Projects
Predicting Fall Risk from Smartphone Videos
Implemented multimodal neural networks to incorporate survey data and at-home sit-to-stand test videos to predict fall risk
Built effective intermediate motion capture CNN to predict joint angles and pose metrics with an OpenPose backbone
Posthoc analyses demonstrated no significant biases across gender, ethnicity
Results suggest the model can serve as a preliminary diagnostic to alert people if they are at an elevated risk of falling
Project completed for CS 231N (Deep Learning for Computer Vision) with Duncan Ross and Spencer Siegel
Automating English Language Proficiency Assessments
Utilized transformer models to create an autograder for English proficiency exams (e.g., TOEFL)
Implemented transformers for written, spoken proficiency (built on DistilBERT and Wav2Vec2, respectively)
Results suggest that transformer models have the potential to be a reliable and efficient alternative to human graders for English proficiency assessments
GitHub Repository: https://github.com/ethan-allavarpu/nlp-toefl-autograder
Project completed for CS 224N (Natural Language Processing with Deep Learning) with Duncan Ross and Spencer Siegel
Predicting the Severity of PG&E Planned Power Shutoffs
Predicted PG&E shutoff length (minutes) for PSPS events based on weather, demographic, and circuit (substation) information by utilizing various regression models, including random forests, boosting, neural networks, and regularization
Posthoc analysis of high prediction errors for the best model (XGBoost) to determine important and volatile features/locations for PG&E power restoration
GitHub Repository: https://github.com/ethan-allavarpu/cs-229-pge-final-project
Project completed for CS 229 (Machine Learning) with Joe Jamison and Umar Maniku
DataFest 2021: Analyzing Drug Misuse Problems
Analyzed drug misuse in the United States in 2018 to determine the drugs most often misused and the demographics associated with misuse
Performed chi-square tests to determine which demographics were linked to higher rates of misuse
Used generalized linear models and random forest (with ANOVA) to see which variables played a significant role in explaining misuse
DataFest 2021 (48-hour hackathon hosted by UCLA) completed with Dara Tan