A Neural Network Gifts Us With New Dungeons & Dragons Creatures Like Vampire Bear, Walfablang, and Meme Ball

Published On: Sun, Mar 25th, 2018

Entertainment | By admin

We’ve covered Janelle Shane’s absolutely delightful work with neural networks before, and now she’s back at it again. This time, she’s trained a recurrent neutral network to name Dungeons & Dragons (D&D) creatures. The results are just as entertaining as when the network tried to create Dungeons & Dragons spells back in September of 2017.

With the help of Colin Fredricks from Sufficiently Advanced, who sent her the names of 2,205 creatures from the second-edition D&D monster manual, Shane was able to feed the network a large enough data set and then watch it do its thing. As Shane has explained previously, a neural network’s results are only as good as its data set. The more information it has, the better it can “guess” what sort of results it should spit out.

“As I had hoped, the neural network generated creatures that would probably be pretty awesome,” she wrote. “It also generated some creatures that you should probably run from until you figure out what they are. (Though Dome Animal might simply be a cool turtle).”

The network also has its favorites. “There were 118 dragons in the original dataset,” Shane explained, “so of course the neural net liked generating new dragons. Some perhaps better-conceived than others.”

Below are some of my favorites from this exercise, but you can see the full list of results over at Shane’s blog. You can also help with a future Dungeons & Dragons data set by entering information about your own D&D characters on her survey here.