Kate Bush is my mom

Life is a costume away, oh, life's a wardrobe change

So I can shift, create a rift with beauty and with grin

Blush sensation creates the foundation, God is in my skiinnnn!

A new paper from NVIDIA recently made waves with its photorealistic human portraits. Called StyleGAN, the algorithm had a new training dataset pulled from Flickr, with a wider range of ages and skin tones than in other portrait datasets. (If your pictures are on Flickr with the right license, your picture might have been used to train StyleGAN). Thanks to that big dataset, a new method of generating images, and a staggering amount of computer power, its human faces are indeed impressive. But, to prove that their method doesn’t just work for human faces, they also generated bedrooms, and cars… and cats.

The cats are so much fun.

In many ways, generating cats is more of a challenge than generating human faces, since cats can be in so many different poses. It has an easier time generating texture than figuring out where all the legs and tails go.

I’ve noticed that for some reason, whenever it generates kittens, one is normal and the rest are haunted. There must be something difficult about pictures containing multiple subjects.

Its humans are even worse, proving that the difference between it and the version that was impressively good at generating human faces was all in the training data.

Speaking of training data, StyleGAN’s training data was something called LSUN Cat. Evidently, LSUN Cat’s images were sourced from the internet, because when it generated some cats, it added meme text to them, assuming that white blocky lettering is just part of what “cat” is.

It also attempted to generate cats with the Shutterstock watermark, although actually spelling “Shutterstock” proved tricky for it.

Another delightful thing about its internet-derived dataset: quite a lot of the cats look an awful lot like Grumpy Cat. She’s only one cat, but she had a profound impact on what StyleGAN thinks cats look like.

There are far more amazing cats than would fit in this blog post. In this bonus material, I’ve collected a few more of my favorites, including several with meme text, perfect for your internet communication needs. You can get them by entering your email here.

So I’ve used neural networks to generate recipes in the past. They’re computer programs that can learn to imitate the data we give them, copying the way that humans drive cars, label images, or translate languages.

That is, they try to learn. They’re called “neural” because they have virtual neurons that work a little like the real neurons in our brains. Their virtual brains, however, are really tiny. Where a human has about 86 billion neurons, the neural networks we use today have hundreds to low thousands - think nematode worm or, optimistically, jellyfish. So they often struggle, and recipes, my friends, are one of those times. (you haven’t lived until you’ve seen a neural net’s attempt at “small sandwiches”)

I turned to cocktails, inspired by Beth Skwarecki’s cocktails bot (and helped immensely by the dataset she sent me).

The neural net’s first attempt was… an attempt.

I let it train for a little longer, and things got a little more recognizable, as it learned to kinda spell more ingredients.

It had, however, not learned when to quit. 64-ounce (2 liter) cocktails are not unheard of at this stage.

It continued to learn on its own, all without input from me, and eventually ended up with cocktails that were almost plausible.

Some of my other training attempts, however, did not go so well. At first the cocktail titles in my training set were in all caps, which confused the heck out of the neural net because capital letters were so rare that it didn’t see enough examples of what to do with them. It struggled with the titles.

You may also notice that the ill-fated all-caps attempt had a lot of repetition - that’s because I was using textgenrnn, which has a somewhat limited memory. It could learn to spell ingredients, but had no idea whether it had already added sugar and cream. When I switched to char-rnn for the lowercase recipes above, I could give it a memory of 50 characters, enough to cut down on repetition. The textgenrnn version, however, became strangely obsessed with creme de cacao. The less said about its cocktails, the better.

For more neural net cocktails (including custom-generated cocktails for any name you care to provide), check out Beth’s cocktails bot!

So I’ve used neural networks to generate recipes in the past. They’re computer programs that can learn to imitate the data we give them, copying the way that humans drive cars, label images, or translate languages.

That is, they try to learn. They’re called “neural” because they have virtual neurons that work a little like the real neurons in our brains. Their virtual brains, however, are really tiny. Where a human has about 86 billion neurons, the neural networks we use today have hundreds to low thousands - think nematode worm or, optimistically, jellyfish. So they often struggle, and recipes, my friends, are one of those times. (you haven’t lived until you’ve seen a neural net’s attempt at “small sandwiches”)

I turned to cocktails, inspired by Beth Skwarecki’s cocktails bot (and helped immensely by the dataset she sent me).

The neural net’s first attempt was… an attempt.

I let it train for a little longer, and things got a little more recognizable, as it learned to kinda spell more ingredients.

It had, however, not learned when to quit. 64-ounce (2 liter) cocktails are not unheard of at this stage.

It continued to learn on its own, all without input from me, and eventually ended up with cocktails that were almost plausible.

Some of my other training attempts, however, did not go so well. At first the cocktail titles in my training set were in all caps, which confused the heck out of the neural net because capital letters were so rare that it didn’t see enough examples of what to do with them. It struggled with the titles.

You may also notice that the ill-fated all-caps attempt had a lot of repetition - that’s because I was using textgenrnn, which has a somewhat limited memory. It could learn to spell ingredients, but had no idea whether it had already added sugar and cream. When I switched to char-rnn for the lowercase recipes above, I could give it a memory of 50 characters, enough to cut down on repetition. The textgenrnn version, however, became strangely obsessed with creme de cacao. The less said about its cocktails, the better.

For more neural net cocktails (including custom-generated cocktails for any name you care to provide), check out Beth’s cocktails bot!

So there’s these computer programs called artificial neural networks that are good at imitating things. By seeing examples of what humans did, they can learn to translate languages, predict product sales, and even categorize text and images as innocuous or explicit (it has a lot of trouble with this last task, as it turns out).

One neural network I use, called textgenrnn, tries its best to imitate any kind of text you give it. I’ve given them paint colors, band names, and even guinea pig names and in each case their results are somewhat… mixed. (Paint colors called Stanky Bean, Stargoon, and Turdly, for example) The problem is that it doesn’t know what any of these words mean - it’s just picking letter combinations that seem likely to it.

This is what happened when I gave it all the cookies from a list of American recipes. This is what human cookies sound like to a neural network.

Now if you’ll excuse me, I’m going to go whip up a batch of Fluffin Coffee Drops.

For more cookies, including the neural net’s strange obsession with “balls,” as well as bonus material every time I post, you can sign up here.

oh yeah baby that’s right 

Francisco Goya, The Third of May 1808 (also known as El tres de mayo de 1808 en Madrid or Los fusilamientos de la montaña del Príncipe Pío, or Los fusilamientos del tres de mayo), 1814

Edouard Manet, The Execution of Emperor Maximilian, 1867