Recently, machine learning (ML) and in particular deep learning models have widely been evolving towards “end-to-end” designs, which naturally reduces their interpretability and behavioural predictability. This project aims to analyse and explain state-of-the-art ML models using tools from frame theory and time-frequency analysis and to design new, more accessible methods based on them.
Methodically, this project reaches from abstract frame theory addressing theoretical questions about stability and reconstructability of model architectures to the application-oriented usage of frames in terms of time-frequency representations being embedded within concrete ML model implementations.
Artificial neural networks have established themselves as very efficient type of ML model, making them a central object of application in this project as well. The full potential of frame theory is employed to study and strengthen these models with solid time-frequency analysis.