Directory of RSS feeds
Statistics

RSS feeds in the directory: 2798

Added today: 0

Added yesterday: 0

Hi-Tech / Internet

Artificial intelligence has learned to see as well as humans

Inside it 20.12.2018 at 07:47

News information technology, technology news, it industry

Despite the significant progress made recently, machine vision systems fall far short of their biological counterparts.

This applies not only to the speed of information processing and reliability. The fact that the brain is able to recognize objects regardless of what part of it can be seen at the moment. And sometimes even based on the context of the overall plan. It is this vision and taught AI experts from the USA, passed the news of it with reference to hi-news.ru.

For the development of responsible scientists from the University of California at Los Angeles and Stanford University, their system can see the world through the same method that I use. According to the editors of the Proceedings of the National Academy of Sciences, the system can identify objects based on some parts that were not available to any AI.

the process of learning a new method consists of three consecutive stages. After the system asked to find an object, it breaks the image into smaller parts. The figure above shows how the system identificeret object on the basis of just such small parts. Then each part is separately analyzed and identifitseerida its relationship with other objects of the picture. Then again the AI "looks" at the whole picture and identifies which parts of the picture are related to the original object and which are not. Well, after the AI started to train on numerous photos and videos taken from the network.

"the Internet has provided us with two things that help our computer vision system to learn as well as humans. This is, firstly, the presence of a large number of images and videos that contain objects of the same type. And secondly, in the network lot of photos and video, where the same objects shown from different angles, at different distances from the viewpoint and under different conditions." — said the study's lead author, Wani Roychowdhury.

At the final stage, the scientists tested the system on more than 9000 images of people and different objects and in all cases it is required to recognize at least no worse than the AI with a huge experience, but trained in the "old" method.