Assessing algorithmic biases for musical version identification | Furkan Yesiler

London Audio & Music AI Meetup (virtual) - 14 Oct 2021 @ 18.30 (BST)

We would like to invite you to our Audio & Music AI Meetup.
Featuring Furkan Yesiler, PhD candidate at Universitat Pompeu Fabra, presenting "Assessing algorithmic biases for musical version identification".

• Agenda:

• 18:25: Virtual doors open
• 18:30: Talk
• 19:15: Q&A
• 19:30: Networking
• 20:30: Close

• Zoom link
  The zoom link will be shared closer to the event date. Please keep an eye on this page and your inbox!

• Abstract
  Version identification (VI) systems now offer accurate and scalable solutions for detecting different renditions of a musical work (e.g., cover songs), allowing the use of these systems in industrial applications and throughout the wider music ecosystem. Such use can have an important impact on musicians and composers regarding recognition and financial benefits, including, but not limited to, how royalties are circulated for digital rights management.

In this talk, I will introduce our recent work where we took a step toward acknowledging this impact and considered VI systems as socio-technical systems rather than isolated technologies. We aimed to quantify the algorithmic biases of 5 systems by investigating discrepancies in their performances on pairs of potentially impacted groups, mimicking a group fairness paradigm. We categorized such groups using the main stakeholders (i.e., musicians and composers) and 6 relevant side attributes (gender, popularity, country, language, year, and prevalence). We found that VI systems may indeed favor certain groups over others, and their behavior may vary depending on whether they are learning- or rule-based, or whether they use melody- or chroma-based input features. We also proposed our hypotheses on the possible causes of the observed disparities and discussed the potential implications of the results on the considered stakeholders from a fairness perspective. We hope that this work will encourage future VI research to incorporate fairness- and algorithmic bias–related evaluation metrics along with the existing accuracy- and scalability-related ones, to get ahold of any wrongful practices.

*Paper
https://arxiv.org/abs/2109.15188

• Follow Furkan
  https://twitter.com/furkanyesiler?lang=en

• Bio
  Furkan Yesiler is a Ph.D. candidate as a part of the MIP-Frontiers project (MSCA Grant No:765068) at Music Technology Group, Universitat Pompeu Fabra (Barcelona). His Ph.D. research is focused on incorporating deep learning techniques to build accurate and scalable musical version identification systems for industrial use-cases. He received his MSc degree in Sound and Music Computing also at MTG, UPF with his thesis on singing voice research. He graduated summa cum laude with two BSc degrees in computer engineering and industrial engineering from Koc University (Istanbul) where he was accepted with a full scholarship. During his bachelor’s studies, he did internships in management consulting, and mergers and acquisitions advisory companies in Istanbul.

• Host
  Kobalt Music: https://www.kobaltmusic.com/

• Sponsors
  AI.Music: https://www.aimusic.co.uk/
  IEEE Signal Processing Society. http://www.signalprocessingsociety.org/