If you raise an eyebrow at the many questionable Halloween costumes currently on the market, just be thankful your outfit wasn’t designed by an AI.
Colorado-based research scientist Janelle Shane has trained a neural network to invent costumes based a dataset set of more than 4,500 crowdsourced submissions, ranging from the classics to the countless ‘sexy’ interpretations available today.
In the experiment, Shane used a type of neural network that learns words without understanding their meaning – resulting in a hilarious assortment of made-up characters, from Aldonald the Goddess of the Chicken to Gandalf the Good Witch.
Colorado-based research scientist Janelle Shane has trained a neural network to invent costumes based a dataset set of more than 4,500 crowdsourced submissions, resulting in a hilarious assortment of made-up characters. Stock image
HOW SHE DID IT
While some neural networks are able to learn the meaning of words, the AI used in Shane’s recent experiment did not have such understanding.
Instead, it ‘learns words from scratch, letter by letter, with no knowledge of their meaning.’
The researcher fed the neural network 4,500 costume names, and then left the AI to its own devices.
Early on, the costumes churned out by the neural network were hilariously nonsensical.
And the early trials reflect the frequency of certain words appearing in the training set, the researcher explained– notably, the word sexy.
In a new blog post, the researcher explains that the Halloween classics were the most popular costumes in the training submissions.
This included witches, ghosts, pirates, Batmans, cats (including sexy cats), vampires, pumpkins, and sexy nurses.
And, there were roughly 300 costumes that contained ‘sexy’ in the name, including a slew of bizarre concepts, such as ‘sexy poop emoji,’ ‘sexy eyeball,’ and ‘sexy Ben Franklin.’
The set also contained those considered to be a ‘technical challenge,’ such as Invisible Pink Unicorn.
Given the creativity of humans, Shane explains in the blog post, this task was going to be tricky for a neural network.’
Instead of taking the ‘sensible approach,’ and using a neural network that knows the meaning of the words its using, the creator opted for ‘something a bit weirder.’
‘So, I used a neural network that learns words from scratch, letter by letter, with no knowledge of their meaning, an open-source char-rnn neural network written in Torch.
‘I simply dumped the 4500 Halloween costumes on it, and told the neural network to figure it out.’
Instead of taking the ‘sensible approach,’ and using a neural network that knows the meaning of the words its using, the creator opted for ‘something a bit weirder.’ Some of the AI-made costumes are listed in Shane’s tweet, above
COSTUMES INVENTED BY AN AI
Aldonald the Goddess of the Chicken
Celery Blue Frankenstein
Dragon of Liberty
A shark princess
Statue of Witch
Giant Two butter
The Twin Spider Mermaid
The Game of Nightmare Lightbare
The Rocky Monster
Statue of pizza
The Spiding hood
A card Convention
The Little Pond
Spice of pokeman
Bill of Liberty
Count Drunk Doll of Princess
Statue of the Spice of the underworker
Sexy the Pumpkin
Sexy the Pirate
Sexy Pumpkin Pirate
Sexy Gumb Man
Sexy The Gate
Sexy Ducty monster
Sexy the Bumble
Sexy the Super bass
Pretty zombie Space Suit
Sexy the Spock
A masked box
The shark knight
Gandalf the Good Witch
Vampire big bird
Early on, the costumes churned out by the neural network were hilariously nonsensical, including ‘Sexy sexy Dombie Sexy Cat,’ ‘Sexy Cot,’ and simply, ‘Hot.’
The early trials especially reflect the frequency of certain words appearing in the training set, the researcher explained – notably, the word sexy.
Later trials produced a series of ‘astoundingly good’ mashups, with recognizable characters or objects combined to be one bizarre entity.
This includes ‘Vampire big bird,’ ‘Panda Clam,’ ‘The shark knight,’ and ‘Punk Tree.’
Later trials produced a series of ‘astoundingly good’ mashups, with recognizable characters or objects combined to be one bizarre entity. This includes ‘Gandalf the Good Witch’ (Gandalf is pictured left), and ‘Sexy the Spock’ (Spock pictured right)
The AI also generated a number of other random costumes, such as ‘The Twin Spider Mermaid,’ ‘Count Drunk Doll of Princess,’ and ‘The Spiding hood.’
And, of course, it churned out a batch of its own sexy costumes as well – though, as it’s immediately clear, it ‘never quite got the hang of it.’
Among the many ‘sexy’ costumes, the AI invented ‘Sexy Scare,’ ‘Sexy Lamp,’ ‘Sexy humblebee,’ and ‘Sexy the Spock.’
Shane has become known for her hilarious neural network experiments, training AI to come up with all sorts of ideas.
In the recent past, she’s trained AIs to invent trendy beer names, ‘serious-sounding diseases,’ and even Harry Potter spells.