PROJECTS
Relevant Projects and Papers
Playing 2v2 Ice Hockey In SuperTuxKart Using CNNs
Final team project for the UT Austin Deep Learning graduate course. Objective was to build an AI that would win a 2v2 hockey game against other AI agents designed by classmates. Gameplay was done in a modified racing simulator built on SuperTuxKart. Our AI algorithm utilized 2 CNNs that were trained to identify the location of the puck, the goals, and other players, then a control algorithm was fine-tuned to play the game based on the output of the CNNs. The resulting AI was able to consistently defeat other AI agents, as well as half of all agents designed by other student teams.
Yahoo Article Headlines Topic Modeling
Final paper for the UT Austin Case Study in Machine Learning graduate course. This study evaluated and compared the performance of 4 common topic modeling techniques, namely LDA, NMF, Top2Vec, and BERTopic, on a news title dataset and a news content dataset. The results demonstrated that BERTopic produced not only the highest topic coherence scores and topic diversity scores across the 2 datasets tested, but also produced coherent and interpretable topics to humans for both datasets. To validate BERTopic’s quantitative performance, qualitative analysis of the topics produced by BERTopic was also conducted using LDA as a point of comparison.
Improving Adversarial Evaluation Set Performance for QA
Final paper for the UT Austin Natural Language Processing graduate course. This paper uncovers common types of prediction errors via adversarial data sets and explores how changing the training data set can improve performance on adversarial evaluation sets. Experiments conducted show that training Electra-small on SQuAD 2.0 instead of SQuAD 1.1 significantly improved its EM and F1 scores on the Adversarial AddSent and AddOneSent evaluation sets, and training on SQuAD 2.0 then Adversarial QA yielded the highest EM and F1 scores on the Adversarial QA evaluation set.