Letting neural networks be weird

When A.I. gets into birding

[Bird illustrations by the talented Carin Powell (10000badframes.tumblr.com  http://carinpowell.wixsite.com/sheanimates)]

I train computer programs called neural networks to imitate datasets of human things. Basically, I give the neural network a long list of things like band names or guinea pig names to look at, and it does its best to figure out the rules that will let it generate more.

The neural network starts with a fresh slate every time, and becomes the World’s Biggest Fan of whatever it’s given - if I give it Pokemon names, it will invent Pokemon after Pokemon; if I give it Star Wars names, it will invent new characters like Darth Tina and Ban Sand.

There was one time when I gave the neural network a list of 37,000 common names of fish, and it invented fish that were, well, honestly no weirder than existing fish names. Fish names are That Weird.

Since that day, and even before, I have been hearing from birders. Any of you who know birders or are birders yourselves will not be surprised. Not only are birders very eager to find out what a neural network would make of bird species, they are also very organized. There is, for example, a downloadable checklist of about 32,000 birds, 14,000 of which have English common names. Big thanks to Kaija Gahm, Dana Terry, and Emily Davis, who sent me this and similar datasets.

The neural network, after reading intently through the entire list about 7 times, is now a dedicated birder.

I asked it to generate some birds. Not too wild - plausible.

Ecuadorian Helmeted Parrot
Slaty White-throated Fairy-bellied Ground-Tyrant
Tree Sunangel
Lazuli Cuckooshrike
Brown-headed Spadebill
Cape Babbler
Three-toed Wren-Babbler

Ground-Tyrant, I asked? Cuckooshrike? It turns out, yes. There are 57 cuckooshrikes, including a Cerulean Cuckooshrike and a Blue Cuckooshrike, but not an actual Lazuli Cuckooshrike. Similarly, that ground-tyrant would pass as plausible among birders.

Fine, I said. What does actual weird look like? I upped the neural network’s creativity level to 1.0, the highest level I usually use.

Nukh’s Dull Gull
Banded Spectacled Snake-Eagle
Thick-knee
Bunticus Gray-chinned Laughingthrush
Ecuadorian Strange-tailed Cuckoo
Violet-throated Tyrant-Eagle
Horned Yellow Flycatcher
Rusty-browed Highlark
Red-capped Lynert’s Leafbird
Dead Flycatcher
Potland Bustard
Beautiful Pulpertory
Fairy Warbler
Hottled Duck
Rufous-fronted Grassy Owl
Chestnut-bellied Ged Parrot-Weaver
Brown-breasted Leaftosser
Green-hooded Hawk
Burrowing Guineafowl
Unicolored Painted Blue-cowled White-browed Cave-Magpie

Ha, I said. Banded Spectacled Snake-Eagle? That’s hilarious.

Yes, the neural network would have replied (if it was equipped for conversation as well as for birding enthusiasm). Banded AND Spectacled? Ha!

It turns out there are 6 snake-eagles, 41 spectacled birds, and 106 banded birds, but nothing that’s banded AND spectacled.

I turned the creativity up to 1.2, a point at which for other datasets, the neural network is emitting unpronounceable strings of letters with only a vague resemblance to the original. Here’s what the birding neural net produced:

Violet-footed Macaw
Blue-winged Cruz White Kiwi
Buttlebird
Red-faced Imperial-Pigeon
Marabor Island Sooty-Whistler
Cornell’s Palm-Goose
Olive Stonechat
Rus Efricans Puffbird
Ringed Wattle-eye
Northern Narrow Lark
Indigo Mungletoe-Tyrant
Wilson’s Blue-eared Hummingbird
Eyebill Kingfisher
Crinete Bor-billed Mountain-eater
Pygmy Sea Shag
Laughing Fig-Warbler
Perplexaquail-Dove

Granted, even I am registering some of these as weird. Mountain-eater? The neural network made that up. But Wattle-eye, puffbird, stonechat, and shag are all real things. Apparently an *Olive* stonechat is just that weird.

At creativity 1.4 the results get stranger but not as strange as you’d think.

Humboldn’s Cuckoo-Eagle-Parrot
Adbau Mungle-Flycatcher
Forest Tern
Red Frogmouth
Freckled Love
Temetai African Pope Catbird
Dufous Partridge
Fawn-bellied Flowerpecker
Lavaranean Hawk
Hothell’s Hummingbird
Eastern Shove Crake
Long-tailed Honey-buzzard
Iceland Reedhaunter
Blood Flycatcher
Mungleh’s Wattle-eye
Slender-eared Chat

For reference, this is what a neural network trained on pie produces at creativity 1.4:

Dibble Ice Fraini Pien Daria Futgo Crustdamamatsna-LiGmeat Pieb
Pe sivle Hed Rice Frozen Mincemeat Mop shb
Impossible Titer: Fiag
Caramel Apple Wime Figl’s Topped Sugrum’s Pumpar

The birders are probably nodding in agreement with the neural network, though. “FOREST Tern? That’s unutterably silly.”

Finally, once I have increased the creativity to 1.6, I declare the neural network’s names to be Obviously Strange.

Maddeel’s Woodhaunter
African White-capped Hocky Owl
Fairy-collared Barblering Bush-Roe Hyra
Dusky Sicky-faced Petrel
White-throated Sukbird
Koopa
paucosian Rivetpecker
Hoaly Titco’s Badwinch
Snowy Mourning Heron-Robin
Javan Clamper Leafbird
Dog-winged Buczardle
Red-bellied Pale-tufted Junkletar
Moustached White Owl
Pacific Three-hong-toed Thick-dee
Buttmanxwecir
Mar Punybill
Ledt’s Cockadoo
Mountain-rumpting-Finch
warn-winged Wood-banded-Black-breasted Stesing-Patein-fronted Crimsonwing

I generated even more birds at even higher creativity levels - if you’d like them, sign up here and I’ll email them to you. I went all the way up to creativity level 1.8, at which point it generated a bird simply called “Strange”.