Projects
Welcome to my side projects page where I commemorate some side projects I have scribbled up over the years on my spare time. Naturally, the amount of the latter highly fluctuates as a function of employment duties, number of children, etc. and so the output is and will always be extremely uneven.
The older ones from 2017/2018 were projects I used to make my first baby steps in ML and more serious programming and in certainly shows. At the same time, looking at them evokes quite some pride and sentimentality about the road taken since then and I decided to showcase them anyway!
After a loooong hiatus caused by, well, the appearance of two rascals and focusing on my proprietary daytime work, I decided to start creating public side projects. Now I just have to keep with it 😅. These days I am mostly interested in keeping up with the Jones's about Generative AI (LLM's, diffusion models).
Retro Diffusion
Ongoing educational journey into probabilistic diffusion models. The goal is to provide a progressively expanding collection of PyTorch reference implementations of the most important diffusion model milestones. Then name implies that by the time I finish, these models will be fairly "retro" (some of them already are).
Tone of Voice Predicts Political Attitudes - paper
Can you tell lawyers ideology by the way they speak? Our work that grew out of a class project in 2018, joining speech samples extracted from supreme court hearings with data on their political donations, indicates that you "probably" can at least to some degree.
Tic-Tac-Toe
Fairly simplistic implmentation about a couple of algorithms to play tic-tac-toe with the computer. In 2018, I was learning Java for data structures and participated in an awesome seminar about AI playing games (this was soon after AlphaGo swept the news). I was very new to everything, but had a lot of fun. Maybe one day I'll give it a facelift.
Population estimation from satellite imagery.
This was also a class project, coincidentally my first ever ML project in 2017. We used Neural Networks, first trained from scratch, then a finetuned VGG16 model, to predict the population (by census) of an area based on satelite imagery scraped from google maps.