Towards lightweight architectures for embedded machine learning in musical instruments

Tools

Kiefer, Chris (2022) Towards lightweight architectures for embedded machine learning in musical instruments. Sound and Music Computing, Saint Ettiene, France, 7 Jun 2022 - 11 Jun 2022. Published in: Proceedings of the 19th Sound and Music Computing Conference. Zenodo

[img] PDF - Published Version
Available under License Creative Commons Attribution.
Download (435kB)

Abstract

It can be challenging to engage with machine learning in restricted computing environments, such as the systems often in use in digital or hybrid musical instruments. We often use low power, low memory devices with limited computational power, and need high frequency models for sensor and sound processing. Conversely, contemporary machine learning and AI can be resource hungry, limiting its use in embedded systems. In previous research, the Stochastic Logic Optimisation algorithm offered a method of lightweight machine learning using two-state logic networks, intended for musical use in embedded systems. This experiment shows how this approach can be expanded on, using random boolean reservoirs, for signal generation. These initial results demonstrate the efficacy of a reservoir computing approach, built only from networks of lookup tables. They show that, for the task of training sine wave generators, reservoirs can be improved if built with hierarchical growth algorithms, and further improved by selecting inputs and outputs based on network centrality. The results also demonstrate successful use of Pulse Density Modulation for signal encoding.

Item Type: Conference Proceedings
Keywords: Machine Learning, Music, Embedded Systems, Logic Networks, Instrument Design, Digital Signal Processing
Schools and Departments: School of Media, Arts and Humanities > Music
Research Centres and Groups: Sussex Humanities Lab
SWORD Depositor: Mx Elements Account
Depositing User: Mx Elements Account
Date Deposited: 16 Jun 2022 14:51
Last Modified: 16 Jun 2022 14:51
URI: http://sro.sussex.ac.uk/id/eprint/106419

View download statistics for this item

📧 Request an update