Illustration: Janelle Shane / AI Weirdness
Electrical engineer Janelle Shane's hobby is giving neural networks — a computing system that loosely mimics a brain — somewhat silly datasets to see what they can create.
Most recently, she input 44,126 thesis titles from MIT. Some of the products — “Atoms and characteristics of monolithic nanocity,” for example — may not immediately jump out as odd. Others, Shane says, are "completely ridiculous."
Case in point: Spacecraft Coal battery induced by mortgage microcontrol.
How it works:
- The neural network basically predicts what the next letter in a word or phrase should be based on the probability of it being in that spot across the dataset it was trained on.
- If there is an 80% chance of an ‘h’ following a ‘c’ and a 20% chance of it instead being an ‘e,’ the responsible neural network would go with ‘h.’
- But if the creativity setting is turned up, the AI may gamble with the less probable territory and see where it takes it, says Shane.
- At a certain point, though, it may output words that don't even exist. (For example, "frook" and "nurler.") “It is an art form with its own connections," says Shane.
Yes but: There is the issue of bias, which is clear in the dataset.
- The word "control" (a favorite of engineers) and the phrase "of a medical device" (a sign of the times) appeared often.
- And because these were MIT theses, tokamak — an experimental fusion device that was a source of institutional pride before it was shut down — turned up often.
- The neural network returns results skewed by the raw materials it was given to start.
- It wouldn't do a good job of detecting and collecting all of the world's theses into a database given only a sample from a school focused on engineering and science.
- And, Shane's goal is for this to be funny, so her own sense of humor colors the results.
Chuckle all you will about the neural network's creations — and they are entertainment in Shane’s eyes — but if what's generated resonates with humans and is then replicated or perpetuated by us, the neural network is arguably creative at some level. An opera company used lines generated by Shane’s neural network and a brewery in Michigan took one of the machine’s suggestions of "The Fine Stranger" in naming one of its beers.
The larger lesson: The AI actually understand very little, says Shane. The neural network she uses is roughly equivalent to the brain of an earthworm in terms of neurons employed. “It is limited and unintelligent when taken outside it's comfort zone," she says. "And, my computer is dedicating 100% of itself to these tasks, but an earthworm has to do other things."