@reading-and-thinking - Tumblr Blog

Superheroes designed by neural network

I trained the neural network to generate superhero names, based on the list from this site. I thought the database was going to be way too small, but the network proved me wrong.

Speet Stank Red Fart Mister Man Rad Food Sapgirl Woop Ann Man Boomss Boark II Supperman Superbore Slonk Lid Man Green Hooter II Starm Surper Shartar Goons Nana Rider Farm Captain In Redink Wolver Man Wizler

Supperman Nana Superbore Red Fart

In Which a Neural Network Learns to Tell Knock-Knock Jokes

The neural network can be trained to write recipes, invent Pokemon, and invent superhero names. But can it learn to tell a joke?

@researchbuzz generously provided me with a list of 200 knock-knock jokes - brief and highly formulaic, they seemed to be the form of joke best-suited for a neural network to reproduce. I figured it would quickly learn the formula, but would never, never manage to tell an actual joke.

It proved me wrong.

Very early on, the network figured out that question marks were important somehow, as were the words “Knock” and “who” in some mysterious way:

Koo k ano? ? Khnck Knock oae?t rA noc? Knock Woock Whecr lwhot h o nwoe hio

In very short order, it learned to begin a knock-knock joke, and almost never got that part wrong again. But the rest of the formula took quite a while to figure out, and the algorithm frequently devolved into playing a game of telephone with itself:

Knock Knock Who’s There? Iane Aatar who? Aaane who? Aan who? Anac who? Iobe who? Irata who?

Next it figured out the formula, but no actual, you know, words:

Knock Knock Who’s There? Bool Hane who? Scheres are then the there

Then it learned the joke about the cow with no lips which became its Favorite Thing Ever, even if it didn’t know how to properly tell it. The punchline to Every Single Joke was ooo ooooo ooo…

Knock Knock Who’s There? Alec Anic who? Wow co poo the tho tho door to the soout oo ooo oooo ooo oo oo oo oo ooo oo oo oo ooo ooo oooo ooo ooo oo oo ooo ooo ooo ooo ooo oo oo ooo ooo ooo ooo oo oo oo ooo ooo ooo ooo…

It made it past this stage, but still had problems that were remarkably similar to two people telling knock-knock jokes over a very bad phone connection, and maybe forgetting who was telling the actual joke:

Knock Knock Who’s There? Alther Water who? Water who who? Who who? Arton and out on the moor?

At last it figured out the basic formula and proceeded to make jokes that, while technically correct, didn’t make all that much sense. Many of them were partially plagiarized from jokes in the training dataset.

Knock Knock Who’s There? Hiwl Howlowwhowhow

Knock Knock Who’s There? Ireland Ireland who? Ireland you money, butt.

And then.

It produced. An actual joke. I checked, and this one most definitely wasn’t in the dataset. There was one about Alec, but the pun was “Alec-tricity”. There was one about knock-knock jokes themselves, but it was “Irish I knew more knock-knock jokes”. And it didn’t produce this just once, but over and over again. I give you possibly the freakiest thing the neural network has done to date:

Knock Knock Who’s There? Alec Alec who? Alec- Knock Knock jokes.

A neural network generates car names

I’ve trained this open-source neural network framework on a variety of datasets, including recipes, Pokemon, knock-knock jokes, pick up lines, and D&D spells.

This time, the dataset was a list of almost 2000 car models, helpfully provided by readers Anya and Dan. Can the neural network learn to suggest plausible car models? (Spoiler: Well, kinda…)

The neural network had to start from scratch, figuring out first what line breaks were, and then how often it should use them (early iterations consisted of long streams of line breaks and nothing else, the network apparently having figured that since it couldn’t yet match the rest of the data, at least it knew how to do line breaks).

Then there was a strange stage where every car name sounded like it was invented by cats:

Maru a Marr Mauu a Mauu e Mauu Maru e Mauu

Followed by another stage where it had figured out how to use other letters and numbers, but not how to use them wisely. It ends up sounding less like cars and more like early internet screennames:

Fatiuln Fo NT-Fi0i Defruti Xo5m Trfamlut Coe0r Mercuren-Be B-Toqu Fozurn G0tcakgn Mabuttsoee E FLlp0oo2dr ST4rr

We proceed from there to what I like to call the Dracula phase, where it begins to try to spell some of the major car makes, but ends up (mostly) sounding sinister instead:

Morcula Sapira Mercult Sarii Deriula Framno Moroa 3-S Darcult Sar Mlrru Pie Nunsgan Caakes

It continues to learn, and now we have knock-offs of the major brands, the sort of fine car you might buy at a very, very discount car lot:

Mosseni-Binz C SL5 Perrari 9r0 Sugkswagen Cepgyst Posda Suzek Subaau Veciro Fonda Caxii Nirkswagen Tiatukrr Vinkswigen Xeama Coyota Slnnae ToyoMa 64

Finally, the neural network has grown more sophisticated, and has learned to spell the major car brands. It can generate the simple model names, the ones that are just a letter or two and a number. However, I would not put it in charge of marketing just yet…

Volkswagen Colon Buick Shoat Buick Crapara Buick Apron Fiat Deter Fiat Coma Fiat S-0-S Fiat Doug

"Fiat Doug"

The Silicon Gourmet: training a neural network to generate cooking recipes

Neural networks are computer learning algorithms that mimic the interconnected neurons of a living brain, managing astonishing feats of image classification, speech recognition, or music generation by forming connections between simulated neurons. I’m not a neural network researcher, but there’s never been a better time to experiment with them, thanks to open-source packages like torch, a scientific computing framework with built-in neural network capabilities. Inspired by Tom Brewe’s neural network-generated recipes, and enabled by the open-source torch add-on for character-based neural networks by Andrej Karpathy, I fired up the neural network code on my 2010 Macbook Pro, and started training it on a bunch of recipes I downloaded from a collection by David Shields. Here’s a recipe my network has generated:

Pears Or To Garnestmeam

meats

¼ lb bones or fresh bread; optional ½ cup flour 1 teaspoon vinegar ¼ teaspoon lime juice 2 eggs

Brown salmon in oil. Add creamed meat and another deep mixture. Discard filets. Discard head and turn into a nonstick spice. Pour 4 eggs onto clean a thin fat to sink halves.

Brush each with roast and refrigerate. Lay tart in deep baking dish in chipec sweet body; cut oof with crosswise and onions. Remove peas and place in a 4-dgg serving. Cover lightly with plastic wrap. Chill in refrigerator until casseroles are tender and ridges done. Serve immediately in sugar may be added 2 handles overginger or with boiling water until very cracker pudding is hot.

Yield: 4 servings

This is from a network that’s been trained for a relatively long time - starting from a complete unawareness of whether it’s looking at prose or code, English or Spanish, etc, it’s already got a lot of the vocabulary and structure worked out. This is particularly impressive given that it has the memory of a goldfish - it can only analyze 65 characters at a time, so by the time it begins the instructions, the recipe title has already passed out of its memory, and it has to guess what it’s making. It knows, though, to start by browning meat, to cover with plastic wrap before chilling in the refrigerator, and to finish by serving the dish.

Compare that to a recipe generated by a much earlier version of the network:

Immediately Cares, Heavy Mim

upe, chips

3 dill loasted substetcant 1 cubed chopped whipped cream 3 unpreased, stock; prepared; in season 1 oil 3 cup milk 1 ½ cup mOyzanel chopped ½ teaspoon lemon juice 1 ¼ teaspoon chili powder 2 tablespoon dijon stem – minced 30 dates afrester beater remaining

Bake until juice. Brush from the potato sauce: Lightly butter into the viscin. Cook combine water. Source: 0 25 seconds; transfer a madiun in orenge cinnamon with electres if the based, make drained off tala whili; or chicken to well. Sprinkle over skin greased with a boiling bowl. Toast the bread spritkries.

Yield: 6 servings

which bakes first, has the source in the middle of the recipe directions, mixes sweet and savory, and doesn’t yet know that you can’t cube or chop whipped cream. An even earlier version of the network hasn’t yet figured out how long an ingredients list should be; it just generates ingredients for pages and pages:

Tued Bick Car

apies

2 1/5 cup tomato whene intte 1 cup with (17 g cas pans or ½ cup simmer powder in patsorwe ½ tablespoon chansed in 1 ½ cup nunabes baste flour fite (115 leclic 2 tablespown bread to ¼ cup 12". oz mice 1 egg barte, chopped shrild end 2 cup olasto hote ¼ cup fite saucepon; peppen; cut defold 12 cup mestsentoly speeded boilly,, ( Hone 1 Live breseed 1 22 ozcugarlic 1 cup from woth a soup 4 teaspoon vinegar 2 9/2 tablespoon pepper garlic 2 tablespoon deatt …

And here’s where it started out after only a few tens of iterations:

ooi eb d1ec Nahelrs egv eael ns hi es itmyer aceneyom aelse aatrol a ho i nr do base e2 o cm raipre l1o/r Sp degeedB twis e ee s vh nean ios iwr vp e sase pt e i2h8 ePst e na drea d epaesop ee4seea .n anlp o s1c1p , e tlsd 4upeehe lwcc eeta p ri bgl as eumilrt

Even this shows some progress compared to the random ASCII characters it started with - it’s already figured out that lower case letters predominate, and that there are lots of line breaks. Pretty impressive!

"Bake until juice."

Fun with a neural net that transforms line drawings to cats

So there’s a neural network framework called pix2pix that can be trained to learn how to transform one type of image into another. It’s been used for example to convert satellite images into line drawings and vice versa (think Google maps satellite view vs the one that show boxes where all the buildings are). There’s also trained version of pix2pix that converts a block drawing to a building facade (although it tends to make the building facades into creepy post-apocalyptic burnt-out shells).

Less useful but arguably even more creepy is the version by Christopher Hesse that turns line drawings into cats.

He trained it on about 2000 stock photos of cats that he ran through an edge-detection filter - he used the edge-detected (line drawing) versions as the starting image, and the original image as the ending image. The neural network then learned to convert from the line drawings into the original cat photographs.

The result? A neural network that tries to make ANY line drawing into a photo of a cat. And, gloriously, you can play with it in your browser here. Here are some results I’ve produced.

Here, it was able to turn my primitive trackpad-drawn scribble into a furry cat, complete with fluffy underbelly. Not bad, neural network.

But, if required, the neural network can make do with less. The results get increasingly creepy, though.

The neural network has not been trained on roosters. The network has only been trained on cats. 12/10 Excellent bright-eyed cat. Would stroke its strange furry head-crest.

This is a strange grabby cat, but at least the neural network has identified the eyes.

Ears? Check. Eyes? Check. Fluffy underbelly? Check. Next!

robotlaughs

tried some of my own

uhhh…

WHAT HAVE YOU DONE

New paint colors invented by neural network

So if you’ve ever picked out paint, you know that every infinitesimally different shade of blue, beige, and gray has its own descriptive, attractive name. Tuscan sunrise, blushing pear, Tradewind, etc… There are in fact people who invent these names for a living. But given that the human eye can see millions of distinct colors, sooner or later we’re going to run out of good names. Can AI help?

For this experiment, I gave the neural network a list of about 7,700 Sherwin-Williams paint colors along with their RGB values. (RGB = red, green, and blue color values) Could the neural network learn to invent new paint colors and give them attractive names?

One way I have of checking on the neural network’s progress during training is to ask it to produce some output using the lowest-creativity setting. Then the neural network plays it safe, and we can get an idea of what it has learned for sure.

By the first checkpoint, the neural network has learned to produce valid RGB values - these are colors, all right, and you could technically paint your walls with them. It’s a little farther behind the curve on the names, although it does seem to be attempting a combination of the colors brown, blue, and gray.

By the second checkpoint, the neural network can properly spell green and gray. It doesn’t seem to actually know what color they are, however.

Let’s check in with what the more-creative setting is producing.

…oh, okay.

Later in the training process, the neural network is about as well-trained as it’s going to be (perhaps with different parameters, it could have done a bit better - a lot of neural network training involves choosing the right training parameters). By this point, it’s able to figure out some of the basic colors, like white, red, and grey:

Although not reliably.

In fact, looking at the neural network’s output as a whole, it is evident that:

The neural network really likes brown, beige, and grey.

The neural network has really really bad ideas for paint names.

staff

1. @lewisandquark is doing brilliant things with neural networks. 2. You should probably take this opportunity to tag yourself.

bbcamerica

Us too.

doctorwho

Same.

Same.

I LOVE this new season of doctor who.

goulds

He’s beauty and he’s grace

Say racist shit and he’ll punch you in the faaaaaaaace

amanitacaplan

Bless this episode

halorvic

✌️