Training Dynamics in Deep Learning for Computer Audition and Beyond
MCML Authors
Manuel Milling
* Former Member
Abstract
Manuel Milling
* Former Member
Abstract
This thesis analyses the interplay of data, neural networks and optimisation routines in their combined loss landscape. The particular focus is on audio data and its comparison to image data. Contributions lie in the development and application of deep learning quantifications, such as sharpness and sample difficulty estimation, to gain further insight about fundamental machine learning aspects such as generalisation, transfer learning, parameter adaptation and dataset characteristics in the given setting.
BibTeXKey: Mil25