Projects

For the CS182 course at UC Berkeley, we explore Transformer models’ In-Context Learning robustness against noisy labels and the effects of noise as data augmentation.

FMLens is an interactive visual analytics system that scaffolds the fund manager selection process to support investors in efficiently evaluating candidate fund managers and unfolding the changes in their investment styles.

VAC-HGNN (Visual Analytics for Comparing HGNNs) is a visual analytics system that enables HGNN practitioners to understand and compare HGNNs at three levels: graph level, group level, and individual level.

ALens is a domain-oriented abstract writing training system that uses rhetorical structure parsing to identify main ideas, evaluates abstracts from different linguistic features and uses visualization to analyze the writing patterns of reference abstracts.

We employed agile development to add features that allows users to learn more about their representatives, political events in their area as well as aggregate, share and view news items in their locality to Actionmap.

The system employs visual analysis solutions to investigate the network assets and correlations of black and gray industry gangs, such as domain names and IP addresses, to identify suspicious activities, understand their operational mechanisms, and develop strategies to combat their detrimental impact on network ecology and social security.

Collected a hand gesture recognition dataset, built DD-Net from research and compress the model with knowledge distillation.

Mitigated data imbalance in tweet-news linkage by utilizing ChatGPT for text augmentation and use Sentence-BERT-based model to link tweet and news.

Use RISC-V to implement the Chrome Dinosaur Game on Sipeed Longan Nano development board.

Added temporal information to static graph representation by GRU and used DQN to discover meta-paths.

CodeCognoscenti is designed to offer developers efficient and non- intrusive ways to interact with LLM when understanding generated code.