Someone used Google software to build an AI system that can identify which restaurant a bowl of ramen is from, and he isn’t even sure how it works.

As Google’s Kaz Sato explains in a post on the company’s blog The Keyword, the AI system was developed by data scientist Kenji Doi. He wanted to see if he could train Google’s AutoML Vision machine learning platform to tell which of the 41 Ramen Jiro shops in Tokyo a given meal is from, and after feeding it a set of 48,000 pictures of ramen bowls, he succeeded – the system could identify which shop a given bowl is from with 94.5 percent accuracy.

The thing is, Doi isn’t even sure how the AI system does it. At first, he thought it could tell by the color or shape of the bowl or table, but the AI system proved to be successful even when confronted with identical bowl and table design. Doi’s new theory is that the model can “distinguish very subtle differences between cuts of the meat, or the way toppings are served.”

It’s a testament to both the accuracy and the alienness of artificial intelligence, which can pick up on visual patterns that are indiscernible to the human eye. And it’s suggestive of the latent power of machine learning AI systems, an area in which Google has demonstrated an intensifying interest as it has sought to bring advanced AI systems to its hardware products and other offerings.

Source: The Keyword