Creating Brand Names with AI
Brains have their limitations.
That’s why I use technology to unlimit my own creativity. Software is indispensible to my naming process, enabling me to find the right words, create new ones, and stimulate idea generation.
Since college I have been AI-curious. I studied natural language processing, learned LISP, and briefly majored in Artificial Intelligence at Clark University before transferring to Linguistics at UC Santa Cruz. Since then, breakthroughs in processing power and cloud computing have brought massively powerful AI within reach of anyone with an internet connection. AI is already everywhere: In our phones, guiding our drives, turning us on to new music. And so much more.
I first learned about Janelle Shane’s AI naming experiments in early 2017 when she trained a recurrent neural network to create car names. They weren’t very good names, but that didn’t matter. AI was being used to create names and I needed to master it.
I asked Janelle if I could hire her to teach me what she knew. She was receptive. Janelle provided me with custom step-by-step documentation and spent hours with me on the phone, patiently bringing me up to speed on AWS, Deep Learning, Torch, and Lua.
Way back in September 2017, working with a recurrent neural network was done exclusively through a Linux command-line interface. All ASCII, no GUI. Linux doesn’t hold your hand and help you get things done. When I got stuck, troubleshooting sometimes took hours. But it was immensely satisfying to struggle through and eventually figure out what went wrong. I grew comfortable (and competent) with my narrow range of tasks on Linux.
It‘s now March 2018, and web-based interfaces for neural network creative generation are popping up. They are limited in functionality and the user’s ability to fine-tune, but they demonstrate the basics, sometimes in really interesting ways.
For now, web-based neural network apps can’t hold a candle to what you can do through a command-line interface.
Here’s how I use AI to create names:
A recurrent neural network is “trained” on data you provide, meaning that it attempts to learn how to generate original ideas that emulate the training data. In my case, that training data is a text file containing hundreds or thousands of words. Sometimes it’s a list of name candidates concatenated from prior projects. It may be an inventory of topically-related words culled from an online corpus. Or I’ll train the neural net on all of the current brand names in a specific category, like luxury goods or hedge funds.
Once trained, you “sample” what the neural net came up with. There’s wide variability in the quality of the output due to all the variables that control it. Training can be adjusted in six different ways. Sampling has four variables to tweak. A small dataset won’t give the network enough information to create its own, original ideas. There's a lot of trial and error. But when everything is fine-tuned just right...most of the output isn’t all that great.
Actually, it’s rather like a human-made naming list. Most of the names in typical namer’s list are crap (mine included!), but some names are, hopefully, really good. With a sufficiently long list there are enough good names identified to survive trademark screening and present to the client.
So it doesn’t really matter that most of the AI’s names are lame, because some of them are good. And that’s good enough.
For example, here are a few good neologisms from a neural network trained on a massive list of Latinate prefixes and suffixes:
In another experiment, a list of compound words inspired my neural net to come up with these:
The neural net generated these based on list of science and astronomy words:
Moon of Action
What impresses me with these examples is their linguistic naturalness. They are new words but feel like exising ones. Like they should already be in a dictionary but through some accident of history, aren’t.
I’ve incorporated AI name development in 6 or 7 projects and presented a handful of those names to clients. No winners yet.
AI isn’t going to put professional namers out of a job. For one, there's a steep learning curve. Neural networks require a lot of fidgeting to generate decent outcome. Their accessibility is limited because a command-line interface is their only interface (though that’s changing weekly). But even when neural networks become easy to access and use for name creation, they will never be able to identify the best names in a list, develop strategic rationale for them, and argue an impassioned case for their value. As my naming peer, Olivier Auroy, said, artificial intlligence can’t “know when a word is brand material.”
AI will prove useful among the other tools used by professional name developers to create new brand names. It‘s helpful to me today as an adjunct to my brain and other software. And, in an historic first as far as I know, Janelle has named a comercially-available beer — The Fine Stranger — using a neural net. Hat‘s off to her for bringing a great, AI-developed name into the world.
A new era of very interesting names is coming.
I, for one, welcome my new creative partner.