Sports teams designed by neural network

One of the many impressive things about neural networks is how well the same basic algorithm can adapt to very different kinds of problems. I’ve used the same framework, Char-rnn, to produce things like recipes, Dr. Who episode titles, D&D spells, story titles, metal band names, Pokemon, and paint colors. With no prior knowledge of what the input dataset is, the neural net looks at it a few characters at a time and learns to predict what comes next. Once a neural net can predict the next character in a set, it’s easy to make it generate text by giving it a single character and asking it “what next?” and then “what next?” again and again.

Here, the neural network has to learn a dataset that would probably be very difficult to explain to any sort of artificial intelligence: the names of sports teams.

“You know how we humans like to run around kicking or throwing balls into nets? Well, we do. Anyways, we need names for the teams so we know what to shout to help them during games. No, we also shout when the games are on television. It’s hard to explain.”

The training data was a list of almost 46,000 high school, college, and professional teams, generously provided by MascotDB.com.

Checkpoint: 8%

By the time it had gone through 8% of the dataset, the neural network had learned about spaces, capital letters, plurals, and line breaks, though not about any actual team names. It thinks e’s are probably common. These would be easy to shout, though.

Moltee Raiees
Raite Ballees
Narte Calaees
Bareen Coilers
Raise Callen Caaeees
Rareen Cartees
Barien Baleees
Corte Colteee Cares

At a higher creativity setting, the names are somewhat… less shoutable.

Mal Ceinrl Melga Carynoads
uamctes Biiihtes
Jaatohi jrpmWhollrioa ongmgis
Beettt Riiara fiicly neadks
Soiiln Uakeas
Cirecs C vieee Sa lBlrrl
Nutl Jatreeh Rthglrrs
Saihei UilanlTr Pgllnms
BecoNviroe Ntnrraintoee RittaotboiRmiirlovatts

Checkpoint: 23%

Now at 23% through the dataset, the neural network has eased back on the e’s. The combination “ers” is now its new favorite. It has not yet managed to produce any actual city or team names yet.

Vande Corte Carters
Galed Cares
Vande Carders
Geen Calders
St. Corthors Callers
Calle Ballers
Dander Banders

Checkpoint: 46%

The neural network is now trying very hard to spell “tigers”, “panthers” and “cardinals” but has not yet figured out exactly how this works. The words “State” and “College” have featured prominently enough, however, that it has successfully learned them. The neural network has also somehow figured out how to apply suffixes to them, although not how to apply the correct ones.

Wanterloo Sarrians
Oulondo Callens
Stateston College Tagers
Staters College Tagers
Dutenton Tagers
Dunger Parthers
Wanernion Cardinans
Statet College Balldons
Sacreter Wardinals

Checkpoint: 107%

In a surprise leap, the Bulldogs have taken the lead. The neural network is still struggling with Cardinals and Tigers. And wherever the heck Furton is, it has been deemed Very Important in sports.

“No, bulldogs and tigers aren’t actually good at sports. They’re just fearsome.
… no, cardinals aren’t particularly fearsome. It’s hard to explain.”

Furton Bulldogs
North College Colders
Marina Cartinals
Furton Rads
Marine Bulldogs
Corton Tagers
Furten Cougars
Burning College Bulldogs

Checkpoint: 198%

Now the neural network can spell things. It will hallucinate nonexistent counties and colleges all day, and assign them tigers and eagles and cardinals as mascots. It has basically solved mascot names - if you want an ordinary mascot name, the neural network can pick one for you. With a small chance that you’ll end up being the Radgers or the Tancans.

Lunter County Tigers
Wort County Bulldogs
Lord County Stars
Bullson Tigers
Torcond Cougars
Portsan Cardinals
Troston Radgers
Turson County Eagles
New York Hankers
Bringson Tancans

At the high-creativity setting, however, the team names get a bit more interesting. You would not want to be on any of these teams.

New mork Punts
Borton Files
Wurt Academy Eiglrs
Pirtsbuugh pantons
Frirg City Corgars
Longer Golden Tuckems
New York Hoomin Lions
Hontbern Academy Land Pontpins
Fart Mirea Chuds
Buncawnia Ragbocks
Tinstol-Hulbbynhibnbla Tigers
Comncamola Slimos
Turreritite Academy Beans

Checkpoint: 3280%

Let the neural network continue learning, and the low-creativity results look much the same, while the high-creativity mascots get more pronounceable but not a bit less weird. Same with the place names.

I would actually like to see the mascot costumes for some of these.

Bleondef College Pheelans
hlel Fratthawks
Artfy Ehole Lancers
Riomport Arabemme-Warly Driton Bopcets
Hickmine Tigers
Liverville Academy S.inars
Bloodman Park Mustangs
Dooming Bearcats
Blooliky Batbles
Lickin Wildcats
Joviet Ponelards
Lameland Cuishalk Gerbuts
Warton Arue Mess
Ever Panthers
Bippinsmore Red Yellowjackets