[Thank you to my brother, Will Greig, for the video soundtrack.]
Through this Kickstarter, I hope to give Neko a brain. The thing all Machine Learning algorithms need is lots and lots of data. I can go through text books and online museum archives on my own, but what I want are individual responses for individual colors. Once all the donations are pledged, Neko will begin at the first tier and work his way up. The higher the donation, the more intelligent Neko will be by the time he gets to it. I'll also work reverse-chronologically, so that earlier donors get the more informed paintings.
I see robots as natural painters. They are patient, immune to toxins, and steady-handed. A robot’s ability to process massive amounts of data allow for new insight into patterns of beauty and symbolism. I think the emergence of a non-human standard of taste would have a beneficial impact on the art world. I've been building painting robots since 2008. I take my work slowly and seriously, only considering one small aspect of painting at a time.
Neko, my second oil painter, began working in the Fall. He's had shows in San Francisco and New York already. In his first iteration, he painted gradients of whatever two site-specific colors I put on his palette. Right now he only knows the pattern of dipping and applying to create a gradient. He's more like an old-fashioned automaton than a robot.
Bringing Neko to Life
Neko needs a learning algorithm. He makes paintings for specific people, in a type of portraiture. Right now I’m doing all the data collection for him: interviewing the sitter, reading books on pigments and color symbolism, testing different pigment blends. I’m collecting all my findings in a comprehensive database. Soon, Neko will be able to consult this database and incorporate his own findings.
I'd like Neko to use both supervised and unsupervised learning. I want him to teach me some novel color associations, so I'll work on a clustering algorithm to go through art history texts and look for frequently paired words and pigments. That's unsupervised learning. When Neko proposes a color for someone's portrait, I'll ask them to supervise his learning by honestly answering "yes" or "no" when he asks if they like it.
I built Neko out of parts from my old neighborhood store, Garber Hardware. To couple these parts to motors and sensors from Sparkfun, I design plastic couplers using a MakerBot Thing-O-Matic. An Arduino Mega joins the electronic components to an old MacBook. The Arduino IDE sufficed for the first version but I’ll be using Python for the database queries in the next implementation. I started a blog to track my progress and research: painterbot.blogspot.com
Where the Money Goes
- Updated sensors and motors
- Vision system
- Canvas, brushes & paint
- Painting & programming lessons
- Research materials
Thank you for taking the time to read my proposal. I hope you enjoyed it, and I look forward to hearing from some of you. The more people I get participating in the project, the richer our data will be.
Risks and challenges
Programming is my biggest challenge! I'm working my way through tutorials now, learning how to get really specific about the tasks at hand. I can usually hack something together quickly, but like to take my time in really getting it right. The reward tiers build up as stepping stones of progress and I think they're realistic. There are experts in a number of relevant disciplines I can turn to for help as needed.
Neko's already at the working-but-not-yet-right stage, both in hardware and software, so at least we can continue making paintings at the current level of complexity. Now it's just a matter of populating the database and getting the Arduino (or possibly Beaglebone) to communicate with the data via Python. I've used AppEngine before and think it'd work well for the project, and remove some of the risks.Learn about accountability on Kickstarter
- (30 days)