neural networking

New paint colors invented by neural network

So if you’ve ever picked out paint, you know that every infinitesimally different shade of blue, beige, and gray has its own descriptive, attractive name. Tuscan sunrise, blushing pear, Tradewind, etc… There are in fact people who invent these names for a living. But given that the human eye can see millions of distinct colors, sooner or later we’re going to run out of good names. Can AI help?

For this experiment, I gave the neural network a list of about 7,700 Sherwin-Williams paint colors along with their RGB values. (RGB = red, green, and blue color values) Could the neural network learn to invent new paint colors and give them attractive names?

One way I have of checking on the neural network’s progress during training is to ask it to produce some output using the lowest-creativity setting. Then the neural network plays it safe, and we can get an idea of what it has learned for sure.

By the first checkpoint, the neural network has learned to produce valid RGB values - these are colors, all right, and you could technically paint your walls with them. It’s a little farther behind the curve on the names, although it does seem to be attempting a combination of the colors brown, blue, and gray.

By the second checkpoint, the neural network can properly spell green and gray. It doesn’t seem to actually know what color they are, however.

Let’s check in with what the more-creative setting is producing.

…oh, okay.

Later in the training process, the neural network is about as well-trained as it’s going to be (perhaps with different parameters, it could have done a bit better - a lot of neural network training involves choosing the right training parameters). By this point, it’s able to figure out some of the basic colors, like white, red, and grey:

Although not reliably.

In fact, looking at the neural network’s output as a whole, it is evident that:

  1. The neural network really likes brown, beige, and grey.
  2. The neural network has really really bad ideas for paint names.

If I let the horse-name neural network pick the most probable character every time, it starts off by generating a bunch of distinct names, some of which are real (e.g., “Real Delight”, “Pleasant Tap” and “Sun Beau”), and some of which definitely are not (e.g., “Pert Pangle”, “Curlind Aches” and “Bew the Dance”).

But after many lines of this, the output eventually converges on a name that, as far as the model is concerned, represents the ideal name for a thoroughbred.

So what is this legendary name, you ask?

Shart Tangle.

modern creepypasta irl basically:

Let’s give a Recurrent Neural Network 10 minutes of anime dialog and tell it to learn how it works. 

after forcing this neural network to listen to and learn this anime for 1000 iterations, the output is…pretty much just empty noise. that’s not too surprising really.

2000 iterations in to the training: the network has learned how to scream.

5000 iterations: it seems to have learned how to produce something resembling chatter or mumbling. 

6000: same as above, except it spent 23 seconds screaming. again.

10000: the screams more closely match human intonation, with descending pitch at the end and breaths in between each scream, etc 

15000: it seems fixated on short laughter (“ehe”) now.

25000: finally the output is starting to resemble human speech, though not with any real words

…but really what the fuck lmao

The neural network has weird ideas about what humans like to eat

So I’ve been training this neural network to generate cookbook recipes by letting it look at tens of thousands of existing recipes.

The generated titles can get a bit odd.

There’s a creativity variable I can set when the network is generating new recipes, and when I set it low, it comes up with its best guess at the most quintessential recipe titles:

Cream Cheese Soup
Cream Of Sour Cream Cheese Soup
Chocolate Cake (Chocolate Cake)
Chocolate Chocolate Chocolate Cake
Chocolate Chicken Chicken Cake
Chocolate Chocolate Chocolate Chocolate Cake
Chocolate Chips
Chocolate Chips With Chocolate Chips

When I tell it to get creative, things get even weirder.

Beef Soup With Swamp Peef And Cheese
Chocolate Chops & Chocolate Chips
Crimm Grunk Garlic Cleas
Beasy Mist
Export Bean Spoons In Pie-Shell, Top If Spoon and Whip The Mustard
Chocolate Pickle Sauce
Whole Chicken Cookies
Salmon Beef Style Chicken Bottom
Star *
Cover Meats
Out Of Meat
Completely Meat Circle
Completely Meat Chocolate Pie
Cabbage Pot Cookies
Artichoke Gelatin Dogs
Crockpot Cold Water

A list of plausible names for new English towns, as generated by a neural network trained on the real thing:

  • Gateshall
  • Buckfidge
  • St Mersey
  • Surby-on-Te-Shame
  • Wood Edmon Amps
  • Market Trey
  • Faxmouth
  • Harrorstampe and Haunton
  • HighClasey
  • Minehampton
  • Burking Neotrin
  • Darkenhall
  • St Cowes
  • Cawndrood
  • Gridston
  • Woldington
  • Bookeardan Terbury
  • Helpston Cortfoe-Zoure
  • Easingwick
  • Lowestool
  • Ripingamer Trey

Video game titles created by a neural network trained on 146,000 games:

  • Conquestress (1981, Data East) (Arcade)
  • Deep Golf (1985, Siny Computer Entertainment) (MS-DOS)
  • Brain Robot Slam (1984, Gremlin Graphics) (Apple IIe)
  • King of Death 2: The Search of the Dog Space (2010, Capcom;Br�derbund Studios) (Windows)
  • Babble Imperium (1984, Paradox Interactive) (ZX Spectrum)
  • High Episode 2: Ghost Band (1984, Melbourne Team) (Apple IIe)
  • Spork Demo (?, ?) (VIC-20)
  • Alien Pro Baseball (1989, Square Enix) (Arcade)
  • Black Mario (1983, Softsice) (Linux/Unix)
  • Jort: The Shorching (1991, Destomat) (NES)
  • Battle for the Art of the Coast (1997, Jaleco) (GBC)
  • Soccer Dragon (1987, Ange Software) (Amstrad CPC)
  • Mutant Tycoon (2000, Konami) (GBC)
  • Bishoujo no Manager (2003, author) (Linux/Unix)
  • Macross Army (Defenders Ball House 2: League Alien) (1991, Bandai) (NES)
  • The Lost of the Sand Trades 2000 (1990, Sega) (SNES)
  • Pal Defense (1987, author) (Mac)

(part one, part two)

The neural network generated pickup lines that are actually kind of adorable

So it occurred to me that the same neural network framework I’ve trained on recipes, Pokemon, superhero names, and Irish tune names might be able to write pick up lines as well.

Gathering the dataset was much more painful than I had expected - I hadn’t really read many of these before, and most were obscene, or aggressive, or kind of insulting. I began to regret the whole project.

But although the neural network figured out the basic forms “You must be a … because….” or “Hey baby, wanna…” it never learned to generate the worst lines - most of these were based on wordplay that it didn’t have a chance of reproducing. 

Instead, it began to generate lines that varied from incomprehensible to surreal to kind of adorable:

Are you a 4loce? Because you’re so hot!
I want to get my heart with you.
You are so beautiful that you know what I mean.
I have a cenver? Because I just stowe must your worms.
Hey baby, I’m swirked to gave ever to say it for drive. 
If I were to ask you out?
You must be a tringle? Cause you’re the only thing here.
I’m not on your wears, but I want to see your start.
You are so beautiful that you make me feel better to see you.
Hey baby, you’re to be a key? Because I can bear your toot?
I don’t know you.
I have to give you a book, because you’re the only thing in your eyes.
Are you a candle? Because you’re so hot of the looks with you.
I want to see you to my heart.
If I had a rose for every time I thought of you, I have a price tighting.
I have a really falling for you.
Your beauty have a fine to me.
Are you a camera? Because I want to see the most beautiful than you.
I had a come to got your heart.
You’re so beautiful that you say a bat on me and baby.
You look like a thing and I love you.

Neural Network Story Name Illustrations 1-4

Images 1-4 of neural network story name illustrations (names from

Nancy Drew: The Last Day - Can Nancy Drew solve the Case of the Malevolent Moon before time runs out?

[Pictured: Nancy Drew stands in the clock tower from Majora’s Mask holding an ocarina. Outside, the evil moon hangs low in the sky and the Skull Kid watches from the background.]

Market That Knave - Dastardly villains, dashing makeovers!

[Pictured: A printed photo of Snidely Whiplash labeled “Before” lies on top of an unseen photo labeled “After”.]

Murder’s Eagle - The most brutal and bizarre bird in America!

[Pictured: A menacing eagle flies while holding a chainsaw.]

American Midnight: Swear Dragon - Are you a bad enough dragon to save the city from ninja related crime?

[Pictured: In a dark city street, a humanoid dragon takes down a sword wielding ninja using RADICAL SKATEBOARD STUNTS while breathing a cartoon swear as a fireball. More ninja goons approach from all sides. A label at the bottom says “Swear Enix“.]
Disturbingly vague ingredients generated by neural network

This neural network, a learning algorithm trained on 30MB of cookbook recipes, generates new recipes based on probabilities. The resulting ingredients, while their words are individually probable, can end up disturbingly vague. “Yeah… I’m pretty sure this recipe’s gonna contain some… chunks.”

¼ cup white seeds
1 cup mixture
1 teaspoon juice
1  chunks
¼ lb fresh surface
¼ teaspoon brown leaves
½ cup with no noodles
1  round meat in bowl

Titles of Wikipedia entries from an alternate universe, as generated by a neural network:

  • The County Route 37th District Championships (Minnesota)
  • Sonic police of the Georgia
  • Berry War
  • Sinister of the Canada and Stars (language)
  • Telephosphate (disambiguation)
  • Great Story Conservation
  • Alan Communication (Australian politician)
  • USS Simple District
  • South Business (The Fish Mool Soundtrack)
  • Community of Battle of the San First Airlines in Montance Regiment
  • Boogo (disambiguation)
  • William Cardinalists (song)
  • Saint-Doctor County
  • Color of Saint-Berlin (disambiguation)
  • Order of Santa (film)
  • Star Sharker
  • Blue High School (District of Historic District)
  • Robert Recomatory (comics)
  • Anti-Family Dendric River
  • Speed Baronet
  • Stricture of the Range (disambiguation)
  • Bear of Santa (footballer)
  • Single and Engineering Act 1982
  • Canton of Dark of Business
  • Shine County Controller

Another set of video games from the mind of a recurrent neural network:

  • Kimaramu (Rage Planets) (2001, Interactive Martech) (Windows)
  • Circus Vampires: Soul Brain Machine (2008, Vivendi) (Windows)
  • Fun Trim World Wail World Risk! (Tetris Deluxe II) (1981, Hewles DF) (WinC5)
  • Barrow Legend (Tom Clancy’s Ghost Heaven) (1983, Project Software) (C64)
  • New Driver: Forbidden Nightmare (2009, New Syscom) (iOS)
  • Super Cautou Bear Tournament (1997, Niogame) (Mac)
  • Resident Evil 3D: Void Edition (1984, Hewson) (Amstrad CPC)
  • Broken Man Simulation (1986, Elite) (MS-DOS)
  • The Wibble Championship (?, ?) (C16/Plus4)
  • Guns 2 (Love and Kids) (1988, SSI) (Apple IIe)
  • One Universal Simulator (2005, Renegade Software) (Windows)
  • Football EXE (1990, ASCII) (Amiga)
  • Pocket Madden Encounter (The Antica Research Space) (2003, -) (Windows)
  • SpongeBob Street Revolution (2012, Telltale Games) (iOS)
  • Space Kinkette (1987, Infocom) (Amiga)
  • Final Fantasy Drivin’ (1992, MC Works) (Game Boy)
  • Shadow Fishing Black (Falline Sword) (1992, Infogrames) (MS-DOS)
  • Cosmic Beat: End Engineer (2000, Interactive Box Games) (Linux/Unix)
  • Pocket Guide to the Missing Forever (2003, ?) (BeOS)

I’m working on a Tumblr/Twitter bot that automatically generates and posts these, so if I generate any further titles I’ll probably fold them into the bot rather than posting them here.