Home | Publications | Kol26

Advances in Deep Learning: Differentiable Sparsity, Overparameterization, and Semi-Structured Regression

MCML Authors

Chris Kolb

Dr.

→ Group David Rügamer
Statistics, Data Science and Machine Learning

Abstract

This dissertation develops a differentiable framework for inducing sparsity in deep learning models, enabling efficient feature, parameter, and modality selection without relying on ad hoc pruning procedures. By leveraging overparameterization, it introduces principled sparsity-inducing methods with strong theoretical guarantees and combines them with interpretable semi-structured regression models, improving both model efficiency and explainability in multimodal learning. (Shortened.)

phdthesis Kol26