When neural networks name planets, they call them Tina

The darkest known exoplanet, TrES-2b reflects just one percent of any light that hits it. Likely of similar bulk and composition to Jupiter, the planet is less reflective than coal. It isn't yet known why it is so dark, but it has been speculated that TrES-2b's atmosphere is filled with light-absorbing sodium, potassium, or titanium oxide.

If a sci-fi film ever names a planet Tina, blame Janelle Shane.

Her hobby is training neural networks on data sets to create amusing names, be it rescue kittens ("Mag Jeggles" and "Snox Boops"), Pokémon ("Tortabool"), and — perhaps most famously — paint colours ("Sudden Pine" and "Turdly").

In the latest round, the neural network was trained on 700 planets from Star Wars. That's worth doing, she says, because "Kepler-452b" isn't catchy enough to shout out to "the ship's engineer during a raging ion storm", and there's thousands of exoplanets that need better names. After feeding in the data supplied by a fan, her neural network spat out better suggestions, such as "Bartan", "Vantos" and "Nananon".

"I was surprised it worked so well, with such a small data set," Shane says. "I wasn't sure it would come up with anything new or just be a mix of characters. But it came up with Star Wars-sounding names without repeating any of the ones that were given."

Indeed, the first batch were a little too good, so Shane flipped the neural network's dial to a more random setting for a bit more creativity, leading to "Duperda," "Rrarar" and "Tina". The system spat out a collection of dull English words, including "Loon", "Robes" and "Sara".

"It's trying out different combinations of pronounceable letters that sound kind of spacey or have something in common with spacey sounding names," she says. "It has no knowledge that its readers are in English or what any English words are, so it will come up with pronounceable combinations that just so happen to overlap with existing words, so you get things like 'Tina' — yeah, that sounds really futuristic."

Shane uses two different open-source neural network frameworks to name kittens, paint and planets, but she's no AI expert — indeed, her background is in electrical engineering and physics. Her day job is steering laser beams.

While AI, neural networks and machine learning are all big business in the tech industry, Shane is in it purely for her own amusement. It all started after she couldn't stop laughing at recipes created by neural networks, reading every one she could get her hands on and dissolving into hysterics. "They were so funny that I wanted to make more of these," she says. "That's what got me interested in learning to train in neural networks. I'd never done anything like that before."

"I've had a lot of people say to me that because of this blog, they're experimenting with neural networks themselves. It's surprisingly accessible. "You don't need a supercomputer. You don't need a sophisticated data set... If you have a list of 700 names of Star Wars planets you can start experimenting with neural networks."

Plus it shows how computers can contribute to creative pursuits. "If you're writing science fiction and you want to generate some new character names or planets, a neural network can help spark that creative process."

Because of that, any sci-fi writer struggling with writer's' block should follow Shane's work — after helping christen planets, she plans to run a data set of Star Wars character names next.

Let's block ads! (Why?)