Skip to main content
School of Electronic Engineering and Computer Science


Project title: Self-supervision in Machine Listening  

Industry partner: Bytedance 

C4DM theme affiliation: Music Informatics, Machine Listening 

Abstract: Recently, deep learning techniques have achieved excellent accuracy on a large number of supervised learning tasks. But models using such techniques are demanding in terms of quantity and quality of data annotation. To address the expensive time and labor cost of annotating, various means of self-supervised representation learning (SSRL) have been proposed.  These approaches train models from unlabeled data itself based on intrinsic similarities among data in various ways to obtain the resulting representations for downstream tasks. Some experiment suggest that the quality of SSRL may still be roughly logarithmic with the size of the annotated dataset, even though these data are not labeled. Such a phenomenon enables SSRL strategies to gain better performance on downstream tasks with smaller amounts of label. 


Back to top