This repository is filled with academic and personal projects. I wanted to showcase them, but they do not have enough substance to be within their own repositories. I might return to each project at some point to add to or retry them. Some of these projects have also inspired me to create a potentially larger project involving them all.
I implemented a generative pre-trained transformer using Pytorch. I followed a tutorial by Andrej Karpathy to create the initial Bigram and GPT models that used characters as their tokens being trained using Little Shakespeare. I decided to go further and use OpenAI's Tiktoken to create a GPT model that used sub-word tokens, but that required me to take steps to make the model more memory effifienct, so I did some research and used a half model, checkpointing, and accumulated batching to reduce the strain on the GPU.
For this project, I created a set of procedural neural networks using Jax to compile each function using jit. The pure-nature of the jit functionality made it such that I needed to think outside of the box to implement the neural network, but I eventually was able to do so and trained the model using mnist classification. The accuracy was all over the place, and I found that I would need to implement parallelization which I still need to do more research on since Jax's jit requires all inputs to be the same size going into vectorization or paralellization on a TPU or GPU.
This project was my first attempt at sentiment analysis, where I used a neural network class I implemented. The implementation was slow, and the training was near impossible on my home computer. This project taught me alot about the necessity of optimization in machine learning. Eventually, I would like to reattempt sentiment analysis.
For this project I experimented with the NLTK Python library to a rudimentary question and answer chatbot. The purpose of this project was to look at the NLTK and see what I could apply to another major project I'm working on, which is also a chatbot, but one that I want to potentially converse with the user.
This project was an experiment with neural networks. It begun with an object oriented approach, where I had a layer class and a neural network class. I was following a Sebastian Lague video to gain a understanding of neural networks in general. Once I figured that out, I decided it would be the perfect opportunity to learn Google's Jax module. That meant I had to convert my object oriented apporach into a procedural approach, since Jax requires pure-stateless functions to use jit. For the sentiment analysis project I'm working on, I'm hoping to combine the class and procedural approach.
I was assigned this project for my Intro to AI course. The idea was to create a reflex agent to handle cars for a ride sharing application. The agent's job was to pick up passengers and then drop them off. The agent prioritizes picking up passengers, and when 1 car is assigned multiple pickup or drop-offs, the shortest path is generated using a recursive search. The passengers are assigned based on the closest car to them.
Whenever I make have ideas for projects, I always like to learn how to make every component from scratch first. This helps me understand the lower-level workings of the concepts before I start using libraries and modules. I decided to learn how to make a neural network before going to use TensorFlow, PyTorch, or any other library. My neural network attempts have since come a long way, as seen in the nn-activation-function-tests project, where I used Google's Jax module.
For this assignment, I was tasked to create a Markov decision making model to play tic-tac-toe. In order to automate the testing and gameplay, I also needed to create a reflex agent to mimic the decision making of a human player. Eventually, I found that using the Markov decision making model wasn't especially beneficial in a game like tic-tac-toe, since the landscape was ever changing. The model became a glorified reflex agent. I would like to try this again. The Markov agent would be put into the into a landscape where it would benefit from being a Markov agent. I would also want to try to make a min-max/alpha-beta pruning agent for the tic-tac-toe langscape.
This assinment, I was required to make a genetic algorithm to find a solution to the travelling salesperson problem. I found difficulty implementing the genetic algorithm. I found that the problem would get stuck on being barely better than average. I would like to come back to this now that I have a better understanding of genetic algorithms.