Deep Learning and Knowledge Integration for Music Audio Analysis (Dagstuhl Seminar 22082)

Given the increasing amount of digital music, the development of computational tools that allow users to find, organize, analyze, and interact with music has become central to the research field known as Music Information Retrieval (MIR). As in general multimedia processing, many of the recent advances in MIR have been driven by techniques based on deep learning (DL). There is a growing trend to relax problem-specific modeling constraints from MIR systems and instead apply relatively generic DL-based approaches that rely on large quantities of data. In the Dagstuhl Seminar 22082, we critically examined this trend, discussing the strengths and weaknesses of these approaches using music as a challenging application domain. We mainly focused on music analysis tasks applied to audio representations (rather than symbolic music representations) to give the seminar cohesion. In this context, we systematically explored how musical knowledge can be integrated into or relaxed from computational pipelines. We then discussed how this choice could affect the explainability of models or the vulnerability to data biases and confounding factors. Furthermore, besides explainability and generalization, we also addressed efficiency, ethical and educational aspects considering traditional model-based and recent data-driven methods. In this report, we give an overview of the various contributions and results of the seminar. We start with an executive summary describing the main topics, goals, and group activities. Then, we give an overview of the participants' stimulus talks and subsequent discussions (listed alphabetically by the main contributor’s last name) and summarize further activities, including group discussions and music sessions.