Yi-Hsin works as Early Stage Researcher at the Technische Universität Berlin (TUB), Germany, and her research topic is “Gated Experts Compression of Light Field Video”.
•Reconstruction of light field data by Gated Experts model with gradient descent optimization
•Incorporate SSIM measurement to train Gated Experts model
•Utilize different initialization algorithm for accelerating training process and improving the quality of the reconstructed results
You can read Yi-Hsin’s introduction here.
Introduction of the research:
My research topic is the compression of high-dimensional video data by Gated experts. The gated expert network is a regression framework that sparsely models data with a set of segments. It is designed and trained to directly explain correlations in high-dimensional data with edge awareness. The main problems in the current gated network are two-fold: the time for the optimization process and the quality of the reconstruction. To accelerate the optimization, the initialization is considered to be crucial so currently, I incorporate block-based SMoE (Steered Mixture of Experts) to verify the advantages and disadvantages of these initialization methods. The initialization is done in a block-based manner so that this process can be on the fly and does not affect the entire encoding time. Convergence speed increased because the initialization quality of block-based SMoE is high. In addition, the excellent initialization locations of the gates also avoid the local minima and improve the quality of the reconstruction.
For further information, you can refer to the slam video here: https://youtu.be/q0YMglUqzFs
Current goals and works:
Goals:
- Improve the reconstructed results
- Speed up the optimization process
- Reduce the amount of data required to be transmitted
Research works to achieve these goals:
- For improving reconstruction results and speeding up the optimization process
- Develop a proper initialization method
- Try different initialization methods
- For reducing the amount of data required to be transmitted
- Merge of gating networks
- Classification of gating networks based on the locations of the kernels and the values of the experts