14:40 - 15:00
Roger Dannenberg
Carnegie Mellon University
Structure and Learning in Music Generation
Music generation by computer has seen renewed interest, but we are still a long way from understanding how humans creative original musical ideas or even how to make conventional-sounding music compositions. Many researchers have been applying deep learning techniques to musical sequences in hopes of modeling music composition. While there are some impressive results, we should remember that Mozart did not listen to the complete works of Mozart before writing his first masterpiece, so the whole idea of training on models is questionable as a model of human creativity. Secondly, sequence learning seems to have great difficulties learning hierarchical structure and relationships that span longer time intervals. In my recent work, I have introduced explicit representations of structure to create conventional-sounding music. I will also share some ideas about how successes with this approach might inform learning-based strategies.
Design downloaded from free website templates.