Optical illusions fool computers into seeing things

日期:2019-03-05 05:01:10 作者:鱼渺 阅读:

By Jacob Aron COMPUTERS can identify objects with near-human levels of accuracy, enabling them to do everything from creating picture captions to driving cars. But now a collection of bizarre optical illusions has revealed that machines don’t see the same way we do, which could leave them vulnerable to exploitation. Image-recognition algorithms learn to recognise objects by analysing numerous images and identifying patterns that mark out a cat from a coffee cup, for example. Jeff Clune of the University of Wyoming in Laramie and his colleagues wanted to know whether a particular type of image-recognition algorithm called a deep neural network (DNN) could be linked to a genetic algorithm that uses a natural selection-like process to evolve pictures. Such algorithms, with humans rating and thus guiding their results, have previously created images of apples and faces, so Clune expected that replacing the human with a DNN would work just as well. “Instead we got these rather bizarre images: a cheetah that looks nothing like a cheetah,” Clune says. Clune used one of the best DNNs – AlexNet, which was created in 2012 by researchers at the University of Toronto, Canada. Its performance is so impressive that Google hired the team behind it last year. It turned out that the genetic algorithm produced images of seemingly random static that AlexNet declared to be depictions of various animals with more than 99 per cent certainty. Other images, generated in a different way, look like abstract art to humans but fool AlexNet into seeing a baseball, electric guitar or other everyday object. The algorithm’s confusion is due to differences in how it sees the world compared with humans, says Clune. Whereas we identify a cheetah by looking for the whole package – the right body shape, patterning and so on – a DNN is only interested in the parts of an object that most distinguish it from others. Focusing on these elements, rather than the image as a whole, leads the DNN down the wrong path. “It’s almost like these DNNs are huge fans of cubist art,” says Clune. Studying these illusions and the differences between algorithms and humans could teach us more about ourselves, says Jürgen Schmidhuber of the Dalle Molle Institute for Artificial Intelligence Research in Manno, Switzerland. “These networks make predictions about what neuroscientists will find in a couple of decades, once they are able to decode and ready the synapses in the human brain.” More immediately, Clune wants to figure out how to help DNNs ignore the illusions. If DNNs can be fooled by static, an attacker may be able to bypass facial-recognition security systems or even trick driverless cars into seeing misleading road signs. “It opens any application that uses this computer vision up to security hacks,” he says. This article appeared in print under the headline “Computer says cheetah” More on these topics: