Neural Network trained on UKC Logbooks - The Results

We recently shared the work of Janelle Shane, who trained a neural network on a database of route names from Joshua Tree (5,633) and Boulder, Colorado (4,527). The results were both amusing and baffling. We wondered how the generated names might differ if we provided Janelle with our much larger database of 432,000 route names, which we split by country.

Funny names generated by a neural network. © Janelle Shane
Funny names generated by a neural network.
© Janelle Shane

A reminder of what a neural network is, for those who are unsure:

'A neural network is a type of computer program that learns by example, rather than being told exactly how to solve a problem. Based on thousands of examples of route names, it had to figure out the rules that let it generate more like them. At a low temperature* setting, it will generate names that it thinks are very quintessential - they'll end up a bit repetitive, but it will mostly be correct. At a higher temperature setting, it will be more daring when it generates names, going with less common sounds and phrases.'

* Temperature is a hyperparameter of LSTMs (and neural networks generally) used to control the randomness of predictions by scaling the logits before applying softmax...apparently...

We jokingly asked if this is the future of route naming for first ascensionists. James McHaffie was firm in his response...

Caff says no. © UKC Articles
Caff says no.

Janelle duly fed the information to a neural network to produce a list. There's definitely a more British feel to the names created using our data. Sheep, tea and even midges make an appearance. Janelle guided us through the process. Check out her blog on this latest project here.


First, for the cleanest dataset possible, I extracted the countries that have mostly English-language route names (about 155k names once I removed duplicates and numbered routes) and now the neural net is churning out one plausible name after another.

The Stuff

Rocket Sheep

Ramp of Lies

Strangershine

Candy Storm

The Dog Sand

Holy Mess

Left Hand Monster

The Scratching One

The Angel's Crack

Suckstone Gully

The Folly Cloud

Burning Doll

Silver Milk

The Cat Bear

Block of Fred

The Limber

Element's Chimney

The Space Special

Bear Box

Smashworm

The Peacher

The Sun Mouse

The Bobble Block

The Rib (Stinkley)

Cry Problem 15

Scary Boulder Start

Solo Gallow Wall (STEXXY

The Sole and Elephant

Crag and Be Bloody

Midge Face

Seven Belly

Wine for the Great Free Man

I'm happy to report that some of the names were indeed even weirder than your typical route name.

You're Not Andrew

Master In Your Tea

Bean on the Pocket

Seven Dry Have Ship

No Rocks Egg

In Arms if the Lords

Parking Store Substance

Over a Wall No Mover

The Very Seven Steps

Robin Time and The Sheep

Captain Purple and Darkness

The Sun Tin's Not Your Winds

There are, of course, a whole category of names that are quite rude...

Joy's Crack

Mantlet Butt's Locket

Ring for A Spank

Finger It

The Slappy Slab

The Slippery Man

Shocker's Crack

Cold Beever

Silver Wally

Handride

More Sex

Sink of Bugger

Welly Will

Enter Crack

Anal The Light Black Straw

The Willy Wind

Middle Cock

The Willies of the Smith

Neural Networks © Janelle Shane
Neural Networks
© Janelle Shane

Next, I trained the neural net on the entire database, just to see what it would make of the non-English names. It definitely struggled more this time - it reported much lower confidence in its results. But it did manage to become multilingual, generating names that were identifiably French, Spanish, or German (these were the most common languages other than English in the dataset, so these were mostly what it learned). Even if many of them didn't make much sense.

La Grimper - French: The 'to climb'

Cascade de l'ange - French: Angel Falls

De l'angle de la surplomb - French: From the angle of the overhang

Steines Schwein - German: Stone's pig

Sin Homble - Spanish: Without Homble

El Pollute - Spanish: The Pollute

Rapute de la vine - Romanian: Rapping of the coming

Danse ton de Barre - French: Dance tone of bar

Sometimes it did end up mixing up the languages, although not as often as I had expected. Maybe it was doing it much more with languages I was less familiar with, and I couldn't tell.

Via Les Rocks (Left - Low Shit Thing)

El Pantes du Petit

La Desire del pierra

Le Chins de Constant (Standing Pub)

El Lope du Pante

Sans Inside Droit

Via de la finger

Les l'Appolena

Placa de Carpet

The Schlang

Its brainpower was spread a bit thin, trying to remember rules for generating multiple languages at once. Unlike human brains, it definitely wasn't built for compartmentalizing multiple languages. This struggle had an effect on the quality of its English names, which actually I rather like.

Boulder 1, Problem the Gorge

Very Up

Fred birthday

Red 1

Blue Boulders Problem 1

No we and Cheese

The Spooning

The Corner Stand of The Little Heart

Cat of the Shallow

Serpent Mars

End Cow

Escapes of the Beach Brother


This post has been read 7,029 times

Return to Articles from 2018 or list other articles


28 Jun, 2018

Missed a trick with the Cat Bear graphic:

https://en.wikipedia.org/wiki/Binturong

28 Jun, 2018

Right, need to find a new Crag X and get to work now. Love those names. 

28 Jun, 2018

Agreed, some are actually quite good.

29 Jun, 2018

Finally, machine learning finds a proper useful application :-)

29 Jun, 2018

They don't look much like route names to me.

But I'll definitely put a fiver on The Cat Bear to win the 14.15 at Uttoxeter.

More Comments