Embedded Files
Welcome! I am a Ph.D. student in the Computer Science and Engineering department at University of California, San Diego, co-advised by Prof. Rose Yu and Prof. Yian Ma. I enjoy learning and building stuff with my growing knowledge. My research interests are Large Language Models, Uncertainty Quantification, Spatiotemporal Modeling and AI for Climate Science.
Research Experiences
Functional-level Uncertainty Quantification for Calibrated Fine-tuning on LLMs
@ Rose Spatiotemporal ML (STL) Lab, UC San Diego
Ruijia Niu, Dongxia Wu, Rose Yu, Yi-An Ma
• Fine-tuned LLMs often exhibit overconfidence due to sparse training data .
• We model epistemic uncertainty with prompt-dependent functional variability via LoRA-MoE.
• Our proposed method reduces Expected Calibration Error by over 25% across five benchmarks.
Multi-Fidelity Residual Neural Processes for Scalable Surrogate Modeling @ Rose Spatiotemporal ML (STL) Lab, UC San Diego
Ruijia Niu, Dongxia Wu, Kai Kim, Yi-An Ma, Duncan Watson-Parris, Rose Yu
• Novel multi-fidelity surrogate modeling architecture to address inaccurate latent representations.
• Model improves the average performance by ~90% from state of art methods for PDE modeling.
• Achieves state of art performance on real world climate observation surrogate modeling (ERA5).
Deep Bayesian Active Learning For Accelerating Stochastic Simulation @Rose STL Lab, UC San Diego
Dongxia Wu, Ruijia Niu, Matteo Chinazzi, Alessandro Vespignani, Yian Ma, Rose Yu
• Novel Spatiotemporal Neural Process Model Structure and acquisition function for active learning.
• Designed and added two new experiments, wrote detailed performance comparisons and analytics.
• Read and analyzed 4 top-notch journal articles with a focus in mathematics.
• Active Learning model performance reached offline (using 100% data) with only 21% of data.
Disentangled Multi-Fidelity Deep Bayesian Active Learning @Rose STL Lab, UC San Diego
Dongxia Wu, Ruijia Niu, Matteo Chinazzi, Yian Ma, Rose Yu
• Proposed local and global information aggregation for each fidelity level.
• Implemented Gaussian Process baselines and pool based active learning.
• Conducted robustness test by separating data at each fidelity level and test independently.
• Model implemented with PyTorch, trained on Linux Servers.
Improving Transferability of Adversarial Patches on Facial Recognition with Generative Models in Physical Space @Institute of Automation, Chinese Academy of Science, Beijing, China
• Physical black-box attack on facial recognition models via deep convolutional generative adversarial networks.
• Read and analyzed 28 top-notch (CVPR 2021) journal articles, combined methods in two of them.
• Optimized the model using PyTorch, Tensorflow, Python, Jupyter Notebook on Linux & Windows platforms.
• Increased model attack success rate by more than 45%.
Work Experiences
ML/Analytics Cloud Engineer Intern @ SuccessKPI, Inc.
• Worked on real-time call transcribing on a serverless SaaS platform with AWS.
• Utilized Docker, Bitbucket, AWS ECS/Lambda Functions/DynamoDB/Transcribe/X-Ray/ECR/etc.
• Implemented X-Ray to monitor ECS activities/logs with CloudWatch dashboard for error reporting.
• Significantly enhanced overall transcribe accuracy with partial results stabilization.
Computer Science & Cognitive Science Tutor @ UC San Diego
• Provided one-on-one and group tutoring for over 60 students on research methods, Java, data structures.
• Collaborated with professors to identify educational goals and initiated strategies to keep plans on track.
• Received excellent rating (10/10) in instructor feedback.
Resume
Brooks_Resume.pdfContact Me: rniu@ucsd.edu
Google Sites
Report abuse