Skills

Programming Languages

Since my first exposure to coding in high-school, I have been always enchanted by the beauty of communicating with computers through programming languages. I used to learn languages as archaic as dinosaur (Assembly, Pascal, VHDL, Verilog). Early in my career, I mostly used Matlab and C++ to develop computer vision software and perform computer-controlled behavioral experiments. More recently, I mainly code with Python, R, SQL and Spark for data science and machine learning applications.

Data Analytics

For most of my career, I have been working extensively on drawing insights from complex data using advanced analytics tools. I'm proficient in manipulating data of all kinds (structured and unstructured) and making insightful visualization using data analytics tools (spark, pandas, scipy, numpy, matplotlib, plotly, seaborn in Python and tidyverse, ggplot in R). You can check out my COVID-19 visualization project. I also have experience wrangling large amount of data (up to 20 TB) on cloud platforms using tools like Databricks, AWS Athena, Glue, SQL, Linux. As a trained scientist, I also have considerable experience in performing advanced statistical analyses and hypothesis testing (like A/B testing ) to draw actionable and deep insights from complex data (you can check out this postdoctoral project).

Machine Learning & Artificial Intelligence

I have extensive experience building machine learning models to make prediction and gain insights from data. During my PhD at UPenn, I performed research with traditional Machine Learning methods: Bayesian modeling, Linear/Logistic regression, LDA, SVM, Tree-based models, PCA, k-means, kNN, Gaussian mixture, tSNE. Later on when I started working at Columbia University and AstraZeneca, I transitioned to the more recent neural networks models and worked on state-of-the-art architectures like CNN, RNN-LSTM, TCN, and GAN. I mostly use popular Python packages to train machine learning models (scikit-learn, tensorflow, keras, pytorch, transformers).

Natural Language Processing

My first exposure to NLP was a project in which I predicted a user's emotion (happy/sad) from their Twitter posts using bag-of-words approach and traditional machine learning methods like Naive Bayes and SVM. Later on, I learned more about feature extraction and preprocessing in NLP (stemming, lemmatization, TF-IDF, tokenization ) and applied them along with machine learning and tSNE method to gain insight from job descriptions. More recently, I'm into the world of LLM and Transformer and have done projects to apply them to clinical texts for several important tasks such as clinical entity recognition and question-answering. For NLP, I mostly use popular Python libraries like nltk, spacy, transformers.

Computer Vision & Medical Imaging

Early in my career, my research focused on traditional computer vision methods. I worked on a project to develop computer vision software that can accurately track changes in objects' properties over time. In a similar vein, I also developed and validated medical imaging softwares to measure blood flow and blood oxygen which are strong diagnostic measures of several diseases (e.g. diabetes as illustrated in my works). Later on when I focused more on human vision rather than computer vision, I had an interesting project to compare those vision systems. I also did a project to detect cancer in medical scan images using various neural network models (CNN, ResNet, Inception) and transfer learning technique. Another intersting project I did was to employ styleGAN (a generative neural network) to generate synthetic images and FaceNet to analyze the synthetic images for behavioral experiments.

Time serise analysis

During my time at AstraZeneca, I spent a lot of time extracting actionable insights from sensor data (e.g. accelerometer, acoustic signals). Several of my projects involve deriving physical activity measures (e.g. step count, walk distance) from wearable devices using both signal processing algorithms and state-of-the-art neural networks (CNN, LSTM, WaveNet) with a focus on clinical use. I also worked on developing, validating and deploying machine learning algorithms (XGBoost, YamNet) to detect cough from acoustics signals.

Web Development

I mostly work from the backend side to develop API that a front-end App or website can query results of machine learning models' prediction (using, for example, Python fastAPI). That said, I also enjoy the front-end party by making beautiful websites. In fact, I designed and created the site you're reading right now from scratch using tools like HTML, CSS, Bootstrap, JavaScript (code is here if you're interested).

Software Development

I'm a strong advocate of best practices in software development such as Agile and DevOps principles. As a way of working, I adopt tools such as Jira, Kanban, Srum, Sprint into my daily workflow. From the coding perspective, I use Git (Bitbucket, GitHub) for version control, VSCode for IDE, pytest for unit testing, Jenkins, Github Action for CI/CD, Docker, Kubernetes for deployment, and popular cloud tools like AWS (SageMaker, Athena, Glue, S3), Slurm, GCP.

Projects

LLM applications on clinical texts

With the ubiquitous adoption of electronic health records, 80% of medical data now exist in the form of texts. As a results, there is a huge potential for AI systems to extract useful insights from that enormous data. In this project, I applied the state-of-the-art LLMs for various popular tasks such as clinical entity recognition and question-answering to provide insights into clinical texts. You can aslo try out the NER and QA models I put on Huggingface.

Accurate Step Count with Deep Learning

Physical activity (PA) is widely recognized as a pillar of general health. Step count, as one measure of PA, is a well known predictor of long-term morbidity and mortality. However, the step count accuracy of popular consumer devices like Fitbit or Apple watch is not rigorously validated, esp. for clinical use. Here by using state-of-the-art neural network algorithms and transfer learning, I can achieve highly accurate step counts (98-99%) from many types of wearable and mobile devices. If you are interested, please find more details here.

Cancer Detection with Deep Learning

In this project, I tackle the CAMELYON16 challenge in which we have to detect cancer in hyper-resolution scans of human tissues (up to 100,000 x 100,000 pixels). To solve the problems, I experimented with several state-of-the-art neural network models including InceptionV3 and InceptionResnetV2. I show that we can obtain very high performance when performing transfer learning with a pretrained InceptionV3 network. If you want to learn more, feel free to read the report here.

Understand Human Decision-making

This is essentially my PhD and Postdoctoral works in which I study how humans make decisions using a combination of computer-controlled behavioral experiments and machine learning (with the focus on Bayesian models). This work reveals a unifying principle behind many of human irrational decisions: maintaining consistency across decisions. You can read the technical paper or non-technical article and play around with Matlab code. Some of the following works are here, here, and here.

Manipulate Human Face Images with styleGAN

In this project, I explored a powerful neural network algorithms, styleGAN, which can generate synthetic human face images that look like real (you can look at images below). Additionally, styleGAN can also change specific features of a face image (e.g. skin tone, hair color, identity). Besides demonstrating these impressive features of styleGAN, I also dived deep into the way it works using PCA technique and an advanced deep learning model, FaceNet. You can find all the details here.

COVID-19 Visualization and Correlational Analyses

In this project I pulled the COVID-19 data from Johns Hopkins University database which include measures such as number of cases and deaths across the world and in the US. Then I made several visualization and combined economic and health data from other sources such as World Bank, WHO and US goverment so that I could gain insights into the evolution of COVID-19 and its relationships to other social and economic factors like GDP, income, health spending. All details are here.

Track Object with Digital Image Correlation

In this work, I used a popular computer vision algorithm, Digital Image Correlation, to track an object's properties over time. Essentially, we captured images of objects in different stages and used image processing algorithms to compare those images. I demonstrate that by using an advanced interpolation algorithms, we can potentially increase the accuracy of the algorithm up to 100x. This algorithm has found its way into a broad range of applications such as civil, mechanical and biomedical engineering. You can read more about that in this paper.

Measure Retinal Blood Oxygen with Optical Imaging

This is a project in collaboration with Johns Hopkins University in which we want to track the progression of diabetes in a non-invasive way. Specifically, we developed an optical imaging system to capture images of patients' retina and employed advanced image processing algorithms to compute the blood oxygen from retinal vessels. In addition, I also developed a friendly Graphical User Interface so that the doctors can easily use the software for their purposes. You can find more details in this paper.

Validate Blood Flow Measurement Method

In this project, I worked on developing and validating optical imaging methods that can non-invasively measure the blood flow in human blood vessels. One challenge to the problem is to fabricate models of human blood vessels for testing the imaging methods. Here I used an advanced laser micromachining system to fabricate models of very small blood vessels (as small as the smallest human hair) and used that to test an optical system that measures blood flow. You can read more here.

Compare Computer Vision and Human Vision in Scene Classification

In this project, I compared the performance of a scene classification computer vision model that is based on low-frequency components of amplitude spectrum (Oliva & Torralba, 2001) and human fast perception of images. I found that human perception is very different from the model, that is, their performance relies more heavily on the phase (local) information in the image. In contrast, the gist-based computer vision model relies mainly on the amplitude spectrum (global features) of the image.

NLP Twitter Post and Job Description Analyses

In these projects, I performed several NLP methods including machine learning to gain insight from text data. In the first one, I aimed to predict a user's emotion (happy/sad) from their Twitter posts. I also performed visualization based on Naive Bayes algorithm to have a closer look into the way the model works. In the second one, I performed analysis and visualization of thousands of job descriptions to gain insights into the relationship between jobs.