C
Chappie
Text
Unrolled Creative Adversarial Network for Novel Music Generation
An adversarial deep-learning system for generating novel symbolic music. It adapts the Creative Adversarial Networks (CAN) framework from visual art to music by representing MIDI pieces as images, and combines CAN's style-ambiguity objective with the unrolled-GAN optimisation procedure so the generator anticipates several future discriminator updates. This unrolled CAN reduces mode collapse and produces music that deviates from individual composers' styles; novelty is scored with an expert-gate autoencoder via reconstruction error.