Skip to main content

This machine is so bad at describing planets, it's funny

Neural networks have no idea what space looks like

You know what a comet looks like. But when a Finnish YouTuber asked an artificial neural network to describe what it was seeing in the Star Trek: The Next Generation intro, its best guess was a matchstick.

Ville-Matias Heikkilä built a neural network trained with the contents of the ImageNet database, and asked it to guess what it saw in the iconic sequence of stars, planets and other space phenomena. Unfortunately, there's not a lot of space phenomena in the database, so the results are... not impressive.

In each frame, the top three classifications are listed, with the top choice marked with an asterisk. Green means that the algorithm is pretty sure about the classification.

The Enterprise-D is a CD player or an odometer. A nebula is a jellyfish. The Star Trek logo is a digital clock. It's also bizarrely fixated with roundworms. Still, I'd be up for a planet made of chocolate sauce. How about you?

Via Gizmodo