post-millenial

There was a paper recently where a research team trained a machine learning algorithm (a GAN they called AttnGAN) to generate pictures based on written descriptions. It’s like Visual Chatbot in reverse. When it was just trained to generate pictures of birds, it did pretty well, actually. 

(Although the description didn’t specify a beak and so it just… left it out.)

But when they trained the same algorithm on a huge and highly varied dataset, it had a lot more trouble generating a picture to go with that caption. Below, I give the same caption to a version of their algorithm that has been trained to generate everything from sheep to shopping centers. Cris Valenzuela wrapped their trained model in an entertaining demo that attempts to generate a picture for any caption.

This bird is less, um, recognizable. When the GAN has to draw *anything* I ask for, there’s just too much to keep track of - the problem’s too broad, and the algorithm spreads itself too thin. It doesn’t just have trouble with birds. A GAN that’s been trained just on celebrity faces will tend to produce photorealistic portraits. But this one, however…

In fact, it does a horrifying job with humans because it can never quite seem to get the number of orifices correct.

It’s fun to ask it to draw animals though. It knows the texture of giraffes, but not quite exactly their shape. And it knows that boats are on the water, but not necessarily that they are boats.

It also (like many other image recognition algorithms) gets a bit confused about the difference between sheep and the landscapes they’re found on. Other algorithms recognize sheep in pictures of empty green fields. And this one, when asked to draw sheep…

That’s different, though, from asking it to draw *a* sheep. In that case, it knows exactly what to do. It draws the sheep, and then just to be safe it fills the entire planet with wool too.

It really likes drawing stop signs and clocks. Give it the slightest opportunity to draw one, and it will chuck those things all over the place.

Other than its horrifying humans, this algorithm can actually be pretty delightful. 

Try it for yourself!

no. tell me what you fixed. i wanna know what the glitch was.

you know those patch notes that are like “fixed an issue where if the player sat in a bush for too long, they’d become the size of a skyscraper” 

i have something set up to subscribe to RSS feeds and add them as issues onto a repo. its like a personalized feed of email newsletters.

Evolving Cellular Automata

Another project from Loren Schmidt, this is a rule-generator for cellular automata.

Cellular automata are handy, compact examples of emergent effects. The most well-known example is Conway’s Game of Life, which is a two-state cellular automata with the rules “B3/S23″. Cells are Born if they have three neighbors, and they Survive if they have 2 or 3. Other rules are possible: “B36/S23″, “B36/S125“, and “B4678/S35678“ are a few of the more interesting known rules.

This rule generator creates three-state automata, which opens up even more possibilities. It also uses evolutionary algorithms to create the next generation of rules: click on a pattern and get a set of rules similar to that pattern.

You can also save the rules you like in a numerical format you can use elsewhere, or go back to a saved rule and try evolving down a different path.

Just a list of resources I’ve found across the web on learning how to code. Feel free to leave an ask/submit if you find anything so I can add it! 

I will continuously add more links so please check back whenever. :)

Just a list of resources I’ve found across the web on learning how to code. Feel free to leave an ask/submit if you find anything so I can add it! 

I will continuously add more links so please check back whenever. :)