UROP Proceedings 2020-21

School of Engineering Department of Computer Science and Engineering 142 Efficient Algorithms to Process Gigapixel Images Supervisor: SANDER Pedro / CSE Student: XU Peiyu / COMP Course: UROP3100, Fall We present a method for rendering acceleration by exploiting the temporal coherency between frames. This project is an optimization of existing algorithm that caches certain information from previous frames, and project them when rendering the next frame. In its simplest form, a regular refresh of the cache is required, which is often expensive. Such refresh cost shall be amortized in application. We managed to propose an algorithm for intelligently refreshing the cache, taking information for each object into consideration. Commonsense Reasoning with Knowledge Graphs Supervisor: SONG Yangqiu / CSE Student: XIANG Tianqi / COSC Course: UROP2100, Fall Presently reported is a progress report of a UROP research trying to accomplish the task 4 of SemEval 2021, a competition in natural language processing area. The project is going to firstly prepare some knowledge of this area. Then it is going to test and evaluate the provided baseline models. Finally, it is going to improve and implement some other models to deal with the tasks and evaluate the effectiveness of the new models. This progress report will show the principle of the operations, the procedure of this research, the progress of this project and the future plan. Commonsense Reasoning with Knowledge Graphs Supervisor: SONG Yangqiu / CSE Student: CHOI Sehyun / COMP Course: UROP1100, Spring Commonsense Knowledge Bases (CSKB) provides a venue to let the machine to understand commonsense knowledge, which are often omitted in human generated text. Such knowledge bases are costly to build in that requires extensive human annotations, and therefore attracted a research direction of automating the process. However, reasoning over such knowledge bases is a difficult task. Previously, CSKB completion aims to introduce missing edges between known entities, whereas CSKB population proposes to introduce both new entities and links between them from other sources. An effective high-quality dataset for the task is proposed, along with novel graph reasoning models that outperformed baselines from previous works.