Google's artificial intelligence subsidiary in London, DeepMind, has created a self-training vision computer that can generate 3D models from 2D images.

DeepMind's AI-Computer Can Make 3D Models From 2D Snapshots

Scientists at DeepMind have created a new system that can teach itself to see and understand a three-dimensional environment.

The Google-owned research company published a new paper in the journal Science on Thursday, June 14, unveiling a new type of system that can create 3D models of a scene just by looking at a 2D snapshot, much like humans.

The system, called the Generative Query Network, can imagine and render the scene without any human intervention or training, according to DeepMind CEO, Demis Hassabis.

Just by looking at a set of images of a scene, such as a room with colored shapes on the floor, the system can reconstruct hidden sides of the objects and create a 3D view from different viewpoints, even taking into consideration factors like textures, colors, lighting, and shadows.

Like Humans, The GQN System Can Teach Itself

Like humans, the AI-powered system mimics the manner in which the human brain understands its surroundings just by looking at it, like how objects get bigger or smaller as the distance changes.

"Much like infants and animals, the GQN learns by trying to make sense of its observations of the world around it," DeepMind researchers noted in a blog post on Thursday. "In doing so, the GQN learns about plausible scenes and their geometrical properties, without any human labelling of the contents of scenes."

The GQN system has two parts: a representation network and a generation network. While the representation part is responsible for taking the agent's observations as input data and encoding it into a mathematical form called a vector, the generative network creates the image based on the vectors and predicts what the scene would look like from a different vantage point.

It's like a person is given a handful of pictures of a room and then told to use their imagination and draw what they would see if they were standing in a particular spot inside the room.

DeepMind's Other Projects

This is the latest breakthrough in a series of high-profile projects for DeepMind, in which AI-systems are demonstrating a previously unknown ability to learn on their own once human programmers set the basic parameters.

In October, DeepMind's Alpha Go displayed its ability to play the uber-complex board game Go. Last month, another DeepMind-developed AI system taught itself to navigate through a maze.

ⓒ 2021 All rights reserved. Do not reproduce without permission.