Google taught a computer to dream and the results are incredibly psychedelic
Dreams often don't make complete sense, but they can be beautiful and spectacular. And now, we can see how Google's artificial intelligence dreams.
The artificial "neural networks" that run Google Images can be told to take what they "know" objects look like, and draw what they think fits on top of other images. The results are sometimes incredible, sometimes terrifying.
Scroll down for the details, but for now, just enjoy these incredible computer images:
These mind-melting adaptations of
You can watch the entire clip in all its unbelievable trippiness on YouTube.
This mental Asian pagoda
The animals that roam this interpretation of a landscape
This alien city dreamed up by an AI
This watercolour-style fantasy realm I want to live in
The visions of DeepDreamBot, which takes random Twitter images and hands them to the dream machine
No, seriously. How does all this WORK?
The computer giant's system for achieving these incredible, surreal images is fairly complex. Essentially, the neural net has been trained with millions of images - so, it it knows what objects look like.
But what it thinks a banana is turns out differently from what you or I see.
But if the computer knows what, say, a screw looks like, then in theory it can a screw.
By giving the machine permission to "draw" whatever it thinks fits on top of another image - whether it's just white noise or a film still, the AI starts to conjure up its own images based on almost nothing - which is why it's called "dreaming" by the engineers.
Then, for the dreamscape images at the top of this post, it's given its own image again - iterating the process over and over. That's why the engineers have called it "inceptionism" - layers and layers of dreams.
Nifty, isn't it?
If you're keen on a more technical explanation, you can read more at the Google engineers' blog.
