The aesthetics of deconstruction: neural synthesis of transformation matrices using GANs on multichannel polyphonic MIDI data

Abstract

Deconstruction, a concept that originated from French philosopher Jacques Derrida's post-structuralist theories, has significantly influenced architecture and graphic design. It has evolved into a rhizomatic network of styles and concepts ranging from new architectural geometry to AI. Despite its limited explicit application in music, this paper explores deconstruction within musical aesthetics, especially in the post-1945 serial school. Computer-Aided Design (CAD) methods have become an important part in the deconstructivist architectural design process; similarly, in music, Computer-Aided Composition (CAC) techniques have become important from generative algorithms to AI techniques. In graphic design, such systems (i.e. GANs, VAEs, CNNs, RNNs, Stable Diffusion, Transformers, etc.) have been experimented with. Google's Imagen or OpenAI's DALL-E 2, which automatically generate images from text prompts given by users, use, for example, diffusion models. Also, Google LM is a music equivalent to these. This paper delves into the novel aesthetics brought about by the application of GANs on multichannel polyphonic MIDI data in music, discussing the conceptual grounding of such usage and overcoming the challenges associated with it. Effectively, using our concept of distance matrices rather than, for example, Hidden Markov Models, is more suitable for GAN generation.