The Deep Learning Revolution in MIR: The Pros and Cons, the Needs and the Challenges

This paper deals with the deep learning revolution in Music Information Research (MIR), i.e. the switch from knowledge-driven hand-crafted systems to data-driven deep-learning systems. To discuss the pro and cons of this revolution, we first review the basic elements of deep learning and explain how those can be used for audio feature learning or for solving difficult MIR tasks. We then discuss the case of hand-crafted features and demonstrate that, while those where indeed shallow and explainable at the start, they tended to be deep, data-driven and unexplainable over time, already before the reign of deep-learning. The development of these data-driven approaches was allowed by the increasing access to large annotated datasets. We therefore argue that these annotated datasets are today the central and most sustainable element of any MIR research. We propose new ways to obtain those at scale. Finally we highlight a set of challenges to be faced by the deep learning revolution in MIR, especially concerning the consideration of music specificities, the explainability of the models (X-AI) and their environmental cost (Green-AI).