I recently tried one of my favorite places for neural networks, a small program that runs on my laptop and knows about the data I feed it. Without Internet training, Char-RNN has no external indicators to draw (for better or worse) but it can still achieve money in new interesting things.
In 2017 I asked Ai Fireirdnessnessnessnessness students to help me CropdsSource List of Halloween costumes and he trained Char-RNN to about 4500 level which I got. Today I'm coming back to both the database and Char-RNN (here's a version running in modern Python), mainly because it's still fun for me. My laptop is more powerful now than the 2010 MacPook I was using at the time, so I can run a large version of Char-RNN. I really can't say if it helps. But I know I'm happy:
Skypug
Hungry boats
Mid Wonka
Srrenous bread cat
Holy Cheesarenda
The Mouth Fairy
The magic slice
Fall dressed monster
The goddess
I looked, and no one going to the training data from 2017 was using “mid” as an adjective, so “Mid Wonka” is a coincidence. The larger version of the Char-RNN was better than I expected at sculpting clothes, producing exciting new characters.
The Science Horse
Lady doo
Captain Gay
Silence the minister
Cheetos Captain
The scorph doo
Undead Mario
Sailor Who
There were quite a few Scooby DOO hats in the original training data, which is probably why the neural net is wearing its hats.
Did you know that it was not in its 2017 training information for 2017? Kpop demon hunters, I have it on good authority will not be an unknown costume in 2025. For fun I asked the neural net to complete the phrases “kpop” and “kpop de”:
Kpop punk
Kpop and the man and the bus
Kpop Bader Ginsburg
Kpop dog
Kpop Donald Science
Kpop devil monster
Kpop Dead Dead
Kpop classification
Kpop dead turtles
This holds for major language models like chatgpt too, of course. If a current event or idea is lost in the training data, it is lost in the outcome.
The full Halloween costume submission form is still open, so if you have a few costumes you've seen or dreamed about lately, you can help bring the training data up to date! If I get enough new representations maybe I will train the neural network again. (The dataset as of Oct 4 2018 is available on my github).
Bonus content for Ai Fireirdness supporters: Many of my learned outfits trained from the 2018 dataset!
