Machine Learning From Scratch

This repository contains a collection of foundational machine learning projects implemented entirely from scratch using NumPy, PyTorch, and other low-level tools, without relying on high-level machine learning APIs. Each project is an independent notebook focused on understanding how core ML algorithms work under the hood.

These implementations were built as part of a hands-on effort to deeply understand key concepts in supervised, unsupervised, and reinforcement learning.

---

Projects

lasso_ridge_reg/

Lasso & Ridge Regression
Manual implementation of linear regression with L1 (Lasso) and L2 (Ridge) regularization using gradient descent. Includes custom cost functions, gradient logic, and weight updates.

• Implements regularization from scratch
• Visualizes weight shrinkage behavior
• Comparison with scikit-learn (optional)

---

Kmeans/

K-Means Clustering
An unsupervised clustering algorithm implemented manually using NumPy — no external clustering libraries used.

• Random centroid initialization
• Vectorized distance computation
• Iterative clustering until convergence

---

GMM_EM/

Gaussian Mixture Models with Expectation-Maximization
A from-scratch implementation of the EM algorithm to train GMMs on a multivariate dataset.

• Uses scipy.stats for multivariate Gaussians
• Updates responsibilities, means, covariances iteratively
• Log-likelihood computation for convergence check

---

SVM-KSVM/

Support Vector Machine & Kernel SVM
Manual implementation of linear SVM with hinge loss, and kernelized SVM using cvxopt for solving the quadratic programming problem.

• Gradient-based linear SVM
• QP-based kernel SVM (polynomial, RBF kernels)
• Dataset split across training, validation, test sets

---

RL/

Reinforcement Learning in a Custom Environment
Implements a basic reinforcement learning agent (epsilon-greedy strategy) in a custom visual environment built with pygame.

• Custom gridworld with start, goal, ice, rocks, and holes
• Q-table logic for action selection and learning
• Visualization using pygame + imageio for GIF exports

---

Goal of This Repo

This repository is not meant to be a production-grade ML toolkit. Instead, it’s a learning lab, a place to:

• Understand algorithms by implementing them
• Experiment with data, math, and code
• Move beyond "import and go" ML workflows

---

Tech Stack

• Python (NumPy, Matplotlib, PIL, PyTorch)
• cvxopt (for QP in SVM)
• pygame (for custom RL environment)
• scikit-learn only for optional benchmarking

---

Folder Structure

Each folder is a standalone project containing:

• A Jupyter Notebook with full code and explanations
• .npy datasets (for projects that require custom or preprocessed data)
• Image assets (for RL environment visuals)

├── lasso_ridge_reg/
│   └── lasso_ridge_reg.ipynb
│   ↳ Implements linear regression with L1 & L2 regularization from scratch

├── kmeans/
│   └── kmeans.ipynb
│   ↳ Custom K-Means clustering with random init & iterative updates

├── GMM_EM/
│   ├── GMM_EM.ipynb
│   └── Dataset(X_data).npy
│   ↳ Gaussian Mixture Model trained using the EM algorithm

├── SVM-KSVM/
│   ├── SVM_KSVM.ipynb
│   ├── Dataset0(X_data).npy
│   ├── Dataset1(Train).npy / (Validation).npy / (Test).npy
│   └── Dataset2(...)
│   ↳ SVM & Kernel SVM using gradient descent and CVXOPT QP

├── RL/
│   ├── RL.ipynb
│   └── img/
│       └── (visual assets like elf, goal, rock, etc.)
│   ↳ Reinforcement learning in a custom Pygame gridworld