22.07.2019, 22:46
Library of Images That is Not Available for AI is Created
Source: OREANDA-NEWS
OREANDA-NEWS Scientists from Berkeley have collected thousands of photos that neural networks cannot cope with recognizing.
Recently, scientists have added another one to the library list - the ImageNet-A group, which contains photos in which even the most advanced AI algorithms are mistaken. New dataset researchers from the University of California at Berkeley presented in an article published in the online library of preprints. According to the authors, the set contains about 7,500 regular images, in recognition of which neural networks make mistakes in 98 cases out of 100.
In fact, as machine vision systems become more and more widespread, technologies for disrupting their work are becoming more and more dangerous. It is shown that even a relatively small manipulation of the picture can confuse even the most advanced AI by organizing a “adversarial attack” on its algorithms.
However, ImageNet-A shows that such images can be obtained without any planned intervention, in a random way - by themselves. Consequently, any machine vision that we are going to entrust human lives to must cope with the appearance of such images. For this purpose, a new dataset appeared.
Recently, scientists have added another one to the library list - the ImageNet-A group, which contains photos in which even the most advanced AI algorithms are mistaken. New dataset researchers from the University of California at Berkeley presented in an article published in the online library of preprints. According to the authors, the set contains about 7,500 regular images, in recognition of which neural networks make mistakes in 98 cases out of 100.
In fact, as machine vision systems become more and more widespread, technologies for disrupting their work are becoming more and more dangerous. It is shown that even a relatively small manipulation of the picture can confuse even the most advanced AI by organizing a “adversarial attack” on its algorithms.
However, ImageNet-A shows that such images can be obtained without any planned intervention, in a random way - by themselves. Consequently, any machine vision that we are going to entrust human lives to must cope with the appearance of such images. For this purpose, a new dataset appeared.
Комментарии