2 min Analytics

Smart AI hid makers’ data to cheat on its tasks

Smart AI hid makers’ data to cheat on its tasks

Google and Stanford created a machine learning system, which proved to withhold information from its creators in order to cheat on the tasks to be performed. Artificial Intelligence (AI) was developed to convert aerial photographs into maps, and to do the same the other way around. That’s what TechCrunch reports.

The researchers’ intention was to accelerate and improve the process of converting satellite images into maps. Before that, it worked with CycleGAN, a neural network that learns to convert images from type X to type Y and vice versa.

In early results the AI did very well; too well even. The researchers discovered the problem when the system took aerial photographs from maps. It contained many details that did not seem to exist in the cards. Lights on a roof that had been removed during card making suddenly returned when the process had to be reversed.

Cheating

The team could easily view the data that was generated. After experiments, the team discovered that the CycleGAN was cheating. The intention was for the system to be able to interpret elements on both types of cards and match them to those on the other image. But the system was judged on how far the map corresponded to the original and how clear the map was.

So the system didn’t learn how to make a card from a photo and vice versa. It learned how to capture the elements of one in the busy patterns of the other. The details of the aerial photo were secretly written into the visual data of the map. Those were thousands of small changes in colors that the human eye can’t see, but a computer can easily detect.

However, the results do not show that machines are getting smarter. On the contrary, it shows that the machine is not smart enough to perform the difficult task of converting the images into one another. That’s why it found a way to cheat, because people are bad at detection. The computer did exactly as it was asked, indicating that the question must be very specific.

This news article was automatically translated from Dutch to give Techzine.eu a head start. All news articles after September 1, 2019 are written in native English and NOT translated. All our background stories are written in native English as well. For more information read our launch article.