about artPainter
current system status : online
update 5.6.05: artPainter is on display on at the Silpe Gallery at the Hartford Art School. For more info go here.
update 4.7.05: We aiming to get artPainter fully operational this weekend or early next week, stay tuned! the physical aspect of the system is nearly complete, it created its first test image today, a self portrait, viewable in the image section of this site. I'm demonstrating the system this weekend at the Trinity College Firefighting Robot Contest (not competing, obviously)
this is the front end for control of artPainter, an artificial neural network based attempt to simulate human creativity.
Neural networks are excellent at pattern recognition, and combined with some long term associative memory, can be used to create categorized lists of almost any kind of data. The specific data used here are vector records of brushstrokes, entered through the input brushstrokes page. Feel free to add your own 'thoughts' to artPainter!
These vectors are categorized by the network, then the network uses a variety of different means to create novel creations from this data.
- neuron death - as the network builds up weights and connections from the data, the effect of each image is upon the matrix of weights which represent the categories the system has created. By altering this map by introducing random deletion of weights and neurons, the system is forced to re-calibrate itself into a new configuration, which results in a unpredictable new take on the existing data, without destroying its existing neural connections through over training.
- self referential novelty - artpainter checks to see if its own creation is novel by attempting to categorize it itself, if it is categorized correctly, it starts again with the randomization process. In addition, each image the system creates which does qualify as novel is introduced into the system, which means it will now be able to recognize things which refer to its creations. while novelty is not a prerequisite of creativity, it does mean it won't be creating the same old picture each time it paints.
- improvisation - simply cycling through the different categories and extracting some data from each would result in novel compositions which reference what the system has learned.
- dynamic categorization - Each category is tweaked by data which falls under its dominion, so a "cat" category would respond to everything it sees as a cat, and alters its boundaries of the "cat" category in order to reflect this new information on cats.
Note that once the creative process is finalized, it will be left as is, to allow the system to work in a completely unsupervised. This allows it to qualify as a Turing test, if the viewer looks at all the existing drawings, and is unable to discern whether or not one is machine generated or not, it can be argued that artPainter has displayed artificial intelligence.