About me

I have recently finished my master’s studies at McGill University and Quebec Artificial Intelligence Institute (MILA), under the supervision of Prof. Doina Precup. My master’s thesis was about reinforcement learning and optimization.

My projects:

Master’s Thesis

Machine Learning Projects

Software projects

Master’s thesis work

SVRG for Policy Evaluation with Few Gradient Evaluations

  1. Worked on developing sample-efficient first order optimization algorithms for policy evaluation in reinforcement learning.
  2. Proved linear convergences of proposed methods.
  3. Conducted experiments and showed proposed methods are more sample-efficient than previous methods, especially in large-data settings.
  4. Links to code and thesis: here
  5. Link to paper on Arxiv: here

Machine learning projects

Image classification on hand drawing data

  1. Designed a conditional random field (CRF) for the task on a small testing dataset. Analyzed why the CRF reaches its optimal performance. Used a residual network to accurately classify images from 340 categories.
  2. Results: here
  3. Language used: Python

Reproducing inverse reward design

  1. Reproduced a Neurips paper, titled “Inverse Reward Design”. Analyzed strengths and weaknesses of the paper. Showed omitted details of the paper in our report.
  2. Report: here
  3. Language used: Python

Image Segmentation

  1. Used Gaussian mixture model and EM algorithm for image segmentation.
  2. Results: here
  3. Language used: Matlab

Aggregators in Markov logic networks

  1. Analyzed how the quality of aggregators relates to prediction accuracies. Developed a Markov logic network that outperforms other baseline methods.
  2. Results: here
  3. Language used: Python

Control of dynamic systems using neural networks

  1. We combined LSTM and DNN, and used it to learn the complex behaviour of model predictive control.
  2. Code: here
  3. Report: here
  4. Language used: Python

Software projects

Course Scheduler

  1. Used a flow network to solve the scheduling problem. Used promises instead of call backs to handle asynchronous operations. Wrote tests that cover most of the code base.
  2. Code: here
  3. Language used: Typescript

Let’s Sell

  1. This mobile application allows users to sell text books and upass (transist pass for students). Users can add items for sales, search items, save items to their watch lists, and log in by authenticating their social network accounts. This app was on the ios app store for a year.
  2. Code: here
  3. Language used: Objective C