Keynote Speaker

Rachel Bittner / Spotify
2023-11-08 | 09:00 (Europe/Rome)
Zoom meeting
There is a considerable gap in the research and engineering methods we use to build MIR systems for academic research and the way we build them for industry-scale systems. This keynote covers some of the many differences and challenges faced when building MIR systems for industry applications. We first discuss the way we define problems in the first place, and why the academic definition of problems is often ill-suited for a particular application. There are also substantial differences in engineering workflows – in particular when multiple researchers and engineers build a single system. We explore differences in academic datasets which are usually “small and clean” to real-world datasets which are “large and noisy”. Academic metrics are useful for us scientists, but they often either don’t match a product use case or mean nothing to product teams. Finally, we dig into deployment considerations including how to run inference flexibly, considering cost and speed, and where the system needs to run. We will explore numerous real-world examples throughout and provide insight into how to build MIR systems within industry.
Rachel is a Research Manager at Spotify in Paris. Before Spotify, she worked at NASA Ames Research Center in the Human Factors division. She received her Ph.D. degree in music technology and digital signal processing from New York University. Before that, she did a Master’s degree in Mathematics at New York University, and a joint Bachelor’s degree in Music Performance and Math at UC Irvine. Her research interests include automatic music transcription, musical source separation, metrics, and dataset creation.

Poster Sessions

Industry Session

Industry Presentations

Moderator: Ilaria Manco

2023-11-08 | 18:30 (Europe/Rome)

Special Sessions

Panel Session: Hybrid deep learning for MIR

Moderator:Gaël Richard Panelists: George Fazekas (Queen Mary Univ) Changhong Wang (Telecom Paris) Zhiyao Duan (University of Rochester) Gus Xia (MBZUAI)

2023-11-08 | 17:30 (Europe/Rome)

Zoom meeting
In MIR, as in many other domain, there is a significant trend towards purely data-driven approaches aimed at directly solving the machine learning problem at hand, while only crudely considering the nature and structure of the data being processed. In the music domain, prior knowledge can relate to the production of sound (using an acoustic model), the way music is perceived (based on a perceptual model), or how music is composed (using a musicological model). These models can usually be encoded with only a few parameters, leading to controllable and interpretable systems that can be exploited in modern neural-based machine learning frameworks, resulting in so-called hybrid deep learning models. The aim of this panel is to illustrate the concept of hybrid deep learning with some specific examples in MIR, and to discuss its limits, merits and potential for future machine learning based music applications.

Social Events

Banquet, Jazz Concert and Jam Session

2023-11-08 | 20:00 (Europe/Rome)