What do neural nets dream of?
Dreams are really weird, when you think about them. Let’s lie still for several hours and maybe vividly hallucinate?
A lot of what people call AI these days are algorithms called artificial neural networks, a type of algorithm that’s modeled after the human brain. They have some things in common with human brains - virtual neurons, and even some of the same mechanisms for analyzing images and navigating mazes. But can they dream?
The answer is yes - if you design them that way. There’s a fun interactive model where you can play an algorithm’s dream version of the Doom level it was trying to beat. By training in its own simulation of the level rather than the real thing, it could focus on the most important parts.
Now here’s another way to get an algorithm to dream: train it on the rambling, semicoherent descriptions that humans give of their own dreams. Pete Majarich asked me what a neural network would make of the over 19,000 dream summaries in The Dream Bank.
I have the answer. And it’s pretty weird.
I used a neural network that can build words syllable by syllable - that way, it can make up its own words, yet they’ll probably at least be pronounceable. (It doesn’t do capital letters, though, so I did those for it.) The trained neural net stuck mostly to existing words, and it even got pretty good at the punctuation. But, like most text-generating neural nets, it has a limited memory, as you can see.
I am shaving into a great dream. I see four quite small people in a huge tank with all kinds of old brass grape liquid on his hair and a red neck like a maze. Of course I help him the kitten through his nose to take a plug and nonchalantly look for a commute we can see her company. He points off Father Andrew and opens the gun. I encourage the old man with a loud sandwich of soap deep signs and a simple juice in its face of mine in a box of wool dog, feeling suspected!
I see a cat on the floor where there are two eels (here east who would grow on the floor). I hear the room I have a new deal of ship by which they are washed to go downstairs the structure and they’d be driving through a short table, either someone else too and we have to fix us. I look annoyed and walk around the strip below from my back and the spirits will be awkward.
There was, oddly, one point in the training where the neural net was obsessed with going to the bathroom.
I am in a small room and see a large piece of stairs in the middle of the house. I am in the water and I see a woman in a large room with a few minutes and I see a small baby man and he says he is going to go to the house. I say I don’t want to be in the dream. I am annoyed and say, “Oh no, I don’t want to get in the bathroom.”
But my absolute favorite part was the incredibly bizarre imagery it would come up with.
I notice my car is now a tennis ball.
I am a beautiful, chubby, big bathtub.
There is a water gun and a clown wall and another coconut object.
I shake my cheek and say, “Oh. I respect sleeping in a large metal cart and record cards.”
I see a huge enemy dress with a small platform of wooden occasional people.
Then my teeth are in Colorado.
It has stairs over Derek’s address, and his name is a plastic sandwich.
I see a lovely maglugated creature with a raging red petacles and jock leaves.
I generated thousands and thousands of words of neural net dreams, and let me tell you, it really loves to ramble on. There’s a “temperature” setting that governs creativity, and at the lowest temperature setting it will describe one repetitive neverending dream for all eternity (I’ve had conversations like this). Read them here!
If you’d like a more selective list of the best bits, become an AI Weirdness subscriber to get them as bonus content!
And finally! Pete Majarich has arranged several dream excerpts into fun prints. Check them out here!