MIT’s research shows AI has own "hallucinations"

Dec 30, 2017, 3:25 AM EST
(Source: ▓▒░ TORLEY ░▒▓/flickr)
(Source: ▓▒░ TORLEY ░▒▓/flickr)

A team of researchers at MIT’s Computer Science and Artificial Intelligence Laboratory recently exposed the vulnerabilities of artificial intelligence by showing how “hallucinations for algorithms” can befool the technology to mistake one image for another, a development that prods AI developers to pull up their socks for new challenges.

Anish Athalye and his team have trained an algorithm to create thousands of such hallucinations to trick Google’s image recognition system, writes New Scientist.

“Adversarial examples” are specially designed optical illusions that can trick AI, a limitation that could be critical, for example if a self-driving car mistakes a stop sign for a speed limit one, notes Wired