Tradeoffs Between Embeddings in Different Models of the Hyperbolic Space

Project done for Stanford's CS 229 (Machine Learning).
There is little theoretical understanding of the tradeoffs in using different models of hyperbolic space and how they are influenced by the properties of the graph they are applied to. In particular, there are four common models that we set out to better understand: the hyperboloid model, the Poincaré disk, the half-plane model, and the Beltrami-Klein model. Through experimentation with various synthesized graphs based on radius, height, and density and also graphs that model real-world data, we compared various metrics for each of the four models when applied to the graphs.

CNN Transfer Learning for Visual Guitar Chord Classification

Project done for Stanford's CS 230 (Deep Learning).
Contrary to previous aural approaches, this project seeks to build a visual classifier to determine whether a guitarist is playing a C, D, Em, F or G chord. The input to our system is a 1080 x 720 RGB image of a person playing a guitar. We then autonomously bound/crop the hands, passing the cropped image into our model to get an output prediction. Note, the data was personally gathered and experimented on ResNet18 and GoogLeNet architectures.

CNNs for Photorealistic Computer-Generated Imagery Detection

Project done for Stanford's CS 221 (Artificial Intelligence: Principles and Techniques).
To counteract the growing danger of CGI’s advancement, our project sought to build a binary classifier that would be able to detect whether an image was a real photographic image or just a photorealistic computer-generated image. Our inputs were RGB images of “real” and “fake” scenes, where we outputted both our classification as well as our confidence.

On the Evolution of Dark Matter Halo Properties after Major and Minor Mergers

Research conducted at UC Santa Cruz's Astrophysics Department.
In this work, we analyzed ~80,000 dark matter halos from the Bolshoi-Planck cosmological simulation and studied halo evolution during relaxation after major mergers. By analyzing their property evolutions and peak distributions, we designed and implemented a novel algorithm for identifying/classifying major and minor mergers. Through our merger statistics, we discovered immense evidence that minor mergers were the culprits behind unaccounted relaxation.

Orbit Modeling of Asteroid 2017 BM31

Conducted at Yale University's Leitner Observatory.
Operating the observatory's 12" telescope, we had to capture nightly pictures of 2017 BM31. Processing those pictures, we then used astro-photometric software MaxIm DL to prepare image analysis and track the asteroid. Later, we coded Python modules to use reference stars and generate coordinate maps for images. Finally, we coded the Gauss-Newton algorithm to determine asteroid orbit, optimizing with gradient descent and a genetic algorithm.

Using a Novel Methodology to Constrain the SMBH-Galaxy Coevolution and Analyze the Selection Bias

Research conducted at UC Santa Cruz's Astrophysics Department.
Compiling datasets of galaxies and supermassive black holes, we devised statistical convolution to estimate redshift evolution, construct number density models, constrain galaxy properties, and affirm a SMBH-galaxy coevolution. Additionally, we addressed the controversial Shankar et al. (2016) selection bias — our statistics found that such a bias has little impact on our SMBH-galaxy relations.