Can artificial brains dream about ponies? aka Science needs your OC

[SasQ 1,379]

Janelle Shane, a research scientist, recently trained an artificial neural network (artificial brain, if you wish) on the data of over 1500 pony names of ponies from the MLP Wiki to see if it will be able to generate new pony names. She shows the result of this experiment on her Tumblr. Some of the names are really crazy but most of them are really funny (My personal favourite was Raspberry Turd )

She brought some of the ponies from the list to life by using GeneralZoi's Pony Creator, and she shows some of them in her Twitter post here:

as well as on her Tumblr page.

At first I thought that she trained her neural network to generate ponies so I was a bit disappointed when I discovered that it was just the names. But then I had an idea: why don't we try to train the artificial brain to actually generate the pictures of ponies? 

You think that it's impossible? Well, not at all! Because, in case you didn't know it, the Pony Creator has an interesting feature in the Advanced tab: it generates a special string of letters and digits, so called ponycode, that succintly stores all the information needed to recreate that pony design again by copy-pasting it into the text field in Pony Generator :> Here's an example pony code that I came up with:

313W000100FE7ABA9E9EFE1010137BC3BN3183700H200020213FCCC7FF7FFF06107F3FCC004CB2

When you copy this ponycode and paste it into the text area in the Pony Creator, you should see a crazy pink Unicorn pony with short cyan mane

Now, neural networks are good at pattern recognition and fiddling with numerical data, so the ponycode is a great representation they can work with. But the neural network requires a lot of training data with example ponycodes to learn how to make good-looking OCs.

And here's where you can help the Science :>

I already contacted Janelle on Twitter and asked about the ponycode trick, and here's what she replied:

So are you ready to help the Science?  What you need to do is to come up with a cool design of a pony in Pony Creator, and then copy its ponycode from the Advanced tab and post it in this thread, along with its name (if you named it)   If you already have your OC ready along with its ponycode and you don't mind it participating in the experiment – even better If you can post multiple designs in one post – better still!

I'll tell Janelle about this thread so that she could collect the data on the go and feed it into the NN.

Just please, make sure your OC will keep up with the standards of the show. If it looks like other ponies you could see around trotting in Equestria, then it's fine. But if it is some crazy design with eye-hurting colour combinations – add an according annotation! (e.g. "BAD OC" or "CRAZY")  Perhaps such data could be useful too, for a separate experiment (e.g. training a neural network to come up with terrible OCs that might be interesting too – how terrible it could be ). But for the main part of the experiment, we better don't let the artificial brain to train on these :q

Edited September 17, 2017 by SasQ
Minor edits.

[Techno Universal 2,575]

1 hour ago, SasQ said:

Janelle Shane, a research scientist, recently trained an artificial neural network (artificial brain, if you wish) on the data of over 1500 pony names of ponies from the MLP Wiki to see if it will be able to generate new pony names. She shows the result of this experiment on her Tumblr. Some of the names are really crazy but most of them are really funny (My personal favourite was Raspberry Turd )

She brought some of the ponies from the list to life by using GeneralZoi's Pony Creator, and she shows some of them in her Twitter post here:

as well as on her Tumblr page.

At first I thought that she trained her neural network to generate ponies so I was a bit disappointed when I discovered that it was just the names. But then I had an idea: why don't we try to train the artificial brain to actually generate the pictures of ponies? 

You think that it's impossible? Well, not at all! Because, in case you didn't know it, the Pony Creator has an interesting feature in the Advanced tab: it generates a special string of letters and digits, so called ponycode, that succintly stores all the information needed to recreate that pony design again by copy-pasting it into the text field in Pony Generator :> Here's an example pony code that I came up with:

313W000100FE7ABA9E9EFE1010137BC3BN3183700H200020213FCCC7FF7FFF06107F3FCC004CB2

When you copy this ponycode and paste it into the text area in the Pony Creator, you should see a crazy pink Unicorn pony with short cyan mane

Now, neural networks are good at pattern recognition and fiddling with numerical data, so the ponycode is a great representation they can work with. But the neural network requires a lot of training data with example ponycodes to learn how to make good-looking OCs.

And here's where you can help the Science :>

I already contacted Janelle on Twitter and asked about the ponycode trick, and here's what she replied:

So are you ready to help the Science?  What you need to do is to come up with a cool design of a pony in Pony Creator, and then copy its ponycode from the Advanced tab and post it in this thread, along with its name (if you named it)   If you already have your OC ready along with its ponycode and you don't mind it participating in the experiment – even better If you can post multiple designs in one post – better still!

I'll tell Janelle about this thread so that she could collect the data on the go and feed it into the NN.

Just please, make sure your OC will keep up with the standards of the show. If it looks like other ponies you could see around trotting in Equestria, then it's fine. But if it is some crazy design with eye-hurting colour combinations – add an according annotation! (e.g. "BAD OC" or "CRAZY")  Perhaps such data could be useful too, for a separate experiment (e.g. training a neural network to come up with terrible OCs that might be interesting too – how terrible it could be ). But for the main part of the experiment, we better don't let the artificial brain to train on these :q

Yeah maybe my OC Techno would be a good subject for the experiment! Though even though I don't have a pony code for him his character page is still linked in my signature!

[SasQ 1,379]

That's nice, but I don't think that the neural network could make any use of just a picture. It's not that advanced to figure out a pony from a picture. That's why the ponycodes are mandatory. Sorry :q  This is the only thing this particular neural network can work with.

But you can always try to recreate your OC in Pony Creator and generate the ponycode for it

Edited September 17, 2017 by SasQ
How to solve it

[Quantum Pony 40]

Would this work? 

2S2S000000E8E8E8BCC99F0020100BB96UN183700000000000CC6868FF7FFF0200CC6262004CB2

As for the name I'm thinking of either Strawberry Tart or John Goodman  

 

[SasQ 1,379]

If you mean the light-gray pony with chocolate hair and aqua eyes, then yes, it works fine Thanks for the data.

As for the name, I would go with the Tart one, because John Goodman is already taken by a certain actor

(BTW nice state vector cutie mark in your profile pic can it be in superposition of states too? )

Edited September 17, 2017 by SasQ
PS

[Flutterlina 21]

Can we try to recreate actual show ponies and post their pony codes?

[SasQ 1,379]

I don't have any objections against that. But I guess we would have to ask Janelle about that (or perhaps drag her into this topic?). I'm not sure how the copyright issues would go with it, but I think there shouldn't be any problems with that, since ponycodes are not Hasbro's property and they are not the actual ponies themselves, just a meta-description in a textual form (something like a source code in programming, or a recipe for baking a cake) And as far as I can tell, a recipe is certainly not the cake Similarly the phrase "A lavender Unicorn with straight dark-blue mane and tail, with pink and violet strips" is not copyrighted either

So maybe for future reference, just mark the ponycode with a tag, like "SHOW" for example, so we could sort the data accordingly and either include such ponies to the dataset or not.  Hashtags so far:

• SHOW           =  a pony taken directly from an episode of the original show
• OC                =  original chacater not appearing in the show
• PONYSONA  =  someone's personal representation in a pony form (implies OC)
• CRAZY          =  weird, crazy, ugly, wacky OC that would never appear in a show  (or should I rename it to GONZO? )
• PONIFIED     =  A ponified version of a real-world character or a character from another universe than MLP.

I also consider adding a hashtag ALICORN, because I suspect there might be quite a number of them

Edited September 17, 2017 by SasQ
Added one more category.

[Quantum Pony 40]

Heres another OC . 

2S2S000000ADEDC7FFC49D000016F00FEUN1837005100000703FABCCFF7FFF00107F3FCC004CB2

This is good practice for making OCs. I feel like the lighter the color of the body (relative to the mane) the closer it looks like the ponies in the show =]. 

2 hours ago, SasQ said:

(BTW nice state vector cutie mark in your profile pic can it be in superposition of states too? )

Your the first pony that noticed!  . And it sure can be a superposition! I defined it to be:

|ψ> = 1/√3  ( | >+ |  > + | > ) 

[SasQ 1,379]

1 hour ago, Quantum Pony said:

I feel like the lighter the color of the body (relative to the mane) the closer it looks like the ponies in the show =]

Your neural network figured it out right   Now I wonder if Janelle's artificial brain will figure it out as well ;>

1 hour ago, Quantum Pony said:

Your the first pony that noticed!

Well, it's because I'm a quantum freak too But with a slightly different approach – I'm not that convinced that probability is the right way to go. I'm rather a follower of the original quantum theory, by Balmer, Schrödinger, Clifford, Wolff etc. in which there are no point particles, only waves of space and their geometry. My ponysona's cutie mark kinda reflects that.

                                                                                                                                      

OK here are some of my OCs I once made for Pony.Town:

Grass Seed  # OC
3830000000B8FE4DFEE89E00201E28239UN1837000200002Y04EAF27FF7FFF1Q107F3FCC004CB2
(I was using her to walk around and plant grass, or make paths through forests, or grass labyrinths for other ponies to play but they usually didn't last long – too many trolls :q )

Double Rainwave  a.k.a. Constructive Interference  # PONYSONA
3T390P410000D0E24F342302000773D3DUN1837508000021G14F2F2F222E4C0P107F3FCC004CB2

Ancient Glyph  # OC
3H2L0D4100EDED9FCEC778020007BAA13MI1911032000001N13D29262D271E32107F3FCC004CB2
A polyglot pony who studies ancient languages and writing systems.

Doctor Manestein  # PONIFIED
3C15074000F4FE9EFFC49D000013B61BFUN393B13E010202D1DEE2FEBDD5D800107F3FCC004CB2
A ponified version of Albert Einstein. This pony is an old physicist working on relativity and time travel

Pictures attached below.

 

Edited September 17, 2017 by SasQ
Corerected the formatting of pictures.

[Skytower 71]

I get the feeling that sooner or later something like this can be used in a VR program where you can put on a helmut and walk around Equestria as a pony and talk and interact with the inhabitants as if they were real.

[Techno Universal 2,575]

15 hours ago, SasQ said:

That's nice, but I don't think that the neural network could make any use of just a picture. It's not that advanced to figure out a pony from a picture. That's why the ponycodes are mandatory. Sorry :q  This is the only thing this particular neural network can work with.

But you can always try to recreate your OC in Pony Creator and generate the ponycode for it

Yeah I'll try to put him in pony creator if I can!

[Quantum Pony 40]

23 hours ago, SasQ said:

Well, it's because I'm a quantum freak too But with a slightly different approach – I'm not that convinced that probability is the right way to go. I'm rather a follower of the original quantum theory, by Balmer, Schrödinger, Clifford, Wolff etc. in which there are no point particles, only waves of space and their geometry. My ponysona's cutie mark kinda reflects that.

That sounds like quite an interesting approach , would love to hear more about it!  . 

The Einstein pony was perfect! You even got the wrinkles right! 

(I tried making a Paul Dirac pony but his asymmetric hairstyle makes it impossible  ) 

Here is an Artist conception 

2S2S006000EEF2C4FFC49D0000100A1FEUN1837108000000C0000000FF7FFF00107F3FCC004CB2

[DJ_Bonebraker 183]

Here's mine: 

Radar Pulse #Ponysona:  3C3C006020AEB9C1FFC49D0040100BBFEUN1837000000000Y10062FEFED50003107F3FCC004CB2

(this is Radar when she's in girl-mode.  In boy-mode, he tends to wear a suit coat, hat & have less fancy mane & tail)

[SasQ 1,379]

@DJ_Bonebraker and @Quantum Pony: thanks for your input
I must inform you, though, that accessories and poses have separate codes, so if you want them too, you need to post the complete information (three codes). I'm not sure though if this would be necessary – accessories are not part of the pony, they can change (with rare exceptions), and the same goes for the pose. So I guess the ponycode for the pony's body should be enough for the neural network to work with

OK, we've got as much (:P) as 8 ponycodes so far, 5 of which were mine :q  That's definitely too few data for training a neural network It's a good start, but we have to try harder :q  I made some more ponies below, this time they're the characters from the show.

Here's all the main six:

Twilight Sparkle   #SHOW:
392S000100D19FE3FFC49D00101622D85UH1D3C000000000P1263773ED438A0N107F3FCC004CB2

Pinkie Pie   #SHOW :
3D2S000000F4B6D1FFC49D002017CCBEDTN1837000000000J0ED458D204E970I107F3FCC004CB2

Rainbow Dash  #SHOW :
3C2P0900209EDBF9FFC49D00001C6006FUN1E33000000000O07F3FCCFF7FFF0M107F3FCC004CB2

Fluttershy   #SHOW :
353B000010FEFAA3FFC49D003011FBAB6UN3D39004000000G0FFB9DBFF7FFF0G107F3FCC004CB2

Rarity    #SHOW :
3B27000100EDF3F4FFC49D005013767AFUN1D3C000000000L04C388EFF7FFF0K107F3FCC004CB2

Applejack   #SHOW :
2S2S000000F7AF54FFC49D0040154C13EUN1837000100000I0FEED94FF7FFF0H107F3FCC004CB2

And some background ponies...

Here's the closest I could get to Octavia Melody  #SHOW :
3D2S000000C4C4B8FFC49D0010176418FON183740000000330464646FF7FFF1Z107F3FCC004CB2

and here's Vinyl Scratch  #SHOW :
066CC6696F4085066CC66066CC66066CC6604E8FBAFFFF8C066CC66066CC66066CC66066CC66

Here's Lyra Heartstrings   #SHOW :
3D2S00010073F0D6FFC49D00001FFC52DTN1837000000000N192DCE8FFFFFF1B107F3FCC004CB2

and the closest I could get to Bon Bon a.k.a. Special Agent Sweetie Drops   #SHOW :
3D2S000000ECF1C2FFC49D00001159CB6TN1837000000001Y1FF8DDF204E971U107F3FCC004CB2

Seriously though, at that rate, we will get to our first hundred of ponies in two years :q
Maybe we should post it on Equestria Daily or something? :q

---

On 18.09.2017 at 6:04 PM, Quantum Pony said:

That sounds like quite an interesting approach , would love to hear more about it!  .

I'll be glad to tell you some more, but I can't do it here due to some ekhm "constraints" of the Poniverse's terms of use (I think you should read them some time too :q you might be surprised). But you can mail me at cloudsdale@go2.pl  from an e-mail account that is NOT Google Mail (their terms of use is messed up too and they scan your e-mails; and I don't want some files on the wall participating in our conversations, if you know what I mean :q ).

[DJ_Bonebraker 183]

I have a couple more for you:

Aurora Shine #OC

2S3C00010000DDFEFFC49D00201FEDD00UN183700000000051CBCB3CFDFF9E0J107F3FCC004CB2

Silicon Chip #OC

3C250K4000A88763FFC49D00000BA3329UN1837000000000C0721E1EFF7FFF0C107F3FCC004CB2

[Quantum Pony 40]

On 2017/9/20 at 0:38 AM, SasQ said:

I'll be glad to tell you some more, but I can't do it here due to some ekhm "constraints" of the Poniverse's terms of use (I think you should read them some time too :q you might be surprised). But you can mail me at cloudsdale@go2.pl  from an e-mail account that is NOT Google Mail (their terms of use is messed up too and they scan your e-mails; and I don't want some files on the wall participating in our conversations, if you know what I mean :q ).

Hey, sorry for the late reply. 

No prob, I made a pony specific email account  

TwilightDiracDash@outlook.com

[SasQ 1,379]

Thanks @DJ_Bonebraker for your input

OK we've got 21 ponycodes in 12 days. At that rate we'll get to 100 in 2 months and 500 in almost a year! :q  We reeeeallly need to boost things up :q

To facilitate that, I made the following picture that shows everything you need to know about the ponycode format in case you wanted to edit it by hand or generate automatically (but not randomly, please :P), or convert some older ponycodes into the latest format.

The format is quite straightforward: it simply serializes all the data from the GUI controls one by one, top to bottom, from each tab. Numbers are encoded as Base36 (hexatrigesimal) with decimal digits 0 through 9 and alphabetical "digits" A through Z, the same way post and comment IDs are encoded on Reddit. Colours are encoded as hexadecimal RGB values, the same way as in HTML or CSS. Checkboxes or two-state radio buttons are encoded as boolean values (0=false, 1=true). Different styles of manes, bodies, tails etc. are numbered left to right, top to bottom, page after page.

I also found this list with quite a lot of ponycodes of the characters from the show:
https://darthwill3.deviantart.com/journal/Show-Accurate-Pony-Codes-for-Pony-Creator-483509252

and a tread in which people post ponycodes for their own OCs along with pictures:
https://generalzoi.deviantart.com/journal/Ponycode-Compilation-Post-254390025

There's a lot of bad OCs in there but maybe after some filtering there will remain at least 200 ponycodes that the neural network could use

If you don't want to post your OCs or help the science by coming up with new pony designs in Pony Creator, then perhaps you can search the Internet for other people's ponycodes at least? The more we can find, the better, because the neural network needs a lot of data to learn from.

 

Edited September 28, 2017 by SasQ
Updated the image