Using reinforcement learning methods to teach an agent to play the game of Limit Hold-Em Poker using the rl-card environment. Wrote Deep Q-Learning agent and Neural Fictitious Self Play agent and formulated recurrent variations of both learning algorithms to induce the modelling bias of sequential strategy in poker gameplay.

Feb - May 2020

Implementation of a self-sufficient appeals system for Amazon Mechanical Turk (AMT). In particular, my partner and I designed a web application through which AMT workers are able to appeal their rejected tasks to be adjudicated by other workers on the AMT platform. A game theoretic analysis of the truthfulness properties of the Mechanism was also conducted after formulating as a peer prediction problem.

Mar 2019 - current

Fine-tuning pre-trained BERT models for performance on the SQuAD dataset using HuggingFace transformers library. Experimenting with BERT, RoBERTa, DistilBERT, DistilRoBERTa to anaylze how different pre-training approaches and distillation of a transformer model affect inference time and performance on the context Q/A task in particular.

Apr - May 2020

Convolutional neural network models to recognize chess gamestate from a photo of the board. Our final model yielded 99.28% accuracy on the testing set. Additionally, we created a hand-labeled dataset of 500 chess boards with various types of noise (lighting, angles, background).

Mar - May 2019

Experimenting with autoencoder-based methods to reconstruct cell lineage tree from cell mutation data. In particular, we use autoencoder models to construct a 16-dimensional latent space for 200 length DNA strands and use various reconstructions methods such as K-means and UPGMA to trace back cell mutation lineages.

Oct - Dec 2019