TELLING a yellow taxi and a pair of binoculars apart is so easy most people could do it standing on their head. Not so for an artificial intelligence: flip the cab upside down and it sees binoculars.
This is just one of dozens of examples that show AI is a lot worse at identifying objects by sight than many people realise.
The examples, compiled by Anh Nguyen at Auburn University in Alabama, raise concerns about the real-world ability of image-recognition systems, for example in driverless cars. 鈥淚t鈥檚 a huge problem,鈥� says Nguyen.
Advertisement
We already know AIs are often flummoxed by doctored pictures that humans can recognise without any problem 鈥� such as a turtle with a special shell pattern that an AI misidentifies as a gun.
However, these so-called adversarial images are designed specifically to trip up AIs. They are contrived and unlikely to be a problem in the real world, or so the logic went. Now it seems you barely need to alter an image for AIs to seriously mess up.
Nguyen and his colleagues took images of common objects from ImageNet 鈥� a database used to train AIs 鈥� and randomly rotated and changed the position of the objects in the pictures. They found this was enough to confuse several state-of-the-art image- recognition systems, including Google鈥檚, 97 per cent of the time, averaged across all of the systems.
In one case, a school bus that was correctly identified in the original image was misidentified as a punching bag when shown in close up, and as a snow plough when upside down in the road ().
鈥淚mage-recognition systems aren鈥檛 as intelligent as we think they are, which raises safety questions鈥�
It shows these systems aren鈥檛 as intelligent as many people think they are and raises important questions about how safely they can be deployed in certain applications, says Jeff Clune at Uber AI Labs in San Francisco.
Nguyen worries what could happen in chaotic situations such as a battlefield, where you want an image-recognition system to be able to identify objects from lots of different perspectives. Similarly, AI-based airport security scanners need to be able to identify objects in bags from multiple angles.
There are big implications for driverless cars as well. It makes sense for a vehicle to avoid any object it can鈥檛 recognise. But if a car slams on its brakes because it thinks a Coke can is a fire engine this could be as dangerous as thinking a fire engine is a Coke can and doing nothing. This is one reason why driverless cars will need to have multiple sensors to fall-back on, says Nguyen.
He, Clune and others have been uncovering these problems for the past few years, but nobody knows how to fix them.
The biggest barrier to progress is that when an AI looks at an image, it can鈥檛 extract rules about the object shown that would help it identify a similar one next time 鈥� for example the rule that horses have four legs. 鈥淭o reach a human level of reasoning, we need a way to extract rules from images,鈥� says Nguyen.
