Nicholas Leland

github.com/nick-leland/ascii_diffusion

Pictures are defined as Arrays (LxWx3 for color, LxWx1 for grayscale) in which each value is a range of 0-255. What if rather then representing colored pixels, we instead trained a diffusion model around individual ASCII characters. In a similar vein to how diffusion models currently learn underlying object definitions, perhaps our model could learn to understand the core values of ASCII art, such as edge definition, shading, and key characteristic identification.

github.com/nick-leland/distortionml

DistortionML was my attempt to solve the problem of image distortions and to explore whether or not a machine learning model could determine if image transformations were present based on the underlying representation of pixels within an image.

While this project wasn't exactly successful in that sense, it id allow me to pursue some interesting ideologies within image manipulation. The fundamental that I built this project on involved first, having to define image transformations. Refreshing my calculus at the time, I thought it would be interesting to define transformations as mathematical functions. We would generate a Vector Gradient Field based on a multivariate function. For instance,

\[ f(x) = x^2 + y^2\]

would represent a simple cone function. After defining several functions, I then created a dataset by cloning the ImageNet dataset and applying a specified transformation to every image within the dataset, following the YOLO localization format.

The next step for this project was to pursue the idea of a machine learning model to learn functions from images . This would be derived off of something like a 2-Dimensional Taylor Series model, starting to lean towards the Universal Approximation Theorem. I do tend to continue to evaluate this, but I think that likely it is not within my grasp of technology to develop.

github.com/nick-leland/jax-facial-emotion-detection

I completed an MIT course on Applied Data Science which ended up being a bit of a let down, but it did give me the baseline for this interesting project. While most of the assignments were very boiler plate like, for the final project, I decided to go off the rails and build a CNN from within the Jax/Flax library. This was quite an enjoyable experience, after dealing with the sharp bits of Jax (;

github.com/nick-leland/rd2l-pred

I previously enjoyed playing Dota 2 (Invoker/Tinker Main btw) and many of my friends continue to play. Many partake in a league known as RD2L, an amateur Dota 2 League. The most interesting aspect of this league is the way that teams are generated. There are two sign ups, Captains or Players. Captains are given an amount of currency based on their MMR (Match Making Rating, A Dota 2 skill level evaluation, similar to ELO) and then must bid directly on players with a live auction system. I developed a model that would take different aspects of the current league and statistics about each individual player, and give a predicted price for that player. Captains would then be able to use this information in order to help quickly evaluate whether or not a player was under or over evaluated and bid accordingly.

Presentation Slides Google Sheets

For my Capstone Project for my Mechanical Engineering Degree, I designed and theoretically tested an aerodynamics component for my 1988 Toyota Supra. This was a really fun project, I had always been interested in Computational Fluid Dynamics and this gave me an outlet to apply my education to a personal passion project.