User details
SZ
SZ
Shuying ZHU
@shuying_zhu_o5jgr6l1m
Abbess
Mount Houston, TX, United States
1Follower0Following0Mutual0Post
Member since: Jan 20, 2025Machine Learning Researcher with extensive experience with designing and building scalable deep learning models for large-scale data on cloud platforms. Equipped with 5 years of ML research and engineering experience (including LLMs). Hands on experience with RAG, Chatbot and AI Agents. Seeking to leverage machine learning and engineering skills to broader business impact.
Objective: To develop a foundation model for time-series data which can be used for various tasks such as risky driving behavior detection.
- Enhanced risky driving behavior detection accuracy by 5% by designing and pretraining a novel self-supervised transformer model for time-series data, inspired by BERT, ViT and contrastive learning in Python and PyTorch.
- Speeded up data processing by 20% with data pipeline development using Spark and SQL on AWS cloud and Databricks.
Reduced user complaints of being mislabeled as risky drivers by 5% by fine-tuning the pretrained model on a small amount of data.
- Cut training time by 50% and decreased cloud costs by 70% through infrastructure optimization with Ray distributed computing and Docker on Kubernetes. Enabled efficient scaling and resource management across multiple GPUs, maximizing computational efficiency.
Objective: To predict the pathogenic effect of single-nucleotide variants on protein-protein interaction.
- Achieved a 40x boost in processing speed by curating and preprocessing over 10 million XML and JSON clinical data entries from the
- ClinVar database using Python, Pandas, Numpy, and parallel computing tools like Dask and Hadoop.
- Improved model prediction accuracy by 10%, reaching 95% AUROC, through extensive experimentation with large language models
(LLMs) such as ESM2-15B and graph neural networks (GAT, GCN, GIN) with PyTorch. (https://github.com/zshuyinggg/ppi_mutation)
Objective: To conduct timely COVID-19 research.
- Analyzed age patterns of COVID-19 cases, adapting daily to evolving case data. Applied hypothesis testing (the Wilcoxon rank-sum test) using R to compare age distributions and age-specific incidence rate between the Hong Kong and Singapore populations, further analyzing the data with a Chi-square test to determine age-specific incidence rates.
- Published two peer-reviewed papers within three days timeframe. (https://link.springer.com/article/10.1186/s13104-020-05178-z) (https://www.sciencedirect.com/science/article/pii/S1201971220321901)
Objective: To propose and implemente a novel contrastive learning model for electronic health records for clinical predictions.
- Identified high-risk women with 5% higher accuracy with the deep learning model than Logistic Regression, XGBoost and Random
Forests using Python and SQL.
- Proposed a novel deep learning framework with PyTorch which successfully predicted in-hospital mortality and length-of-hospital-stay with 10% higher accuracy than existing models, incorporating Autoencoder, attention mechanism and contrastive learning with LSTM and CNN as backbones. Used MySQL to extract 17 features within 48 hours after the admission to ICU from EHR data MIMIC-III.
- Models were published on a high-impact journal, with the potential to support hospitals prioritizing high-risk patients when limited
resources are available.
(https://github.com/zshuyinggg/CPAE)(https://www.sciencedirect.com/science/article/pii/S0169260723001505)