Portfolio

Deep Learning

• Developed a library for deep learning-based visual similarity search, clustering, and image embeddings [Python, PyTorch,fastai, Flask]:
  • While building the library, contributed to several open-source packages (including fastai, a popular open-source deep learning library);
  • To showcase possible use cases of the library built a Flask application with Docker and deployed it on AWS;
  • Ensured reliability of the library and the project by implementing a continuous integration pipeline in CircleCI using pytest. Also implemented end-to-end browser tests using Selenium;
  • Ported open-source RasterFairy library to python3, packaged it and pushed to PyPI.
• Created an eye-tracking system [Python, fastai, Flask]:
  • Built a web app for data collection and labeling, collected data;
  • Trained a deep learning model that predicts which part of laptop screen a user is looking at;
  • Accuracy less than 2 cm;
  • My team productized this technology at a Beijing Startup Weekend in May 2019 and got third place.
• Built a leafy greens classifier [Python, fastai, Docker]:
  • Scraped more than 8000 images across 29 classes of leafy greens;
  • Built a classifier to tell them apart.

Traditional Machine Learning

• Made a predictive model for breast cancer diagnostics [Python, scikit-learn]:
  • Performed feature analysis and engineering;
  • Finetuned and validated a variety of models;
  • Best model achieved 98.1% accuracy, with more than 96% TPR and more than 99% TNR.
• Performed sentiment analysis of Twitter feed [Python, R]:
  • Analyzed 200 gigabytes of Twitter feed data to extract sentiments;
  • Built an index that was predictive of stock market movements.

Consulting Projects

• Created a proof of concept of deep learning in industrial defect detection:
  • Put together a data collection pipeline (industrial cameras and lighting setup), collected and labeled data;
  • Trained a convolutional neural network with performance on par with QA employees (97% accuracy).
• Built an automated sales pipeline proof of concept for a client:
  • Performed data cleaning and analysis for auto dealers data all across China, with millions of data points;
  • Estimated customer return dates.
• Delivered two AI product feasibility analyses:
  • Researched and prepared papers summaries, compiled state of the art results on the topics of 3D pose estimation with deep learning and emotion recognition in voice data using deep learning;
  • As a result, one of the clients proceeded to develop a solution for voice emotion recognition.

Writing

• Why Swift may be the next big thing in deep learning - an overview of advantages of Swift in deep learning applications (also published on Medium).
• CapsNet Series in four parts (1, 2, 3, 4) — a popular introduction of the novel neural network architecture. This series was originally published on Medium (1, 2, 3, 4), where it got a lot of attention.
• Stochastic Weight Averaging — a new way to get state of the art results in deep learning.
• Deep Learning Vision for Non-Vision Tasks — three case studies about creative application of deep learning vision models to non-vision tasks
• Key Deep Learning Architectures for Visual Object Recognition — a series of posts about the most important CNN architectures.

Autonomous Driving

Below are the projects I did while taking Udacity’s Self-Driving Car Nanodegree:

• Lane detection using computer vision techniques [Python, OpenCV]
• Traffic sign classification using deep learning [Python, Tensorflow, OpenCV]
• Behavioral cloning using deep learning [Python, Keras, OpenCV]
• Advanced lane detection with camera calibration. In addition road curvature and vehicle offset relative to the center of line are calculated [Python, OpenCV]
• Computer vision-based real-time vehicle detection system [Python, scikit-learn, OpenCV]