The Uncanny Wormhole
July 9, 2015 10:34 AM   Subscribe

DeepDream, Google's code for visualizing neural networks, is being used like some unholy Lovecraftian Instagram Filter to produce disturbing, surrealistic photos and videos,including upping the psychedelic ante in Fear and Loathing in Las Vegas. Many, many of the photos and videos are Not Safe For Life, Work, or Sanity. Almost all of them are very weird, especially the food ones.
posted by Cookiebastard (94 comments total) 68 users marked this as a favorite
 
KlingonSnackFilter
posted by a lungful of dragon at 10:36 AM on July 9, 2015 [3 favorites]


I love this tool. The images it pukes out are the closest representation of what the world looks like through psychedelics that I've ever seen.
posted by Evstar at 10:38 AM on July 9, 2015 [9 favorites]


That photos link features user submitted pictures, for example a dick pic of that guy with two penises, in case you are thinking of looking at the links at work.
posted by Elementary Penguin at 10:39 AM on July 9, 2015 [2 favorites]


I guess Cyriak's out of a job, then.
posted by Jon Mitchell at 10:40 AM on July 9, 2015 [10 favorites]


I saw a collection of porn photos that had been run through Deep Dream last night... that thing really likes dogs. Like, a LOT.
posted by palomar at 10:41 AM on July 9, 2015 [9 favorites]


The images it pukes out are the closest representation of what the world looks like through psychedelics that I've ever seen.

Wouldn't it be something if all this stuff led to a confirmation that psychedelic substances work by making your brain's neural network run backwards, just like this software?
posted by JoeZydeco at 10:44 AM on July 9, 2015 [5 favorites]


Ready the thinkpieces about how we've reached Peak Weird
posted by prize bull octorok at 10:48 AM on July 9, 2015 [2 favorites]


I guess Cyriak's out of a job, then.

NEVAAAAAAAAAAAAAAAAAAHHHHHH!!!!!
(makes A's and H's from bouncing fractalizing phallus-sheep dancing to oompah-techno)
posted by lalochezia at 10:48 AM on July 9, 2015 [9 favorites]


Evstar: "I love this tool. The images it pukes out are the closest representation of what the world looks like through psychedelics that I've ever seen."

Indeed. I'm used to seeing weird pictures on various forums, where people will often comment, "I bet that's what it looks like when you're on LSD or mescaline. LOL."

And I always reply, (because I'm an asshole), "No. That looks nothing like an acid hallucination. Trust me."

But this stuff does. It is as close to an acid visual that you can get while stone cold sober. I love it.
posted by Splunge at 10:49 AM on July 9, 2015 [4 favorites]




The damn thing loves dogs.
posted by dilaudid at 10:52 AM on July 9, 2015 [3 favorites]


When do the open source programs come out for us to download and play with?????
posted by ian1977 at 10:53 AM on July 9, 2015 [1 favorite]


Looking at these, I realize that one of the uploaded photos is of a relative. In waking life, my relative has only the normal number of eyes, limbs, etc, and only in the normal configurations.
posted by Frowner at 10:53 AM on July 9, 2015 [4 favorites]


Cuz Skynet hates cats!
posted by ian1977 at 10:53 AM on July 9, 2015


Not to be limbonormative or anything.
posted by Frowner at 10:54 AM on July 9, 2015 [7 favorites]


It loves dogs, yes, but even more than that it seems to love eyes.

(I'm guessing because we love taking pictures of dogs and other things that have eyes.)
posted by A dead Quaker at 10:54 AM on July 9, 2015 [3 favorites]


Yes, why so many Schnauzers?
posted by Omnomnom at 10:58 AM on July 9, 2015 [2 favorites]


So it is going to be a waitlist of 102,000 selfies isn't it? Good grief.
posted by ian1977 at 10:59 AM on July 9, 2015


When do the open source programs come out for us to download and play with?????

Google has actually put up an IPython notebook you can use to run deep dream yourself, and Adrian Rosenbock has used it as the basis for a new module, fittingly called bat-country.
posted by Going To Maine at 10:59 AM on July 9, 2015 [8 favorites]


Enhance 224 to 176, ENhance...EnhAAAnce... ENNHAECEE... hencanhe... c@4nadee... d@h4dee... do45ree.. d@ggee
posted by smidgen at 11:02 AM on July 9, 2015 [17 favorites]


If it's been trained on Internet images, why isn't it full of cats?
posted by iotic at 11:03 AM on July 9, 2015 [5 favorites]


that thing really likes dogs. Like, a LOT.

Yeah, why is it so prone to find animal parts (and especially faces, and especially dog faces) in things?
posted by escape from the potato planet at 11:07 AM on July 9, 2015


My god, it's full of cats!
posted by I-baLL at 11:07 AM on July 9, 2015 [1 favorite]


It seems to add extraneous eyeballs everywhere which actually is appropriate for a tool created by a company whose main profit center is Advertising. (relate to previous post 'The Cookie Conundrum')
posted by oneswellfoop at 11:08 AM on July 9, 2015 [2 favorites]


Oh, crap, iotic got it first. Or maybe I'm just seeing a pattern where there is none?

/mind asplode
posted by I-baLL at 11:09 AM on July 9, 2015


Previously: teaching the machine to hallucinate

Code for generating your own: Dockerized web app that processes images and videos (much eaiser than installing caffe etc.), DeepDream Animator for creating animations.
posted by jjwiseman at 11:09 AM on July 9, 2015 [4 favorites]


If it's been trained on Internet images, why isn't it full of cats?

Well, let me read the programmer's statement. Oh, there. That's why. Let me quote a section:

Much eyes. So scare. Wow. So mystery.
posted by maxsparber at 11:09 AM on July 9, 2015 [11 favorites]


So, we know the answer now .... Android dreams of electric dogs, eh?
posted by TheLittlePrince at 11:12 AM on July 9, 2015 [2 favorites]


Omg there's a gangbang on the second link O.O
posted by sexyrobot at 11:15 AM on July 9, 2015




I think he asked if we like biscotti.
posted by maxsparber at 11:17 AM on July 9, 2015 [1 favorite]


JoeZydeco: "Wouldn't it be something if all this stuff led to a confirmation that psychedelic substances work by making your brain's neural network run backwards, just like this software?"

Or may be psychadelics make neurons fire randomly and our brain just tries hard to make sense of the random firings.
posted by TheLittlePrince at 11:19 AM on July 9, 2015 [8 favorites]


Don't eat anything with a face ... aaagh!
posted by scruss at 11:21 AM on July 9, 2015


Revealing that all our mythology, all our philosophy, all our understanding, are simply variants on sea lions...
posted by smidgen at 11:22 AM on July 9, 2015 [1 favorite]


especially the food ones.

Ripley, I think we're gonna need a bigger flamethrower.
posted by prinado at 11:40 AM on July 9, 2015 [3 favorites]


simply variants on sea lions...
GO AWAY SEA LIONS.
SERIOUSLY.
SERIOUSLY.
posted by oneswellfoop at 11:46 AM on July 9, 2015


I wish you could change the setting on this thing from "Melty Dog" to "Lisa Frank."
posted by Metroid Baby at 11:52 AM on July 9, 2015 [12 favorites]


I am too lazy to actually fool with installing a python package and dinking around with it. But I have this horrible idea, which I will share here in the hopes that someone else decides to run with it.

1. Train a neural network on a new set of images derived solely from pornography.
2. Run it backwards on some actual porn.

It is my hope that this will result in images that are titties all the way down instead of horrible dog-worm-beast entities. I mean these too will probably be horrible Lovecraftian conglomerations of body parts.
posted by egypturnash at 12:09 PM on July 9, 2015 [9 favorites]


Or may be psychadelics make neurons fire randomly and our brain just tries hard to make sense of the random firings.

Or maybe this is what the world actually looks like and Deep Dream and psychedelics strip away that veil of delusion...
posted by frimble at 12:19 PM on July 9, 2015 [5 favorites]


I'm pretty sure that's how you summon the Diana of Ephesus to the prime material plane, egypturnash.
posted by prize bull octorok at 12:19 PM on July 9, 2015 [2 favorites]


titties all the way down

Gigertty Gigertty
posted by Potomac Avenue at 12:32 PM on July 9, 2015 [3 favorites]


I saw one on BoingBoing yesterday that started with noise and iterated. Near the end there were things that looked like Combine foot-soldier faces.
posted by infinitewindow at 12:35 PM on July 9, 2015 [3 favorites]


I challenge you to convince me that this is a bad idea, prize bull octorok.
posted by egypturnash at 12:45 PM on July 9, 2015 [1 favorite]


I tried the IPython notebook the other day, but I still didn't install it all correctly, the next day someone on my FB posted the Docker link, but then that's a whole other thing I have to deal with. Looking into it, it seems like I didn't get numpy and scipy installed yet, so I was missing the pre-reqs.

Anyways, how DO you train this thing? Because yes - it is sad that it only, apparently, comes already trained on certain things. And it's great, don't get me wrong, but I would like to be able to generate my own deepdreamREALITYTUNNEL, thanks.

Then we need to link up all our ANN's to each other to create an even deeper dream.

INCEPTIONONON.

Also? Boobs all the way down? I do art like that. Though it ends up way more abstract. But if you really want, I can do that for you. I like boobs.
posted by symbioid at 12:47 PM on July 9, 2015


I've been looking at this stuff for days, and while body parts arranged impossibly (even descriptions thereof) is something that triggers a visceral panic reaction in me, somehow, these are pleasant and soothing. I looked at it expecting to have to back away, or worse, to be haunted by images of it every time I closed my eyes for weeks (like happened when I started looking at trypophobia-related images and didn't stop.)

Instead, I know that the images should normally freak me out, but don't, and am enjoying them immensely and grabbing the IPython bundle to see what I can make with it myself.
posted by frimble at 12:52 PM on July 9, 2015


More realistically: as an artist who likes to use her computer to extend her powers, Deep Dream feels like something full of potential. Right now it is feeling a little banal; "oh hey look I put the dog nose filter on another image" and "oh look it found a bunch of buildings in the compression artifacts of a nearly blank image" gets old surprisingly fast. It's almost as comedically oldmeme as sticking the Mandelbrot set into everything.

But what happens when this matures into a thing one can control with some finesse? What happens when a non-programmer has an easy-to-control set of switches they can run on an image, or parts thereof? What happens when there's a tool that lets me type a keyword in, which it then trawls google images/flickr/whatever for pictures of, trains a neural network to look for it, then lets me deep dream for those particular things?

Feed it a bunch of pictures machinery and bones; do you get a H.R.Giger filter? What do you get when you train it on photos of Gundam models and run it on a photo of a lizard? Take these results, chop them up, move them around, make something even more beautiful and strange by iterating on it with your trained human brain.

I am not sure I am the artist who will explore this. But I hope someone is.
posted by egypturnash at 12:57 PM on July 9, 2015 [6 favorites]


I hope they have a 15 minute snail sequence on Hannibal processed this way.
posted by ian1977 at 1:04 PM on July 9, 2015


I saw one on BoingBoing yesterday that started with noise and iterated. Near the end there were things that looked like Combine foot-soldier faces.

My favourite aspect of that video is that if you watch it intently and then pause it, or let it end, the stretched membrane of the image seems to flex, as if the video is still moving. Or breathing.
posted by frimble at 1:10 PM on July 9, 2015


Or may be psychadelics make neurons fire randomly and our brain just tries hard to make sense of the random firings.

Far from it. Some of the geometrical imagery generated by psychedelics actually reveals the underlying structure of our visual neurobiology by causing ordered waves of activity in the visual cortex. The imageries produced are called form constants. This can be (and has been) analysed mathematically, so we have a very good idea of what's going on without any need for randomness or 'interpretation'.
posted by xchmp at 1:10 PM on July 9, 2015 [5 favorites]


How about a real time app that does this for Google Glass????? ACK MY BRAIN BROKE
posted by ian1977 at 1:15 PM on July 9, 2015 [3 favorites]


EVERYTHING IS EYES
posted by augustimagination at 2:21 PM on July 9, 2015 [2 favorites]


Omnomnom: "Yes, why so many Schnauzers?"

If I had a buck for every time I heard this...
posted by Splunge at 2:33 PM on July 9, 2015


Whoa.
posted by homunculus at 2:56 PM on July 9, 2015 [7 favorites]


"I am too lazy to actually fool with installing a python package and dinking around with it. But I have this horrible idea, which I will share here in the hopes that someone else decides to run with it.

1. Train a neural network on a new set of images derived solely from pornography.
2. Run it backwards on some actual porn.

It is my hope that this will result in images that are titties all the way down instead of horrible dog-worm-beast entities. I mean these too will probably be horrible Lovecraftian conglomerations of body parts.
"

Here's the NSFW link that you're looking for:

http://motherboard.vice.com/read/what-computers-dream-of-when-they-look-at-porn-nsfw
posted by I-baLL at 2:59 PM on July 9, 2015 [4 favorites]


I have a theory that dogs are fundamental to the universe, and this is just revealing the truth to us.
posted by moonmilk at 3:11 PM on July 9, 2015 [2 favorites]


Whoa.

Wait until you see the vet bill.
posted by a lungful of dragon at 3:28 PM on July 9, 2015


Please ignore my ignorance but... Is this of any value other than creating weird images that quite frankly look almost all the same to me. It's like some fad/meme photoshop filter that everybody is using...
posted by njohnson23 at 3:28 PM on July 9, 2015


Well, it's having a big impact on the rare Pepes market
posted by prize bull octorok at 3:35 PM on July 9, 2015


What I haven't seen a lot of discussion of is how hard it is to train your own model, how many images and how much processing time is required, etc. The dog / eye / shoggoth filter it currently does is fine, but it gets repetitive. I want real horrors, which means I want to train it on the google image search results for "chitinous", "cockroach", too many legs", "centipede", "giant isopod", "deep sea horror", "bobbit worm", and "piers morgan".

What do I need to do to make this happen?
posted by Joakim Ziegler at 4:02 PM on July 9, 2015 [4 favorites]


I-baLL: "Here's the NSFW link that you're looking for:

http://motherboard.vice.com/read/what-computers-dream-of-when-they-look-at-porn-nsfw
"

That's actually just the normal dog / eye filter run on porn. Which is amusing, but not as recursively dream-within-a-dream (my dreams, anyway) as adding more porn to porn.
posted by Joakim Ziegler at 4:09 PM on July 9, 2015


It seems like every day people are releasing easier-to-use versions of the software. I've got my fingers crossed that training will be made easy soon. I'm told that it can take weeks to train a network, though!
posted by moonmilk at 4:10 PM on July 9, 2015


Here's the NSFW link that you're looking for:

That's just the same "dog eye" model looking at porn. What egypturnash is talking about is retraining the model using porn images, and then looking at other images such that the model will produce bits of porn. To retrain you're going to need a lot of images, as many GPUs as you can get, and some time.

Fortunately, for this purpose, I don't think you need to use supervised (labeled training data) training because you don't really care about the categories, just that there are some. So just deciding on the number of categories in your output vector is enough. In fact, it might make the images much more abstract and interesting.
posted by smidgen at 4:10 PM on July 9, 2015 [1 favorite]


JoeZydeco: "Wouldn't it be something if all this stuff led to a confirmation that psychedelic substances work by making your brain's neural network run backwards, just like this software?"

It's not exactly backwards. It's more of a feedback loop in neural circuits designed to recognize specific things. Which honestly is not a terrible theoretical model for visual hallucinations, I think.
posted by Joakim Ziegler at 4:11 PM on July 9, 2015


It doesn't matter if it takes weeks to train a network, as long as networks can be fairly easily shared, which it seems they can (the bat-country module talks about downloading pre-trained models). If someone makes easy to use wrappers for training (point at a directory of images, point at a Google Images search, etc.), I expect models for all sorts of stuff to start popping up.
posted by Joakim Ziegler at 4:13 PM on July 9, 2015


So what's the deal with dogs? Are we to conclude that DeepDream likes dogs, and not that it's been coded to like dogs? Of all the images of all the myriad things it could be "dreaming," it just stops at dogs.
posted by zardoz at 4:16 PM on July 9, 2015


This is mostly dogs and eyeballs.
posted by turbid dahlia at 4:18 PM on July 9, 2015


As I understand, it's been trained on "images of animals". I'm betting a lot of photos of animals in Google Image search's database are of dogs, and even more of them are of animals with (visible) eyes.
posted by Joakim Ziegler at 4:27 PM on July 9, 2015 [1 favorite]


More realistically: as an artist who likes to use her computer to extend her powers, Deep Dream feels like something full of potential. Right now it is feeling a little banal; "oh hey look I put the dog nose filter on another image" and "oh look it found a bunch of buildings in the compression artifacts of a nearly blank image" gets old surprisingly fast. It's almost as comedically oldmeme as sticking the Mandelbrot set into everything.

I too am really excited about the potential uses of this as people start to apply it more creatively (honestly some of the animations I've seen, where it's apparent that the recursive zoom effect on display was generated by slightly scaling the network's last output and reapplying the process on that, struck me as a step in that direction -- the product itself isn't especially interesting, but new techniques can inspire new ideas), although I'm maybe a little pessimistic about just how much labelled data would be necessary to make your own effective training sets. The notion of using image search as a source seems like a pretty good answer to that, though, in that you could search for a label and then tell the network that every image you found in your search is an instance of that label.
posted by invitapriore at 4:47 PM on July 9, 2015


turbid dahlia: "This is mostly dogs and eyeballs"

Then again, what isn't?
posted by Joakim Ziegler at 4:51 PM on July 9, 2015 [3 favorites]


Oblig. two minute obscure psych soundtrack to this thread - "Dogs in Baskets" by Geranium Pond.
posted by Devonian at 5:16 PM on July 9, 2015


OK, so maybe I don't want to try acid after all.
posted by oluckyman at 5:19 PM on July 9, 2015 [3 favorites]


I've gone on and off Nortriptyline a few times over the years, and on the days when it's building in my system or tapering off I can have this weird thing where I start dreaming before I fall asleep. I'll lie down and close my eyes, and as I'm getting drowsy I'll get this rapid-fire surreal imagery, like I'm watching a crazy movie in fast motion, projected on the inside of my eyelids. It can go on like that for hours but it will stop the instant I open my eyes. And as soon as I close my eyes again, it's crazy dream vision time.

Those waking dreams look exactly like that filtered Fear and Loathing clip. I mean, exactly, with the scratchy electric swirls and the dead staring eyes and the animals out of nowhere. I never imagined I'd see such visions in the waking world, and it gives me the goddamned creeps. Humans weren't meant to see that shit awake and sober.
posted by Ursula Hitler at 5:26 PM on July 9, 2015 [3 favorites]


One time we went to a sushi place on acid...
posted by Cookiebastard at 5:47 PM on July 9, 2015 [1 favorite]


I wish you could change the setting on this thing from "Melty Dog" to "Lisa Frank."

Metroid Baby, I don't think this is quite what you had in mind, but...
posted by Trinity-Gehenna at 7:45 PM on July 9, 2015 [5 favorites]


Wow, what is most fascinating to me is the number of people here who are saying this replicates their psychedelic experiences. I tend to think of neural nets as a useful hack for learning statistical categories, not a real model of brain function. Are we closer to AI than we think? Whichever way you interpret that question, it's kind of striking.
posted by smidgen at 9:00 PM on July 9, 2015 [3 favorites]


SO MANY EYES. WHY SO MANY EYES?
posted by bendy at 9:11 PM on July 9, 2015 [1 favorite]


Eyes Where They Ought Not Be is one of the few things that really gives me the jibblies. Gives me bad flashbacks to Cobra Commander's origin in the GI Joe movie from the 80s (and the HELL I'm linking to that scary nonsense).
posted by MrBadExample at 9:37 PM on July 9, 2015 [1 favorite]


MrBadExample, that sounds terrible! Please enjoy this soothing clip of a snake man kitten.
posted by Going To Maine at 9:46 PM on July 9, 2015 [1 favorite]


The "zomg so many eyes" thing is something I've also encountered a lot on psychedelics, and I'd always assumed this was because we're highly optimised for seeing faces and eyes in things, since it's very useful to know if, eg, a predator is looking at you, so I guess something similar applies to these images.

Without venturing too far into Thomas Nagel territory, I wonder what it would be like to take acid if you have prosopagnosia.
posted by doop at 10:59 PM on July 9, 2015 [1 favorite]


I love this tool. The images it pukes out are the closest representation of what the world looks like through psychedelics that I've ever seen.

Note to self: do not ever take psychedelics. These eyeballs-everywhere things make me want to vomit. Very similar to the phobia of holes-in-things (can't look that up now but I know MeFi knows what I mean).
posted by harriet vane at 11:57 PM on July 9, 2015 [1 favorite]


Joakim Ziegler: "What I haven't seen a lot of discussion of is how hard it is to train your own model, how many images and how much processing time is required, etc. The dog / eye / shoggoth filter it currently does is fine, but it gets repetitive. I want real horrors, which means I want to train it on the google image search results for "chitinous", "cockroach", too many legs", "centipede", "giant isopod", "deep sea horror", "bobbit worm", and "piers morgan".

What do I need to do to make this happen?
"

Slow down, Satan.
posted by Splunge at 8:05 AM on July 10, 2015 [3 favorites]


all of these organic dogs appearing, this all just reminds me of the art of my friend Stefan Thompson.
He so deep dreamed YEARS ago. The computer's just riffing on his work now. ;)
posted by Theta States at 9:11 AM on July 10, 2015 [4 favorites]


imagine if this were applied to sound identification/translation or (driving for now ;) concepts!

Watson Adds Deep Learning to Its Repertoire - "Applying learning from one area, such as vision, to another, such as speech, is known as a multimodal approach. It could make future AI systems far more useful and could yield fundamental insights into the nature of intelligence... 'If I talk to you about a dog, it's very hard to have an understanding of what a dog is without having an experience of that dog, which you get through a multimodality view of that.' "
posted by kliuless at 5:19 PM on July 10, 2015 [2 favorites]


I'm going to regret this, but... has anyone put a Minion through this treatment?
posted by harriet vane at 1:48 AM on July 11, 2015


DeepDream is going to define the next couple of years as much as Kai's Photo Tools defined a 90s visual style.
posted by quartzcity at 3:02 PM on July 13, 2015




AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAH
posted by Going To Maine at 6:51 PM on July 13, 2015


(Well, I don't keep either dogs or eyeballs down there.)
posted by Going To Maine at 6:52 PM on July 13, 2015


Square.

(Vanilla?)
posted by Joseph Gurl at 7:31 PM on July 13, 2015


Zhora our Swedish vallhund doing a jump in Nightmare Agility. I am especially happy with the chicken head in her tail.
posted by ROU_Xenophobe at 8:29 AM on July 17, 2015


We Hunted The Mammoth is going off its core mission, but that's okay: Deep Dream Reveals the Shocking Truth About Donald Trump’s Hair
posted by Going To Maine at 11:56 AM on July 17, 2015


Another service to create your own: http://deepdre.am/
posted by jjwiseman at 12:47 AM on July 18, 2015


In reply to my comment above about how Stefan's thompson's art was deep dreaming long before google, I submit exhibit A:
What happens when we DeepDream his artwork? His animals just wake up.
posted by Theta States at 9:04 PM on July 20, 2015 [1 favorite]


That's pretty much what happens to lots of images, though. 80% of the ones I ran through that site did pretty much exactly that.
posted by Joseph Gurl at 3:45 AM on July 21, 2015


« Older Will Graham would never, ever judge you for...   |   I Was An Invisible Girlfriend and Boyfriend Newer »


This thread has been archived and is closed to new comments