Research Projects

Category Prediction in OGBN-Products Dataset — PyTorch Geometric, Graph Convolutional Networks Implemented a graph convolutional network (GCN) to predict the category of a product in Amazon’s product co-purchasing network. Trained on a subgraph with 100,000 nodes and achieved an accuracy of 91%.

REINFORCE Algorithm for Continuous Cart Pole Environment — Reinforcement Learning, Gym, PyTorch Deployed the REINFORCE algorithm with target and policy networks such that the agent learned to strategically accelerate and reach a goal state on a continuous sinusoidal hill. The agent solved the task every five attempts.

Deep Q-Learning for Frozen Lake Environment — Reinforcement Learning, Gym, PyTorch Trained an agent to reach the goal state in the maximal, 8x8 slippery (stochastic) environment using deep Q-learning. The agent learned to complete the task with 100% accuracy after approximately 200 iterations.

Generative Adversarial Network for Face Generation — Generative Adversarial Networks, PyTorch Developed a generative adversarial network (GAN) that generated new faces after training on a dataset of characters from The Simpsons.

Autoencoders for CIFAR10 Recoloring — Autoencoders, U-Nets, Convolutional Neural Networks, PyTorch Implemented an autoencoder and U-net that recolored grayscale-converted images from the CIFAR10 dataset. Minimized average loss to approximately 0.01.

Multi-Modal Digital Assistant — Computer Vision, Natural Language Processing, JavaFX Collaborated with 5 teammates to create a multi-modal digital assistant that launched the GUI and addressed the user upon face detection and recognition. Implemented the program’s response to text inputs through CFG, CYK, NER, LSTM, and BERT. Extended the application with Google and OpenAI API, and increased usability with hand gesture commands.

3D Bounded Knapsack Problem with Polyominoes — Knapsack Optimization, Genetic Algorithm, JavaFX Coordinated with 6 teammates to develop a recursive algorithm that provided exact cover solutions to a 2D grid, given the dimensions and available pieces as input. Adapted the program to solve knapsack problems in a bounded 3D space visualized with JavaFX. Extended the application to a Pentris game and trained a bot that outperformed human players using genetic algorithms.