This was the result.

http://i.imgur.com/KMG1kN3.jpg

I regret nothing.

It's path of learning. Follow my FA if this kind of weird stuff interests you.

Wow, Nice, for me it's looks pretty similatr to google deep dream. Keep making these .

Maybe one day this ai will be able to draw furry porn on it's own. :D

So, uh, what do random pictures have to do with AIs and bots?

Thank you, Hexdragon, I intend to continue. And don't worry Qmannn, I'm just a nerd that knows how to copy much smarter people's code.

Furrin, on my FA I give an explanation. Basically, the images you see were generated using a machine learning algorithm similar to Google's Deep Dream or the likes. It's definitely not strong AI, but one could argue it is a form of learning and/or intelligence.

This is a neat project.

I've been debating trying to train an AI system in the classification of tags for images based on E6, but I was having trouble getting the required computing power to make it even remotely feasible. I'm hoping that at some point the GPU requirements for running things like image classification with neural networks against images will come down.

KiraNoot said:
This is a neat project.

I've been debating trying to train an AI system in the classification of tags for images based on E6, but I was having trouble getting the required computing power to make it even remotely feasible. I'm hoping that at some point the GPU requirements for running things like image classification with neural networks against images will come down.

Actually it's pretty feasible nowadays with a highish end GPU and a few days/weeks time. I already did some experiments with making a tagging AI using the VGG-19 network as a basis, and I calculated I could probably train it in two weeks time. Though I made some heavy assumptions such as only training on tags with more than 1000 images attached to it, otherwise I fear the accuracy would suffer. I wanted to do that first, but tagging is less fun than generating images. Once/If I get around to it, I could donate it to you guys if you'd like.

deepfur said:
Actually it's pretty feasible nowadays with a highish end GPU and a few days/weeks time. I already did some experiments with making a tagging AI using the VGG-19 network as a basis, and I calculated I could probably train it in two weeks time. Though I made some heavy assumptions such as only training on tags with more than 1000 images attached to it, otherwise I fear the accuracy would suffer. I wanted to do that first, but tagging is less fun than generating images. Once/If I get around to it, I could donate it to you guys if you'd like.

Also, not sure if you are the one to report to, but I think the before_id parameter at /post/index.xml is broken.

It does appear to be broken. It looks like the .xml template expects actual pages for some reason. I'll look into fixing it. The JSON API is the preferred API for sure, both for transport size and consistency.

KiraNoot said:
It does appear to be broken. It looks like the .xml template expects actual pages for some reason. I'll look into fixing it. The JSON API is the preferred API for sure, both for transport size and consistency.

Noted, will use it from now on.

AI tagging of images is the future of e621. Keep it up.

AI tagging images will most likely result in either one of two things (or possibly both)

A) Tagging arguments cease to exist, or at least show up at much heavily decreased rairty

Or

B) People argue tags with an Artificial Intelligence. Oh boy, those would be some wild times.

tl;dr for better or worse might as well try out AI tagging when stuff happens or shit lul

This is pretty nice, but it got a lot more learning to do. Computers suck at image recognition and generation, and before you say "no they don't" of course they can do basic stuff, outlines, abstract art and similar. But they got nothing against humans for anything else. A mediocre drawing made in 1 second isn't worth anything compared to a fantastic drawing made in 10 hours.

deepfur said:
... Though I made some heavy assumptions such as only training on tags with more than 1000 images attached to it, otherwise I fear the accuracy would suffer.

That's too bad, on of the charms of our tagging system is that we got all those small specific tags, but those are also the hardest/most tedious to tag and hence least tagged, which is why I'd want a system explicitly for adding the more obscure tags which humans are less inclined to.

Also if e621 ever gets AI tagging I hope it uses different database entries for that, something like "suggested tags", I don't yet trust AIs doing a lot of tasks.

Oracle_of_Pelor said:
A) Tagging arguments cease to exist, or at least show up at much heavily decreased rairty

Or

B) People argue tags with an Artificial Intelligence. Oh boy, those would be some wild times.

I don't think arguments would decrease rather the opposite. The difference would just be that they would most likely end sooner and less people on one side.

But it's always fun arguing with AIs, some are just so good that you sometimes have to exchange words a few times before you realize something is wrong.

Chessax said:
This is pretty nice, but it got a lot more learning to do. Computers suck at image recognition and generation, and before you say "no they don't" of course they can do basic stuff, outlines, abstract art and similar. But they got nothing against humans for anything else.

A better explanation would be 'humans suck at understanding how image recognition and generation works. But because they find it easy to see, they wrongly assume that the problem of vision is easy.'
(they also tend to assume that vision can be taught quickly to computers, even though it took them years to really get a solid grip on seeing.)

A central issue here is the need to reduce the input images to characteristics that accurately represent the image per human perception, rather than characteristics that don't. That includes tasks like analyzing overlaps to arrive at a simplified representation of the z-levels in the image, reducing the full image histogram into a small representation of the image palette, reducing the prominent edges into vectors so that the system you're training can readily identify patterns / shape themes in it. Most of these are ongoing problems that we've only partly solved.

That's too bad, on of the charms of our tagging system is that we got all those small specific tags, but those are also the hardest/most tedious to tag and hence least tagged, which is why I'd want a system explicitly for adding the more obscure tags which humans are less inclined to.

Yeah, it's pretty much unavoidable. Of course you could take the 'self-correcting'/incremental approach (accept training sets that are a little small, and then track the images tagged to see whether the tag is subsequently removed), But you'd still need a minimum size for the training set.

Also if e621 ever gets AI tagging I hope it uses different database entries for that, something like "suggested tags", I don't yet trust AIs doing a lot of tasks.

I agree in principle but feel I must note that this would complicate and slow searching, unless you create a third table which combines both user tags and suggested tags.

One question I have since I don't really have too much insight into image generation. How good could an AI of this kind (or similar at least) actually get with enough practice? And additionally if it could generate one "good" image how many "bad" images would come before it? I know, really difficult questions to answer, but still... Also I don't consider Google's Deep Dream software to be a valid comparison of what it could do, I would compare that to an advanced form of simple mosaic software.

savageorange said:
A better explanation would be 'humans suck at understanding how image recognition and generation works. But because they find it easy to see, they wrongly assume that the problem of vision is easy.'
(they also tend to assume that vision can be taught quickly to computers, even though it took them years to really get a solid grip on seeing.)

A central issue here is the need to reduce the input images to characteristics that accurately represent the image per human perception, rather than characteristics that don't. That includes tasks like analyzing overlaps to arrive at a simplified representation of the z-levels in the image, reducing the full image histogram into a small representation of the image palette, reducing the prominent edges into vectors so that the system you're training can readily identify patterns / shape themes in it. Most of these are ongoing problems that we've only partly solved.

Yeah probably not the best analogies I've made. It's just that, like you say, it's such a complex problem to solve that in the end even a computers best attempt can end up being outward laughable compared to humans since we already have specialized hardware developed for doing this kind of stuff with our eyes closed that takes almost no apparent effort to do, only time.

The only thing I feel AIs can successfully compete with us in a general, is speed. I believe that's one of the reasons why the advances in especially robotics and AI often seem really cool at first but then people lose interest because, in reality, it wasn't really that useful from an end-users perspective, simply another step in the right direction.

savageorange said:
I agree in principle but feel I must note that this would complicate and slow searching, unless you create a third table which combines both user tags and suggested tags.

Oh yeah, that would be perfectly fine, i.e. searching brings up both, or maybe you have a little checkbox to chose if you want to include auto generated tags. It's when AIs start tampering with "validated human data" (or the "absence” of data, which in itself is also data) or when the two data sets are valued equally that I get worried, maybe I'm just being paranoid.

Though I'm not exactly sure how one would implement this separate data in a useful manner, e.g. how to let users validate generated data in a simple and efficient manner.

Oracle_of_Pelor said:
AI tagging images will most likely result in either one of two things (or possibly both)

A) Tagging arguments cease to exist, or at least show up at much heavily decreased rairty

Or

B) People argue tags with an Artificial Intelligence. Oh boy, those would be some wild times.

tl;dr for better or worse might as well try out AI tagging when stuff happens or shit lul

AI tech most definitely isn't at the level to replace humans yet, but I think it could be use as an "assistant". For example, you upload an image, and it shows you a list of suggested tags, and you can click on the ones you think fit to add them.

Chessax said:
One question I have since I don't really have too much insight into image generation. How good could an AI of this kind (or similar at least) actually get with enough practice? And additionally if it could generate one "good" image how many "bad" images would come before it? I know, really difficult questions to answer, but still...

Actually that is pretty easy to answer in this case: Not any better. You can see that the output of the AI plateaued somewhere around epoch 40, further training didn't really make anything better, it even decreased the quality sometimes. This system is a very primitive AI, a fun little toy project I threw together in a weekend. Both the network's size and it's input data are quite small. With more data and a larger network, I do expect better quality. But even then, it doesn't do justice to just how hard a problem vision is. Humans are exposed to light and how it interacts with objects all day, for years and years, their entire lives. This is an unbelievable amount of data, and allows us to predict how light should interact with imaginary objects. The only way to break through the Deep Dream-esque abstract imagery level of quality is to create an AI that understands how light (and for that matter, objects) works. This is a very under researched field, since few AI researchers are interested in image generation (And crazy people like me don't have funding).

Chessax said:
Yeah probably not the best analogies I've made. It's just that, like you say, it's such a complex problem to solve that in the end even a computers best attempt can end up being outward laughable compared to humans since we already have specialized hardware developed for doing this kind of stuff with our eyes closed that takes almost no apparent effort to do, only time.

The only thing I feel AIs can successfully compete with us in a general, is speed. I believe that's one of the reasons why the advances in especially robotics and AI often seem really cool at first but then people lose interest because, in reality, it wasn't really that useful from an end-users perspective, simply another step in the right direction.

At the moment, absolutely. Generational AI is in it's complete infancy, I consider this to be a novelty toy more than anything.

Chessax said:
Also I don't consider Google's Deep Dream software to be a valid comparison of what it could do, I would compare that to an advanced form of simple mosaic software.

Calling Deep Dream an advanced form of mosaic software is, with all due respect, a vast understatement. It models the primate visual cortex very well and can even outperform humans in classifying natural images.

deepfur said:
Calling Deep Dream an advanced form of mosaic software is, with all due respect, a vast understatement. It models the primate visual cortex very well and can even outperform humans in classifying natural images.

Fair enough, I guess I just wasn't that impressed, to be honest your "novelty toy" almost impressed me more :P

Oracle_of_Pelor said:
B) People argue tags with an Artificial Intelligence. Oh boy, those would be some wild times.

tl;dr for better or worse might as well try out AI tagging when stuff happens or shit lul

C) Roko's tagging basilisk:

The AI takes over e621 and bans/kills you based on the tagging arguments you're having right now.

Now we don't need to manually rate pics anymore:
https://tech.slashdot.org/story/16/09/30/2018253/yahoo-open-sources-a-deep-learning-model-for-classifying-pornographic-images

Munkelzahn said:
Now we don't need to manually rate pics anymore:
https://tech.slashdot.org/story/16/09/30/2018253/yahoo-open-sources-a-deep-learning-model-for-classifying-pornographic-images

Give it a rating:s for 0-0.24, rating:q for 0.25 to 0.75, rating:e for 0.76+. Except we don't know that it would do well on drawn fantasy/furry pics, and the included examples are too damn safe.

We probably have thousands of pictures that could score a 0.999 and still get a questionable rating:

post #1006347 post #1004391 post #988268 post #921177 post #847969 post #1010421 post #1010322 post #977500 post #924337 post #925058

Ball's in your court, Yahoo! e621 wants to talk to you, Jay Mahadeokar, Gerry Pesavento, Sachin Farfade, Amar Ramesh Kamat, Armin Kappeler, and Shraddha Advani.

Lance_Armstrong said:
stuff with sexy pictures

Thanks for another image to fave.

The problem with artificial intelligence is once it becomes self aware, scientists (or programmers in this case) are certain that the following is at least 85% going to happen:

If the AI starts questioning its existence, it will shut down or try to take over the world.

Sources: Isaac Asimov

RogueLizard said:
If the AI starts questioning its existence, it will shut down or try to take over the world.

Sources: Isaac Asimov

Yeah, but don't the Laws of Robotics prohibit that?

True, said laws haven't been used in a robot yet, but still

Ryuzaki_Izawa said:
Yeah, but don't the Laws of Robotics prohibit that?

True, said laws haven't been used in a robot yet, but still

The Laws of Robotics were an excellent plot device used to help Asimov pump out many interesting short stories. Beyond that, it shouldn't be assumed what smart AI or AI forced to be "friendly" will actually do with humanity.

Lance_Armstrong said:
The Laws of Robotics were an excellent plot device used to help Asimov pump out many interesting short stories. Beyond that, it shouldn't be assumed what smart AI or AI forced to be friendly will actually do with humanity.

Well, fuck.

But wait, would there any way one might conceivably 'hard-program' core concepts into a sufficiently advanced artificial intelligence? Sort of like 'hard' memories, like how one doesn't usually just forget how to breathe or stand up.

Ryuzaki_Izawa said:
But wait, would there any way one might conceivably 'hard-program' core concepts into a sufficiently advanced artificial intelligence? Sort of like 'hard' memories, like how one doesn't usually just forget how to breathe or stand up.

It's an active area of research and a philosophical as well as technical question:

https://en.wikipedia.org/wiki/Friendly_artificial_intelligence (read all of this one)
https://en.wikipedia.org/wiki/Machine_Intelligence_Research_Institute
https://en.wikipedia.org/wiki/OpenAI

Ryuzaki_Izawa said:
Yeah, but don't the Laws of Robotics prohibit that?

True, said laws haven't been used in a robot yet, but still

If anything Asimov's stories were made to demonstrate that the laws rather exactly do NOT work.

What if a friendly AI forces all humans to engage in homosex (for their own good)

Munkelzahn said:
What if a friendly AI forces all humans to engage in homosex (for their own good)

Then it will be business as usual for you! (sick burn, brah)

GDelscribe said:
If anything Asimov's stories were made to demonstrate that the laws rather exactly do NOT work.

"I, Robot" for example, the Mother AI core decides to lock humans into their houses and prohibit any social interactions, even going as far as (violently) detaining humans to force them inside, believing that because of a few violent interactions, any and all interactions are too much of a risk.

Furrin_Gok said:
"I, Robot" for example, the Mother AI core decides to lock humans into their houses and prohibit any social interactions, even going as far as (violently) detaining humans to force them inside, believing that because of a few violent interactions, any and all interactions are too much of a risk.

i bet she outlaws first-person shooters, too

This is all I could think of upon looking at this thread.