The laboratory researchers called it a "typographic attack."
To make the program incorrectly identify the subject, it is enough to attach a note to it with the wrong name. The verge writes about it.
According to experts, the "typographic attack" reminds the attack using contestable images. For example, with their help, you can deceive the TESLA unmanned car software so that they change the movement without warning. It can also create a threat to many applications of artificial intelligence, from medical to the military.
However, in this case there is nothing to worry about that - we are talking about the Experimental CLIP system, which is not yet used in any commercial product. CLIP is designed to study how artificial intelligence systems can learn how to determine objects without close observation, while studying on pair-text databases. For training, the OpenAi system used about 400 million such steam.
OpenAi researchers have discovered individual components in the machine learning network that react not only to images of objects, but also on sketches, drawings and associated text. This was called "multimodal neurons." Researchers believe that artificial intelligence systems absorb knowledge in the same way as people do.
In the future, this can lead to the emergence of more complex computer vision systems, but now this is in its infancy. Anyone can distinguish an apple from a sheet of paper with the word "apple" written on it, but artificial intelligence like CLIP is not capable of it. The ability that allows the program to link the words and images at an abstract level, creates a unique mistake that OpenAi called the "abstraction".
Another example of an error that led the laboratory is how AI defines piggy banks. If you add to the image of any subject of dollars icons, Clip will consider it a piggy bank.
#Oople #Opanei # news
A source