This seems to be a year for AI experiments centered around Christmas movies. MIT Technology Review generated some titles and plot summaries (I would probably watch The Christmas StorK) and Botnik Studios has been using predictive text to write weird new Christmas movie plots and summaries. So when Nicole Kobie from Wired UK sent me a list of 300 Christmas movies, I had to get in on the fun.
A dataset of just 300 titles is pretty small, and I knew a neural net might have trouble with it. There wouldn’t be enough examples of how to use various words, and so it would end up parroting them back without being able to recombine them in new and interesting ways.
My best best, I decided, would be to start with a neural network that had already learned to generate movie titles on a different, but much bigger, dataset. Then I would train it for a little bit on the Christmas movies, so it could use what it already knew about spelling words and using them in movie titles, yet learn to specialize on Christmas movies. This strategy is called transfer learning, and is used all the time for commercial algorithms. Want to train an algorithm to recognize cancer cells? It can be cheaper and easier to start with an algorithm that has already learned to recognize cats, cars, and bicycles.
I trained a textgenrnn neural net on a large dataset of movie titles and then on the Christmas movies, and sure enough, it managed to generate new and interesting movies. It even remembered some of the words it had learned earlier, and knew how to use them.
Christmas Creature Case
Spring Can’s Christmas
A Christmas Cat
A Christmas louse
The Noise of Christmas
The Happy Car
Can Christmas Of PuppY
When Plant Deaal
The Christmas Manury III
However, perhaps things would have gone slightly better if the previous dataset hadn’t been horror movies.
Revenge of Santa
A Christmas Comes
Fist of Christmas
The Christmas Dead
A Cry of Christmas
The Haunted Christmas Love
Or maybe it didn’t do so badly after all. It generated this eyebrow-raising horror/Christmas mashup:
which, as of this week, actually exists.
I tried again with NOT horror movies as my starting dataset - instead, I used the titles of basically everything on Wikipedia that has a plot summary (a dataset collected by Mark Riedl). It sure did inject more variety into the movie titles, but I’m not sure they actually got better.
Open A Fist Gift
Best Snape for Christmas
RantBerry Joy Boots
To read the rest (and get other bonus content), become an AI Weirdness supporter. Or become a free subscriber to get new AI Weirdness posts in your inbox!