Defense of Doctoral Thesis with with Yi‑Hsin Li
Yi-Hsin Li will present her thesis "Steered Mixture of Experts for Compact and Edge-aware Representation: From 2D Image Regression to 3D Radiance Fields" in Technical University. Berlin, and on Zoom.
About the thesis
Visual computing requires increasingly detailed yet effective representations. At the same time, 2D models get stuck in rigid grids and many 3D methods only become realistic at the cost of long training times and high resource demands. This thesis therefore addresses a central question: how do you create scalable, high-quality visual representations without sacrificing efficiency?
The thesis focuses on Steered Mixture-of-Experts (SMoE), a modular architecture with local cores and good interpretability, but which has long been plagued by heavy training, large models and limited support for high-dimensional data. The work answers three questions that aim to make SMoE fast, compact and usable in 3D.
First, it introduces a segmentation-based initialization that aligns experts with semantic image regions, reducing redundancy and greatly shortening training. Subsequently, the thesis replaces inefficient gradient optimization with a rasterized training schedule inspired by Gaussian splatting: images are split into blocks and only relevant cores are activated, significantly lowering the computational cost without losing precision. Finally, SMoE is generalized to 3D through reparameterized cores and differential rendering, enabling compact and high-quality scene build-ups even with sparse data.
Experimental results show clear improvements in speed and reconstruction quality compared to previous SMoE models and competing 3D methods. Taken together, these contributions make SMoE a practical framework for effective and high-fidelity visual representation and point towards broader principles for structure-aware, rasterization-friendly learning systems.
More information
Date: February 5, 2026
Time: 14:00
Location: Technical University, Berlin and via Zoom.
Doctoral thesis: "Steered Mixture of Experts for Compact and Edge-aware Representation: From 2D Image Regression to 3D Radiance Fields".
Defendant: Yi-Hsin Li.
Supervisor and chairman: Professor Mårten Sjöström, Mid Sweden University and Professor Thomas Sikora, Technical University, Berlin.
Review Committee:
Professor Thomas Sikora, Technical University, Berlin
Professor Ulf Assarsson, Chalmers University of Technology
Professor Peter Lambert, Technology Ghent University