Telling a yellow taxi and a binoculars (双筒望远镜)apart is so easy that most people could do it without a second thought. However, it is not so for an artificial intelligence: if you turn the taxi upside down, it sees Binoculars.
This is just one of dozens of examples that show AI is a lot worse at identifying(识别) objects by sight than many people realize. The examples, collected by Anh Nguyen at Auburn University in Alabama, raise concerns about the real-world ability of AI image recognition systems.
Nguyen and his colleagues took images of common objects from the Internet and rotated (旋转) and changed the position of the objects in the pictures. They found this was enough to confuse several AI systems, including Google's. In one case, a school bus that was correctly identified in the original image was misidentified as a lunch box when upside down in the road. It shows these systems aren't as intelligent as many people think they are.
Nguyen worries what could happen in real-world situations. For example, it makes sense for a driverless car with AI system to avoid an object it can't recognize. But if the car stop unexpectedly because it thinks a Coke can is a fire fighting truck, this could be as dangerous as thinking a fire fighting truck is a Coke can. This is one reason why driverless cars will need several sensors (传感器) to provide more information, says Nguyen.
His team has been discovering these problems for the past few years, but nobody knows how to fix them. The biggest problem to progress is that when an AI looks at an image, it can't extract (提炼) rules that would help it identify a similar one next time – such as the rule that horses have four legs. "To reach a human level of reasoning, we need a way to extract rules from images," says Nguyen.