Until recently, modern computers with artificial intelligence technology did not know how to intuitively recognize an image - computer vision cannot create a “mental” image of an object the way a person does. Such systems are not capable of self-learning, so they are trained by demonstrating full-size objects, each of which is then identified by a robot.
Until now, computer vision systems did not have the ability to create a complete picture of an object based only on its parts. For this reason, AI can be misinformed by showing a familiar object in an unfamiliar environment.
This lack of machine vision is one of the problems that scientists are trying to fix by creating a system that can recognize objects as it does in humans. For example, a person who sees the tail of his pet is able to understand where his head, paws, etc. are located, that is, based on a part of the image, mentally create his complete image.
An artificial intelligence system, developed jointly by scientists at California and Stanford University, has learned to recognize objects based on some of its parts. The approach used in the system is identical to the human perception of objects. To help computer intelligence, the authors of the method immersed it in a virtual copy of the human environment.
The method by which machine vision has learned to see as a person consists of three stages. In the first step, the AI divides the picture into small pieces. Next, the computer learns to determine how small parts can form combinations with each other, creating a solid object. At the third stage, the machine captures accompanying objects that are present in the foreseeable space and their connection with the primary object.
In the final, the developers tested computer vision, showing several thousand pictures of people and other objects. Artificial intelligence has managed to recreate the image of a person in detail. Similar tests were done with images of cars, airplanes and motorcycles. As a result, the computer was able to identify objects no worse than a generation of AI systems trained in a different way.
The Topic of Article: Artificial intelligence has learned to see as a person.