Neural networks can raise a smile, but do they have a sense of humour? Nicole Kobie reveals the world of AI-created paint, Pokémon and recipes
Let computers think for themselves and they’re funnier than people ................
tummy Beige. Dorkwood. Sindis Poop. Turdly. These are not paints you’d choose to slather on the walls of your front room, but it’s what Janelle Shane’s neural network spat out after being trained on 7,700 Sherwin-Williams colours.
Artificial intelligence (AI) is serious business. It’s threatening to take our jobs and change our lives beyond all recognition, from self-driving cars to policing by robot and beyond, potentially one day overtaking our own brains and leaving our biological intelligence in the dust. However, if you read Shane’s Tumblr blog showing the results of her research, you won’t be quivering with fear but laughter.
MAKING “SUDDEN PINE”
Shane isn’t an AI researcher by trade – she works in optics – but plays with neural networks because they crack up her and her blog’s many readers (lewisandquark. tumblr.com). Using an open-source framework called char-rnn and developed by Andrej Karpathy two years ago, she feeds in a dataset to train the neural network, eventually letting it make up its own paint names, Pokémon characters and more. After studying the data file, it constructs words by guessing what’s likely to be the next character, aiming to make words that match the original training text.
“It’s looking at a sequence of a certain length… trying to predict what the next character should be,” she told PC&TA. “When it’s done that, it moves to the next one.” Not only has her neural network created paint colours such as “Sudden Pine” and “Greenwater Chamiweed”, but 1980s action figures (“Battle Command Master Cramp”), Dungeons & Dragons spells (“Gland Growth”), and even Doctor Who episode titles (“The Dalek of the Daleks”).
While a more serious researcher would be looking to get an accurate result from a neural network, Shane is out for laughs. “I’m not experimenting in a systematic way, as you would if you were optimising or solving a problem, [aiming to get] the closest match [to the original text],” she explained. “In my case, having the neural network work not entirely well
can be good for comedic e ect. I will sometimes stop the evolution early if I’m getting more interesting results.”
It all started when she stumbled on a collection of neural network-written recipes. “I thought they were hilarious and read through them all, and then they were done and the only way to get more was to do them myself,” Shane said, adding that she had never worked with neural networks previously.
Bending neural networks to goofy humour is certainly one way to learn. “Being able to explore the neural network for a simple dataset gives me a real appreciation for the more complicated problems and more sophisticated neural networks that modern research is using,” she explained. “I am definitely learning things about neural network size and about dropout and all these variables that are traditional in neural networks, but I’m learning them through experimentation rather than systemic study.”
While Shane was seeking a giggle, none of this means comedians are at risk of losing their jobs to AI. Julia Taylor Rayz, assistant professor in the computer and IT department at Purdue University, said it’s possible to teach computers to create jokes, but the results are limited.
There are two ways to train AI to make jokes. “One is we explain to a computer the rules that jokes are based on,” Rayz said. “There are quite a few theories of humour… about what makes a joke a joke.” On top of those rules of humour, you’ll also have to give the system a “knowledge of the world” so it has some material to work with.
The second technique is to tell the AI a lot of jokes, letting it learn by example. “You’re feeding a computer as many jokes as you can, and you hope it will find patterns or logic or features that will let it di erentiate jokes from non-jokes,” she said. “If you’re doing that, the result is going to be very much dependent on how well you select your training corpus, how well you select the jokes [and non-jokes] you are feeding it.” Keeping it niche will o er the best results, she added.
Of course, Shane’s neural network isn’t trying to be funny – she’s letting it loose and collecting the results of its unintentional comedy. “Computers definitely make mistakes that are hilarious,” Rayz said. “For a computer to purposefully become a stand-up comedian – writing its own script and delivering it such that people will find it amusing – that is a little bit further away.”
Sometimes intentional and unintentional comedy can combine. Shane tasked the neural network with creating knock-knock jokes, a niche format that according to Rayz should be relatively successful. Its first attempts were in the right structure, but the words were nonsensical – it read one example with a punchline of a cow mooing, and took a while to get variants of mooing as an answer out of its system.
After many tries, Shane was shocked that the network managed a completely original joke. Here it is: “Knock Knock.” “Who’s There?” “Alec.” “Alec who?” “Alec- Knock Knock jokes.”
We didn’t say it was that funny. Indeed, Shane said she was surprised how many people find her neural network’s output amusing. “I thought I had a weird or di erent sense of humour, but it turns out there are a lot of people who find this equally as funny,” she said. “It may be that it’s tapping into something that’s pretty common to a lot of people, the same way that kids’ sayings or drawings can be funny.”
Few humans would come up with paint colours of “Queen Slime” or “Porchtingle Grey”. Does that mean Shane’s neural network is showing real creativity?
“I think the neural network is a tool… in the same way that using splatterpaint technique would give you an unpredictable pattern,” Shane said, adding she’s in truth the main creative force, as she chooses the dataset with the most potential for hilarity and manipulates it to the right degree for comedy.
But humour is just one potential artistic output of Shane’s neural network fiddling. “I have been contacted now by people who are artists, painting and so forth, and they want to use the neural network as a tool for their artwork... taking it from paint to text to paint again,” she said.
Here’s hoping a clever artist combines their own creativity with Shane’s goofy paint creations – we can’t wait to see what the first Jackson Pollock of the AI world manages to produce with colours such as “Flumfy Gray” and the blue “Pester Pink”.
Janelle Shane was inspired to begin the project after reading hilarious recipes generated by a neural network