The truth so far in the format of the experiment – to implement the function in the social network is not planned.
Researchers from Facebook published the results of the experiment with artificial intelligence. He was taught to “open” the eyes of users on the pictures.
Employees of the company trained a neural network on the basis of images with the eyes open, so that she understood how the human face looks and did not spoil it when retouching. In the process of learning, the system has learned to recognize the characteristics of the eyes, such as color and shape, and take into account various factors, for example, the human pose and lighting.
To train artificial intelligence, it took 2 million 2d-images. Among them there were about 200,000 individual people who had at least one photograph with their eyes open. Also used photos of 17 thousand celebrities who were taken from the Internet.
According to the results of the study, ordinary users in most cases could not notice that artificial intelligence “opened” the eyes of people in the pictures. In 54% of cases, they were sure that the eyes in the photographs were initially open.
However, the system has its own limitations: for example, it is hampered by glasses and some hairstyles. The neural network is also sensitive to “noise” in the photo and unusual postures. In addition, to work for her in any case, you need a picture of a particular person with open eyes.