Profile
Project title:
Deep learning for low-resource music
Abstract:
Self-supervised learning has been successfully used in a variety of deep learning applications to learn data representations by leveraging vast mounts of unlabelled data. Given the comparative scarcity of manually-annotated data in the field of Music Information Retrieval (MIR), there is potential and interest in using self-supervision to learn music audio representations that can then be used in various downstream MIR tasks.
In this PhD I am interested in investigating how self-supervised learning methods applied to music audio can be more robust and versatile. I believe that one of the underlying considerations to this question is what is the relation between a representation's versatility and its ability to perform well on individual tasks.
C4DM theme affiliation:
Music Informatics