Unrolled Creative Adversarial Network For Generating Novel Musical Pieces

TL;DR

This work tackles the challenge of generating novel music with adversarial networks by extending Creative Adversarial Networks (CAN) to the music domain and introducing Unrolled CAN to mitigate mode collapse. It uses three MIDI-based datasets (jazz, classical by composers, and Arabic) and converts MIDI into image representations to train GANs and CAN variants. Results show that CAN achieves strong stylistic deviation but can suffer from mode collapse, while Unrolled CAN improves diversity and novelty (MSE up to ~0.043) by incorporating future discriminator dynamics. The study demonstrates that unrolled adversarial training can yield more creative and varied musical outputs, with potential applications in AI-assisted composition and playlist generation.

Abstract

Music generation has emerged as a significant topic in artificial intelligence and machine learning. While recurrent neural networks (RNNs) have been widely employed for sequence generation, generative adversarial networks (GANs) remain relatively underexplored in this domain. This paper presents two systems based on adversarial networks for music generation. The first system learns a set of music pieces without differentiating between styles, while the second system focuses on learning and deviating from specific composers' styles to create innovative music. By extending the Creative Adversarial Networks (CAN) framework to the music domain, this work introduces unrolled CAN to address mode collapse, evaluating both GAN and CAN in terms of creativity and variation.

Figures (17)

• Figure 1: Symphony No.46 in B major, Haydn
• Figure 2: Mozart Piano Quartet in G minor
• Figure 3: Treble clef and Bass clef
• Figure 4: Same-Different exercise example schon2001naming
• Figure 5: Example of standard CAN output with $k=0$, showing convergence to nearly identical samples after 20 epochs.