Selected Projects

======

• Towards Fair Language Modeling via Parameter-Efficient Methods by Machine Feedback 2024
  • Mitigation of social biases in large language models (i.e., T5, BERT, LLaMA 2) based toxicity detection and hate speech detection
  • Train LLM to learn fairness and mitigate bias using reinforcement learning (RL) and parameter-efficient tuning methods (i.e., LoRA, P-tuning)
• MEGAWATT: MAST for Evaluating Generative AI in Worker-Automation Team Tasks 2024
  • Apply MAST (AI trust assessment tool) to evaluate baseline performance, inform improvement, and appropriate adoption of OpenAI’s GPT-4, to assist in intelligence and analysis (I&A) type tasks
  • Conduct human subject studies to assess whether off-the-shelf or improved outputs can lead to appropriate use, including correct rejections of model outputs
  • Improve quality of GPT-4 responses with prompt engineering and retrieval-augmented generation (RAG) for general conversation and various NLP tasks (e.g., text summarization, entity recognition)
• Automated Evaluation of Machine-generated Summaries using RLHF 2024
  • Trained a LLM-based classifier to evaluate a document-summary pair through multi-class classification and reinforcement learning with handcrafted human preferences dataset
  • Conducted expert evaluations on the output scores to validate the effectiveness of the proposed learning method
• PADTHAI-MM: A Principled Approach for Designing Trustworthy, Human-centered AI systems using the MAST Methodology 2023
  • Developed a novel AI design framework, addressing the challenge of designing trustworthy AI systems
  • Demonstrated the effectiveness of the framework through the development of the AI-enabled decision support system, with the framework positively impacting trust perceptions among users
  • Conducted association analysis between participants’ ratings and trust-impacting information, providing a theoretical basis for the framework’s effectiveness in enhancing AI system trustworthiness
• READIT: REporting Assistant for Defense and Intelligence Tasks 2022
  • Trained and developed a text summarization system for use in intelligence analysis, utilizing Transformer-based models
  • Implemented a user-friendly web interface for the text summarization system using NodeJS and the Google Cloud Platform, allowing analysts to easily access summarized reports, enhancing their workflow and productivity
• Facewise: An AI-based Face ID Verification System 2022
  • Engineered a robust and accurate face ID verification system, ensuring a reliable and efficient means of identity authentication in security screening scenarios
  • Implemented face matching algorithms with Convolutional Neural Networks (CNN) and ResNet and fine-tuned model parameters to optimize the system’s performance, thus enhancing the overall security and user experience
• Interpreting Text Classifiers with Counterfactual Explanation 2021
  • Completed as the final project for CSE 472 (Social Media Mining)
  • Implemented counterfactual models for a multi-layer neural network used in text classification
• Biomedical Entity Relation Extraction 2017
  • Extracted Biomedical entities and identify their relation existence
  • Utilized the Comparative Toxicogenomics Database (CTD) dataset, which provides chemical-gene, chemical-disease, and gene-disease relation data collections through distant supervision due to the lack of training data
  • Implemented and trained a tree-RNN based model, SPINN, in conjunction with a word-character embedding model