Wednesday, September 2, 2015

Inceptionism

I recently read a blog post by Google about "inceptionism" where they describe a way to gain insight into the behavior of neural networks for image recognition by asking the networks to show them what they "see" in images of random noise and arbitrary scenes. The results of this process are called "deep dream" images for short, and they are fascinating but also more than a litte freaky. Popular science called them nightmares, and I don't disagree.

Actually the images really remind me of the visions I used to have as a young child when I'd rub my eyes really hard just before going to sleep. Strange fractalized swirling noise with recognizable objects blending together in infinite spinning repetition. They freaked me out then too, but I couldn't help myself. I'm not sure whether it's exciting or terrifying to think that Google's deep image learning tools have apparently managed to mimic the human brain so accurately that they can even mimic the patterns we see where none really exist.

Search google for more inceptionist images.