Ming Qi
Me, Ming!
I'm a sophomore majoring in Computer Science and Physics at UPenn. I like playing soccer, going to the gym and playing video/board games with my friends. I would love to be a part of Google Deepmind, specifically working on something like Alphastar, and make agents to play my favorite games, like Civiliztion 6, Counter Strike and Coup.
Relevant Links
My Resume LinkedIn GithubProjects
Deep Reference Prior Implementation
Implemented deep reference prior model as outlined in Professor Chaudhari's paper linked below, which maximizes mutual information between weights and data for a set of atoms, allowing the neural networks to train on very small amounts of labeled data. Conducted semi-supervised learning on MNIST data using deep reference prior, achieving an 82% accuracy with just 3 labeled images per category.
Deep Reference Priors Paper Priors Github RepoAirline Delay Prediciton Project
Investigated whether bad weather could predict airline delays. Joined a Kaggle dataset on 2015 flights with a US county weather dataset to get the weather at the time of all flights. I preprocessed the data using one hot encoding for the airlines and standardized the condinuous features, then used GridSearch to hyperparameter tune a Random Forest XGBoosted Model. I also made a neural network, with a custom loss function that did binary cross entropy but allowed higher positive class weights to offeset the class imablance of the dataset. We found that weather was not a reliable predictor of delays, possibly due to airlines taking the necessary precautions in advance for bad weather.
Airline Delays Github Repo