Computer Science Student
I'm Michael, an aspiring machine learning engineer/researcher studying Computer science.
I plan to pursue a PhD, focused on research in Machine Learning algorithms and methods. I'm still
exploring literature to develop my specific research interests.
I care about collaborating, learning, and finding solutions to even the smallest things. In
my free time, I enjoy deejaying and producing electronic music,
skateboarding,
playing video games (Overwatch, Zelda, & SSBU), and baking (bread, pizza, & sweets). At school,
I run an EDM specialty show called Two Step Authentication
and another called Electronic Soundsystem
the 100% student-run WREK 91.1FM,
and film for the Georgia Tech Skateboarding Club.
Python (TensorFlow, Requests, Pandas/NumPy), Java, HTML/CSS, Bash
Git, Agile Development, Communication, Responsibility & Timeliness
Formal, Reliable, and Succinct
A computer vision model that classifies mating mechanisms and other important behavioral interactions between cichlid fish. Built on DeepLabCut, we use coordinate and velocity information to classify 6 different behaviors. Currently, we are analysing our dataset and automating the creation of our behavioral dataset. In the next steps, we will develop the architecture for our network. This project is supported by the McGrath Lab in the school of Biological Sciences at Georgia Tech.
Figure 1: Unfiltered pose predictions on a short clip of our cichlids.A collection of machine learning models using TensorFlow and Scikit-learn libraries trained on data from cognitive and behavioral testing sourced from the PPMI database. Some methods used include CNNs, XGBoost, and SVMs. Our most robust model trained on biomedical voice measurements from 219 patients performed at 88% accuracy (# correct predictions / # truth). I left this project in May 2023 as the semester ended...
With this project, we intended to develop low-cost solutions to help detect Parkinson's through a variety of tests. We chose to develop a model for several cognitive tests that are easily replicable at home, such as the spiral test (as pictured below) and Boston naming test. However, we struggled with small datasets and general unfamiliarity with a lot of the concepts we were working with. In retrospect, we could have done more with our data, e.g. use different architecture, as well as focus more deeply on strengthening our models. Of course, it is also possible that these tests (Boston naming test and trail-making test) are simply not good indicators of whether someone has Parkinson's.