Many failures are not due to limited capacity, but because they do not adhere to the end -- No Game No Life
-
Oregon State University, Corvallis, OR
-
Master of Science in Statistics
-
Oregon State University, Corvallis, OR, June 2022
-
Master of Science in Mathematics
-
Oregon State University, Corvallis, OR, June 2017
-
Bachelor of Science in Mathematics
Online Classes
- Computer Science and Programming Using Python at MIT
- Computer Science 101 at Stanford University
- Machine Learning at Coursera
Research Experience
- Denoise Speech Using Deep Learning Networks and Wavelet, Advisor: Adel Faridani, Ph.D. Department of Mathematics, Oregon State University
- Built Encoder-Decoder neural network models by using Convolutional layers for audios denoising
- Applied wavelet for deconstructing and reconstructing noised audios
- Implemented Mallat's algorithm to construct Daubechies wavelets
- Sentiment Polarity Identification in Financial News (on-going)
- Applied Beautiful-Soup to web-scrap daily financial news from major news websites
- Built a robot to collect Twitter daily top news
- Sentiment Analysis
- Potential Statistical Method
Kaggle Data Science Competition
- Avito Demand Prediction Challenge
- Top 2% out of 1873 teams
- Applied pre-trained Convolutional neural network models to extract features from image data
- In this competition, we used a single layer stacking of different Neural Network and LGN models, and then trained our stack with XGB and blended the results with LGB stacker 1:1. Since there are not many word embeddings for Russian texts, we applied TF-IDf for feature engineering and then used PCA to reduce the dimention from 2000 to 200.
- Toxic Comment Classification Challenge
- Top 2% out of 4551 tems
- Challenge: Don't have computers with good GPUs and enough RAMs for training.
- Since this was our first Natural Language problem, we stacked and blended our Neural Network models (over 20 models) to achieve our final result. Our best single model was a RCNN model that bases on bi-gru-cnn-poolings with some tunnings.
- Mercari Price Suggestion Challenge
- Top 8% out of 2380 teams
- We applied tf-idf from sklearn to extract features from texts and then built a ridge model with all original numerical features and tf-idf features for predicting the correct item price from an ad
- Quora Insincere Questions Classification
- Top 20% out of 1401 teams
- This was a kernel (13Gb RAMs and a Tesla 80k GPU) only competition.
- Topic modeling, a small projects for studying NLP
- Final model was a RNN model with poolings and attention. Because kernel only runs up to 2 hours, we spent a lot of time on improving training time. Our solution for that was inserting a Capsule layer between GRU layers that groups outputs into k capsules. In order to get better results, we applied stochastic weighted average to average weights after 3 epochs.
Web Development
- Bitcoin Price
- This project is a Time-Series study (because coindesk only provides close prices data for free, we just do a simple pattern analysis). D3 is the main module to create each line chart. In order to zoom in specific time range, we modified a function from D3 gallery. The major obstacle of this project was replacing new charts with new json files. If we simply update a json file and then create a new chart, previous one will maintain on the background; therefore, we removed all svg structures of previous chart and then added new svg structures to generate a new one.
-
Graduate Teaching Assistant at Oregon State University - (Sept. 2019 — Present)
-
Teaching assigned College level Statistics/Mathematics courses
-
Assessing and correcting students’ programming code
-
Leading class discussions and answering student questions
-
Evaluating students' written answers on assignments