Home → News → The Best Image-Recognition AIs are Fooled by Slightly... → Full Text

The Best Image-Recognition AIs are Fooled by Slightly Rotated Images

By New Scientist

March 20, 2019

[article image]

Researchers at Auburn University have identified dozens of examples of how artificial intelligence (AI) is much worse at identifying objects by sight than many people realize.

For example, distinguishing between a yellow taxi and a pair of binoculars seems simple, but if an image of a taxi is flipped upside down, many AI systems see a pair of binoculars.

The Auburn researchers took images of common objects from ImageNet and randomly rotated and changed the position of the objects in the pictures. The team found this minor change was enough to confuse several state-of-the-art image-recognition systems 97% of the time, averaged across all of the systems.

The biggest hurdle to progress for this technology is that when an AI system looks at an image, it cannot extract rules about the object shown that would help it identify a similar object in the future.

Said Auburn researcher Anh Nguyen, "To reach a human level of reasoning, we need a way to extract rules from images."

From New Scientist
View Full Article


Abstracts Copyright © 2019 SmithBucklin, Washington, DC, USA


No entries found