Multiview 3D-VAE-GAN

CMPSCI 674: Intelligent Visual Computing final project

A popular method in 3D object classification is to use multiple 2D images (views) of the 3D object and pass through independent CNN networks. Recent research in generative 3D deep learning includes generating 3D voxelized objects using the VAE-GAN architecture (3D-VAE-GAN). In this project, I extend the existing 3D-VAE-GAN to include multiple views of 2D images for generating voxel based 3D objects. I found that my method had a lower average reconstruction loss on the test dataset.
[report] [code] [slides]

Learning to drive with RL and VAEs

CMPSCI 682: Neural Networks final project

Using the DuckieTown self driving car environment, the agent learns policies to drive in a straight line using the DDPG reinforcement learning algorithm. I experimented with two different state representations: one with raw images directly from the simulator, and the second with a VAE compressed version of the image to a single dimensional vector. The idea was that if the agent learns from a lower dimensional state representation, it may learn better policies or learn policies quicker than using raw images. My results showed that this was not true, and using raw images outperformed the latent representations.
[report] [code]

Real World Robotic Projectile Catcher Using Reinforcement Learning

Senior Design Project at Wentworth Institute of Technology

How could a robot learn to catch a ball thrown in its direction? This was the main question we aimed to tackle for our senior design project. Our robot was a XY style plotter robot with a cup attached to a gantry plate. To create a controlled projectile, we constructed a wooden ramp to roll ping pong balls down. The robot's goal was to go to the landing position of the ball before it landed. It used Q-Learning with a state representation of (x,y,velocity) of the ping pong ball and received a reward based on the Euclidean distance between the ball and the gantry plate. The ball position was determined by a camera above the ramp and two laser pointers measured the velocity. A Unity simulation of our enviornment was used to train the agent then transfered to the real life robot.
[paper] [code][slides]

Hip Hip and Gospel Music Language Modeling

CMPSCI 585: Introduction to Natural Language Processing final project

This project experiments with language models to generate both hip hop and gospel lyrics. We scraped the LyricWiki website to generate our dataset, and experimented with bigram, trigram, and LSTM language models.
[report] [code]

The Adventures Of Juk Cesium

Undergraduate OOP final project

Top-down fantasy RPG created in C++ with the SFML graphics library. The player can fight monsters, level up, explore the world, and fight bosses!
[code]