New insect species, dreamed by neural network
3 min read

New insect species, dreamed by neural network

New insect species, dreamed by neural network

Neural networks are a kind of machine learning program that imitate the way the human brain learns. They learn by example, rather than by being given rules by a human programmer. And because they can get very, very good at what they do, neural networks are now the algorithms that run a lot of the technology we interact with - everything from facial recognition to language translation to self-driving cars.

The tasks I ask them to do, however, are a lot more silly. I’ve asked them to generate new paint colors, new names for guinea pigs, and knock-knock jokes. Their results are usually very weird.

Recently, I asked them to generate new species of fish, based on a list of thousands of existing fish names. It worked very well, mostly because the fish names in the training data were so darn strange to begin with. Then I did the same with bird species.

Know what else is strange? Insect names.

Inspired by Mark Schwister’s suggestion, I contacted Joe Rominiecki of the Entomological Society of America, who kindly sent me a set of 2283 common names from abbreviated wireworm to Zimmerman pine moth.

Sure enough, the neural network learned to produce insect names that were (to the untrained eye) indistinguishable from the originals.

woollyborred wideworm
twolined blister beetle
cattle leafhopper
hairy spittlebug
black pine needleminer
redbacked elm shortworm
blueberry trumpet mite
ten pine billbug
pipevine weevil
smaller cherry leafhopper
threestriped flea beetle
American hairy spider mite
goldeneyed ash borer
woolly billbug

One interesting/infuriating thing that the neural network did was pick up on the ONE example where “leafroller” was misspelled in the dataset. It was spelled correctly 25 times and once it was spelled “leafroIler”. Depending on your font, you may not see that this was a capital i instead of a second L. The neural network noticed, though. And it did this. Not just one time, but every single time.

peach leafroEler
longlogged sucking leafroDer
ivberry leafroFler
redbanded leafroJler
inglavora pear apple leafroBer
lrueberry leafroMeed scale

It never once spelled “leafroller” correctly. Yet it had rather impressively figured out that capital letters are related in some way, based on the very few times that proper nouns were used in this mostly-lowercase dataset. The Entomological Society of America has promised to correct the error - maybe that’s all the neural net wanted.

Other than its thing with the leafrollers, the neural network did pretty well. Insect names are weird to begin with, and so they’re easy to fake. But I would not advise trying to get these past a roomful of entomologists:

soybean mall fly
magicking ant
dan beetle
chicken caterpillar
twostriped eat bug
goat pot weevil
so beetle
Friedfruit weevil
western black norpher
large moth
leaffooted caterpillar
smalled bear mite
woollybacked wap bug
sheep pine cone borer
black tumple beetle
Slender dumpler
More bug
leaf crumpet bumble bee
chicken leafcutting lady beetle
ungwood clumpler
bluster flower leafcutting s clothes moth
aurmegod shogheaded borer

And these seem like species to avoid.

King mosquito
Cussy caterpillar
doo bug
black lick beetle
tomato fire weevil
brown goat chafer

Become an AI Weirdness supporter to get bonus content! Or become a free subscriber to get new AI Weirdness posts in your inbox.

Subscribe now