Growing MIDI Music Files Using Convolutional Cellular Automata

This paper describes a novel generative system based on cellular automata (CA). The main contribution of this work is CARLA, a Cellular Automata Rule Learning Algorithm that learns CA update rules with lower computational overhead (i.e. a smaller neural architecture) than other current approaches. A formal model is developed, situating the algorithm in terms of machine learning. The algorithm is then validated on the 256 possible elementary cellular automata rules. Next, the CARLA algorithm is incorporated into a larger generative system that learns musical patterns by transforming MIDI-encoded music into cellular automata. A novel postprocessing algorithm called Tendril is then developed and integrated with CARLA. This system is then used to generate a novel MIDI composition based partially on update rules learned from Chopin's Waterfall Étude, demonstrating the utility of this new algorithm for creative applications.