Google Trains AI To Trick Humans Into Seeing Things That Aren’t There

By Prosyscom
In March 2, 2018

Image via Shutterstock

It’s well-established that robots might outsmart people one day, but who knew it could happen this soon?

A key objective of machine learning is that it should work as closely to the human brain as possible—which means if a person isn’t bamboozled, AI shouldn’t be either.

A team of researchers at Google—led by Ian Goodfellow—recently published a paper titled Adversarial Examples that Fool both Human and Computer Vision, which acknowledges that tools used to trick an AI’s perception usually won’t work on humans. It elaborates that this information can help drive these computers to become smarter.

“Here, we create the first adversarial examples designed to fool humans.”

The Next Web points out that just last year, a research team from MIT attempted to con Google’s AI with a code that would influence it to believe some turtle images were pictures of rifles instead. The computer succumbed, but a young child would know the difference.

Of course, the neural network wouldn’t truly be at fault for failing the test. It doesn’t have eyes, and real-life perception isn’t something you can teach with the click of a button.

Working to overcome this barrier, Google is studying both human and computer brains. By doing so, it hopes to “conclusively” pinpoint the aspects of imagery deception that AI currently accepts but people reject.

The researchers explain that on top of applying computer algorithms to their dataset, the photographs were also manually tweaked with “transformations performed by the human eye” to train the models to process images like a human visual system would.

Image via Google

With AI and human assistance, the scientists were able to manipulate an image of a cat (left) to look like one of a dog (right), and successfully misled participants for an extended period of time. “[In] most cases, our adversarial examples fool humans only after a brief exposure, [but] the example depicted has a strong effect even for long viewing duration.”

If you were to look more closely, though, you would be able to notice that the “dog” still possesses cat-like features, such as whiskers.

The experiment confirms one thing: nothing beats the human eye—yet. The team adds that it will be possible to completely trick viewer’s eyes should research finally determine “human-meaningful features” down pat.

Image via Google

[via The Next Web, images via various sources]

قالب وردپرس