
The Generative Query Network can pick up basic info to create 3D spaces
Processing images and imagining scenes without given information is one of the most important characteristics of the human brain. However, this has always been incredibly hard for computers to reproduce, but researchers from Google’s project DeepMind have just changed that. By using a Generative Query Network, they managed to achieve better image processing for AI.
Computers had a hard time building 3D images from scratch
Humans are remarkable creatures because they can pick up information about the world just by taking one single look at it. For instance, if someone tells you to think of an object, you’ll start imagining all the other items associated with it. A computer usually has a hard time imagining a space if it doesn’t receive external information on it.
Therefore, Google researchers decided to teach an AI image processing and how to imagine 3D spaces. More precisely, the machine will receive a special training that helps it make inferences. The Generative Query Network was able to construct a 3D space after looking at two 2D images, just like the human brain. In other words, both the brain and the machine assume what information should come in missing spots.
The Generative Query Network can learn on its own from basic information
Here’s how the Generative Query Network functions. The first component transcribes the images into a code version and feeds them to the AI. Based on its previous training, the network generates all the missing information. Therefore, researchers proved it was possible to teach a machine about perspective, lighting, and other principles typical of 3D spaces.
The research is still in its prime, so there’s a long way until the Generative Query Network can work quickly. However, this is one of the biggest achievements in terms of machine learning. The research has many implications. From now on, researchers will be able to feed AIs brief data and watch them learn on their own.
The study on this remarkable mechanism was published in the journal Science.
Image source: Public Domain Pictures