When he's not pushing the boundaries of virtual reality as part of his day job, Casey Rodarmor is something of a nostalgist for old-school computer graphics. He loves ANSI art, a more advanced form of ASCII art that supports both foreground and background colors, as well as a larger character set that includes several characters specifically designed for drawing. ANSI art first became popular in the pre-Internet days of dial-up bulletin board systems, because it was a low-bandwidth way to draw colorful graphics on text-only terminal screens. And while the BBS era is long dead, the ANSI art scene is still going strong.
What turned Rodarmor on to the idea of teaching a neural network to draw using ANSI art was the sheer ingenuity of the form. Like ASCII art, it takes incredible skill to create a recognizable picture using just 256 characters; as such, every ANSI artist has his own hallmark tricks to get the effects he is looking for. "Human-made ANSI art often contains very clever or unique details which would be very difficult for a neural network to learn," Rodarmor explains. "Neural networks need to see many, many variations of a pattern in order to learn it, so a clever flourish that appears just a few times is unlikely to be picked up by a neural network."
Not knowing what would happen, Rodarmor decided to feed 35 years worth of ANSI art into a neural network to see if it could learn the form, which in turn amounts to an almost inconceivably meager 32MB of data. He stripped out color to make it easier for a neural network to learn from the ANSI art, then gave it four days to train. After 96 hours, Rodarmor's Artnet started spitting out ANSIs of its own.
"It's always a rare treat when I write a program whose behavior surprises me, and this was one of those instances," says Rodarmor. "I had absolutely no idea what the output would look like. When it started to produce good stuff I was pretty thrilled."
Artnet's ANSIs are admittedly a little basic. Like a lot of ANSI art, many of Artnet's finished pieces resemble low-res graffiti tags, but unlike graffiti tags, you can't read them. They're all abstract geometry, which, if you squint a little, you can imagine as stylized lettering you can't quite make out. In addition, Rodarmor says his neural network is proficient at producing "huge meandering clouds of block letters" which are its version of ANSI paintings. Unlike traditional ANSI paintings, though, Rodarmor says his neural network hasn't quite become proficient at creating the Rat Fink-like monster and human characters that are a mainstay of the form.
Even if the results are a little abstract, Rodarmor says people are still connecting with his neural network's ANSI art, for many of the reasons that low-res ANSI art became popular in the first place: People are really, really good at creatively interpreting shapes. "I spent a fun 15 minutes today arguing with a co-worker who swore that one of the images looked like a picture of a coquettish woman staring longingly at Nigel Thornberry," Rodarmor says.