A Fully-Unsupervised Generative Method for Choreo-Musical Translation
(2021)


Based on successful work in building cross-domain translation systems using Variational Autoencoders to translate between MNIST handwritten digits and spoken utterances of those digits, I attempted to build a model that could translate between quantized MIDI musical data and choreography (represented as quantized sequences of pose data).








Lightening the Load: DeLighT Blocks for Faster QA Training
and Ensembling
(2022)


One of the biggest risks to the future development of artifical intelligence is the consolidation of AI research within just a handful of institutions and corporations due to the ever-rising computational and data demands of our increasingly complex statistic models. That's why, in my attempt to build a competitive model for the SQuAD 2.0 question-answering task, I focused on lightweight alternatives to the popular transformer-based approach dominating much of the current NLP landscape.






Visual Price Estimation for Real Estate
(2021)


After building much of my AI experience in projects around audio and/or reinforcement learning, I was excited for the opportunity to branch out in my recent work. This paper represents one such exploration in a new direction, and probes the applicability of transfer learning, in combination with textual descriptors of real estate, to provide signals for visual price estimation.




Policy-GNN/Policy-GAT
(2021)


Graph neural networks (GNNs) were a fascinating new technology to dive deep on in my first quarter at Stanford. After discovering the application of reinforcement learning to the node preduction in the form of Policy-GNN, I attempted to further augment the model performance by replacing the GNN component with a state-of-the-art, attention-based, GAT network.




Impact Sound Neural Network
(2018)


My first research project enabled me to hit the ground running in applying deep learning to audio problems. Using a physical-modelling approach to sound synthesis for digital models, we attempted to build a system that could solve the inverse problem - identifying the source geometry from a sound of the object being struck.




An Inverse Reinforcement Learning Approach to Generative Music
(2020)


For my senior honors thesis at UNC Chapel Hill, I sought to bring together my experience with reinforcement learning from my time at Berkeley's AI and Robotics research group with my love for music. I was interested in exploring inverse RL approaches due to their high sample efficiency - enabling the development of an AI model trained on the sensibilities of an individual musician.