Half‑time seminar with Yi‑Hsin Li

Thu 13 Jun 10.00–12.00
L401 Campus Sundsvall, Technical University Berlin and Zoom

Welcome to half-time seminar where PhD-student Yi-Hsin Li will present her work on The Steered Mixture of Expert for High-Dimensional Data Compression.

Yi-Hsin Li

Title: The Steered Mixture of Expert for High-Dimensional Data Compression

Respondent: Yi-Hsin Li

Opponent: Prof. Peter Lambert (Ghent University, Belgium)

Supervisors: Prof Mårten Sjöström (Mid Sweden University), Prof Thomas Sikora (Technical University Berlin, Germany), Prof Sebastian Knorr (HTW Berlin, Germany)


Mid Sweden University, Campus Sundsvall: L401

Technical University Berlin: TBA

Zoom (Zoom link)

Time: 10.00-12.00


The relevance of data compression in society underscores the importance of efficient compression techniques across various domains, including communication, entertainment, healthcare, and beyond. Steered Mixture of Experts (SMoE) offers unique advantages in data compression by simultaneously optimizing edge-awareness and smooth reconstruction, mitigating the trade-off often encountered in previous methods. However, SMoE's joint optimization of the entire number of components escalates computational resource requirements, particularly as the number of components increases. To address this limitation, the research focuses on segmentation-based initialization as a means to allocate kernels efficiently. By automating the initialization process and estimating the optimal number of kernels based on input data characteristics, the study aims to enhance compression gains, optimize computational resources, and improve overall efficiency. The hypothesis suggests that automated initialization can ensure effective resource utilization and enhance network adaptability, leading to improved reconstruction quality and scalability across various datasets and applications. The future study further investigates the application of segmentation-based initialization for enhancing the performance of SMoE models in high-dimensional data compression, and its scalability on various shallow networks, proposing to extend the application of segmentation-based initialization.

The page was updated 5/28/2024