Like many nerds, I’ve spent a lot of spare time over the past week playing with the open-source code for Google’s DeepDreams visualisations. This runs a visual-recognition neural network—basically, an artificial model of a visual cortex which has been trained on a big image database—through a feedback loop, which adds all sorts of psychedelic hallucinations to images. (If you are interested in the technicalities, here’s my post on how I got it running. If you want to get started with less fuss, I’d recommend Ryan Kennedy’s containerised version.)
Most of the output has only used a single layer of the net: there are actually
nine available layers [actually, as it turns out, there are stacks more than nine, but these are the ones named “output”] so I set up a loop to apply each of them to a single image (some randomly-generated colour noise) and see what the different effects were. Here they are, ranked in ascending order of eldritch horror.
3a. This is the lowest-level layer, so it’s all rather abstract and decorative in a 1950s-public-art way. Like bacteria, or tapeworms, as my daughter said.
4b. When I set this experiment running, I expected the results to get steadily more nightmarish as they became more figurative, but I was wrong. These incredibly stoned dogs are almost cute.
4a. The level preceding 4b is similar except that the dogs look pissed-off. I call this one “The Seer”
4c. This is the default layer which the open-sourced code uses: “puppyslugs”, or, I guess, eventually, “DeepDreams classic”. Here we see a single puppyslug in its native habitat of trippy as hell eye-cobwebs.
If this gives you the creeps, close the tab now. We’re not even halfway. The Google researchers mercifully withheld the worst.
5a. While layers 3b, 4a, 4b and 4c all show a distinct tendency to furriness, at 5a we’ve left the mammalian world far behind. This is the second-last layer, but even though it’s a mass of exoskeletons and scales, it’s still creating things that you could imagine seeing. Through a microscope, in a drop of very bad pond water. When you’d dropped acid.
3b. This is a step back to the lower levels of abstraction, and at first it seems like interesting fuzzy spirals and loops. Until you notice the… well, they’re not exactly dogs. Doglike features. Cells, even. Eyes. Emerging.
5b. The Elder Gods. This is the final layer in the neural-net stack, and it’s pretty bad. At least the flea-fish-microbe things of 5a had three-dimensional-ish bodies. At least the puppyslugs had… faces. I miss the puppyslugs. I miss their little weird faces.
WHY DID I SAY I WANTED FACES. NO MORE FACES.
Are those… apes? Eating a piano-accordion. Inside cobwebs. (Don’t look too closely at the cobwebs.) This is a Max Ernst nightmare.
4d. China Miéville’s fictional fantasy world Bas-Lag has a region call The Torque, a relic of an ancient war, which makes things go wrong. I never really cared to see anyone try to visualise its effects.