Comment by geor9e
3 years ago
My hypothesis is that the neural network was trained on a lot of labeled photos, so somewhere inside the network, when you see the moon, it has some moon=0.95 confidence number, and whatever label has the highest confidence, it tries to bring it up to 1.0 akin to how deepdream makes images of spaghetti have more dog faces. Samsungs marketing department interprets that as technically enhancement of images and not faking the moon specifically. So perhaps if it sees Jupiter, it will try to make it more jupitery.
No comments yet
Contribute on Hacker News ↗