This repository contains an assortment of my projects in machine learning. Here's a brief overview of some of what you'll find here:

The ml_from_scratch folder contains notebooks where I implement various machine learning algorithms from scratch using only NumPy. Each notebook contains detailed documentation, mathematical explanations, and demos to experiment with the algorithms.

I built these projects to practice the skill of taking mathematical formulas and turning them into vectorized code. As such, I did not review anyone else's code related to these algorithms but rather studied their mathematical formulations.

Currently, the algorithms I have implemented are:

• Neural network
• K-Means clustering
• Linear regression
• Logistic regression

A configurable fully-connected neural network that can be used for regression or classification tasks. This is an all-purpose algorithm that can be applied to many tasks like word embeddings, time series predictions, or multiclass classification (especially when using tabular data).

Key algorithms used include: backpropagation and gradient descent, cross entropy loss, L2 regularization, sigmoid and ReLU activations, forward propagation, normalization, parameter initialization, and a training and validation loop.

Results on sample data:

Classification	Regression

A vectorized implementation of the K-Means clustering algorithm that groups an unlabeled dataset into $k$ clusters (labels) and can then classify new data based on proximity to the cluster centers. Uses include: finding related items in a dataset, segmenting customers, creating color palettes from images, grouping text documents, quantizing a dataset, categorizing transactions.

Key techniques used include: Euclidean distance, cosine similarity, vector normalization (L2), and a lot of plotting code for animated visualization.

Generate color palettes from images

Fit to any numerical data

A logistic regression model for binary classification using gradient descent to optimize the model's parameters. Uses any number of input features.

Key algorithms include: binary cross entropy loss (log loss), sigmoid, and threshold selection using the inverse of sigmoid.

Training progress on a sample dataset, showing how the decision boundary is updated through gradient descent:

A linear regression model trained using gradient descent with optional early stopping. Uses any number of input features (i.e., multilinear regression).

Key algorithms used include: mean squared error loss, model training through gradient descent.

Results from a multilinear regression demo: notice that y_hat (color gradient) closely matches y (vertical axis).

In my dominant color extraction notebook, I use the K-Means algorithm to estimate the main color for an image, which I then apply to the task of detecting a vehicle's color given an image of that vehicle. I used this technique in my vehicle specs project as part of a feature engineering pipeline to augment a dataset with vehicle color information.

A comparison of different Gradient Boosting Machine (GBM) implementations on a classification dataset, including visualizations.

In this notebook, adapted from a tutorial by Usman Malik, I implement a method for extractive text summarizing and use it to summarize Wikipedia articles.

I created a sort-of API to generate images using the Craiyon text-to-image model, available at: https://www.craiyon.com/

Note: I last updated this notebook in fall 2022; the website may have changed since then and the web requests might not function as intended.

I took a computational linguistics course and compiled my work on course assignments into a single notebook. In this notebook, I compute precision and recall for various NLP algorithms, compare categories and word frequency distributions across a corpus, classify sentiment, use part-of-speech tagging, predict categories using synsets from WordNet, and use LDA to visualize topics, compute the number of topics using perplexity and coherence, and use a lot of RegEx, web scraping, NLTK, and topic modeling with Gensim.