Deep learning approaches in music information retrieval

Abstract

With the increasing popularity of deep neural-based architectures, the results of deep architectures have been significantly improved recently in several areas. Due to the popularity and success of these deep approaches based on neural networks, other symbolic and hierarchical approaches are no longer the focus of researchers. In this article, we review the recent progress of deep and compositional approaches in the field of music information retrieval. Furthermore, we deliberate on the most notorious issues in the field and highlight problems where deep approaches based on neural networks have not yet been successfully applied. As an alternative to such approaches, we provide an overview of hierarchical models and describe the compositional hierarchical model as an alternative deep architecture. The latter shows great usability with the presented problems. We conclude this review with a discussion of the future of deep models compared to other approaches.