Quick, Draw! is yet another publicly-facing interesting neural network. It's pretty fun: Like Pictionary, you're given a sequence of things that you need to draw within 20 seconds, and the game tries to guess what you're drawing. I'm told to draw a house, I make a quick sketch of a house, and the game says "It's a house."
From a practical point of view, here's what's happening: Quick, Draw! is a computer that's been trained on a lot of different correlations. By the time I played the game, thousands of people had been told to draw a house. The game probably got the first few dozen of those wrong, but as time went on it became able to correlate that two-dimensional box with a triangle on top with the concept of house. Think of it like a spreadsheet where certain shapes make up columns and rows. When enough form together, the game can reliably recognize "house."
For this same reason, Quick, Draw! had a very hard time recognizing the pig that I drew. I did it from the side, a classic landscape pig drawing, and the game just wouldn't believe that I wasn't drawing a dog. At the end of a play session, the game presents you with images that it was "looking for," and the lack of recognition made sense; the majority of people have been drawing pig portraits.
The game application of neural networks is interesting like that. They're constantly reflecting the world that they exist in, and stuff that doesn't fit that set of parameters just never enters their computer minds. DeepDream, one of the most famous public neural networks, illustrates this perfectly. When you train your brain to look for dogs, the whole world starts looking like it's made of dogs.
It's probably worth thinking about how the way we're all teaching Quick, Draw! what a pig looks like is a kind of labor. We're working in a Montessori school for a weird little robot who just loves to play games, but we're also teaching a system how to better recognize normalized differences between shapes, lines, and figures.
In a time when policing and some of the largest caches of facial images are becoming more and more attached, the kinds of norms that neural networks look for have real, material consequences, and the inability for the machine to recognize my different angle on a pig points to where the gaps in neural network abilities lie.
I know, I know. It's a game. It's a pig. It's fun. Still, I'm always wary here, because it's easy to imagine a dystopian project of, say, identifying what a criminal looks like. When hit new games take on issues like these, we get excited for characters to smash the establishment and reveal the clockwork machinery of an unfair world. But there are other, smaller things that often get ignored, and paying attention to the details of small things always pays off in the long run.
But right now we're drawing houses and pigs.