objects in photos or help generate realistic computer speech, but Nvidia has successfully built a neural network that can create an entire virtual world with the help of a game engine. The researchers speculate this “hybrid” approach could one day make AI-generated games a reality.
The system build by Nvidia engineers uses many of the same parts as other AI experiments, but they’re arranged in a slightly different way. To goal of the project was to create a simple driving simulator, but without using any humans to design the environment.
Like all neural networks, the system needed training data. Luckily, work on self-driving cars has ensured there’s plenty of training footage of a vehicle driving around city streets. The team used a segmentation network to recognize different object categories like trees, cars, sky, buildings, and so on. The segmented data is what Nvidia fed into its model, which used a generative adversarial network to improve the accuracy of the final output. Essentially, one network created rendered scenes, and a second network would pass or fail them. Over time, the network is tuned to only create believable data.