In the last twenty years, the field of music information retrieval (MIR) has undergone rapid developments in terms of the problems considered, the methodology, and its applications. Using conceptually simple tasks and methods evaluated on small and idealized datasets in its beginnings, MIR now contributes to a wide range of concepts, models, and algorithms that extend our capabilities of accessing, analyzing, understanding, and creating music. Given the complexity and diversity of music, MIR research considers various aspects such as genre, instrumentation, musical form, melodic and harmonic properties, dynamics, tempo, rhythm, and timbre, to name a few. Furthermore, music is inherently multimodal, incorporating speech-like signals (e.g., singing), videos (e.g., of live performances), static images (e.g., scanned music scores), and text (e.g., lyrics and reviews). This wealth of data makes MIR an interdisciplinary and challenging field of research, which closely connects to technical disciplines such as signal processing, machine learning, and information retrieval, as well as mathematics, musicology, psychology, and the digital humanities.
In this Dagstuhl Seminar, we aim to advance technology and education in these disciplines using music as a challenging and instructive multimedia domain. Thinking of data-driven machine learning approaches, we will discuss recent deep learning (DL) approaches and their ability to learn from training examples to make accurate predictions for previously unseen data. Furthermore, by learning from the experience of traditional engineering approaches, our objective is to better understand existing and build more interpretable DL-based systems (e.g., through integrating prior knowledge).
Beyond these technically oriented perspectives, an essential focus of our seminar is to approach the concept of learning from other perspectives, including a pedagogical, educational, psychological, and user-centered one. We argue that music is an essential part of our lives that most people feel connected to. Therefore, music yields an intuitive entry point to support education in technical disciplines. In particular, we will explore how music may serve as a vehicle to make learning and teaching signal processing and machine learning an interactive pursuit.
From an application perspective, we want to learn with and from domain experts about musical works and their recorded performances while exploring the potential of computational tools by considering complex music scenarios of musicological relevance. Finally, through a dialogue with social scientists, we want to gain a deeper understanding of human perception and interpretation of music and technology.
In all perspectives on learning, the question of reproducible research, including open access to data and software, is becoming increasingly important (so future insights can be built upon existing ones). As an overarching topic of our seminar, we will discuss questions about open science and good scientific practice. This is a key issue, especially in higher education, where the open exchange of best practices and teaching materials significantly impact the interdisciplinary and transnational education of the next generation of researchers.
In summary, in our Dagstuhl Seminar we want to approach and explore the concept of learning from different angles. Besides considering technological developments, the seminar equally addresses aspects of data and model understanding, transdisciplinary methodology and applications, science communication, and education.
- Machine Learning
- Music information retrieval
- signal processing
- deep learning
- user interaction