Center for Computer Research in Music and Acoustics

Abstract: Nowadays deep learning-based approaches have become popular tools and achieved promising results in machine listening. However, a deep model that generalizes well needs to be trained on a large amount of labeled data. Rare, fine-grained, or newly emerged classes (e.g. a rare musical instrument or a new sound effect) where large-scale data collection is hard or simply impossible are often considered out-of-vocabulary and unsupported by machine listening systems. In this thesis work, we aim to provide new perspectives and approaches to machine listening tasks with limited labeled data. Specifically, we focus on algorithms that are designed to work with few labeled data (e.g. few-shot learning) and incorporate human input to guide the machine.