Machine Learning Projects 🚀

Welcome to my Machine Learning repository!

⚠️ Note: This is an active, working repository. I am continuously learning and experimenting, so I will keep updating this repo and adding new project files, models, and notes over time.

This repository contains various ML models, exercises, and projects I have built while learning Machine Learning concepts—ranging from basic regression and unsupervised learning to more complex classification tasks.

---

📂 Current Projects

1. Seoul Bikes Regression

A regression model to predict bike rental demand at noon in Seoul based on weather conditions.

• Dataset: UCI Seoul Bike Sharing Demand — 8,760 hourly records filtered to 365 noon observations
• Techniques: Exploratory data analysis, feature selection, univariate & multivariate linear regression, TensorFlow/Keras neural network (32→32→1 architecture)
• Key Result: Multiple linear regression achieved R² = 0.528; linear regression outperformed the neural network on test MSE, highlighting that the relationship is largely linear

---

2. MAGIC Gamma Telescope Classification

A binary classification task distinguishing high-energy gamma rays from cosmic ray background noise using telescope imaging data.

• Dataset: UCI MAGIC Gamma Telescope — 19,020 samples, 10 Hillas parameter features, imbalanced classes (gamma vs. hadron)
• Techniques: RandomOverSampler (SMOTE-style), StandardScaler, KNN, Naive Bayes, Logistic Regression, SVM, TensorFlow/Keras neural network with grid search hyperparameter tuning
• Key Results:

Model	Test Accuracy
KNN (k=1 & k=5)	82%
Naive Bayes	74%
Logistic Regression	79%
SVM	87%
Neural Network	88%

---

3. Seeds Unsupervised Learning

An unsupervised learning project applying clustering to a wheat seeds dataset to discover natural groupings without using labels.

• Dataset: UCI Seeds Dataset — 210 samples, 7 geometric features (area, perimeter, compactness, etc.), 3 wheat varieties
• Techniques: K-Means clustering (k=3), PCA dimensionality reduction to 2 components, pairwise scatter plot exploration, cluster vs. true-label comparison
• Key Result: K-Means cleanly separated the three seed varieties; PCA projection preserved the cluster structure, confirming strong natural groupings in the data

---

4. MNIST Image Classifier

A classic image classification project training models to recognise handwritten digits from the MNIST dataset.

• Dataset: MNIST — 60,000 training / 10,000 test images (28×28 grayscale, 10 classes)
• Techniques: Two PyTorch architectures compared — a fully-connected LinearMNIST baseline and a SimpleCNN with 2 convolutional layers, ReLU, and max pooling; trained with Adam optimizer and CrossEntropyLoss
• Key Result: SimpleCNN reached 99.09% test accuracy in just 5 epochs with minimal overfitting

---

5. Add Numbers ML Model

An introductory PyTorch exercise demonstrating how a neural network can learn a simple mathematical relationship from scratch.

• Dataset: 100,000 synthetically generated random number pairs in [−100, 100]; target = x₁ + x₂
• Techniques: PyTorch 2-input linear network, MSELoss, SGD optimizer, 5,000 training epochs
• Key Result: Learned weights converged to ≈ [1.0, 1.0] — the correct coefficients for addition — with loss dropping from ~2,725 to ~0.016

---

6. Linear Regression & Gradient Descent — Visual Teaching Notebook

A step-by-step educational notebook that builds deep intuition for linear regression and gradient descent through interactive visualizations. Designed for beginners with no prior ML background.

• Dataset: 200-point synthetic dataset generated from a known true line (y = 4x + 7 + noise), chosen so the reader can verify that the algorithm recovers the correct parameters
• Concepts covered: MSE loss function, convex loss surfaces, partial derivative derivations (hidden in collapsible cells for those who want the calculus), gradient descent update rule, learning rate selection
• Visualizations:
  • 3D interactive loss surface (bowl shape illustrating convexity)
  • Dual-panel interactive slider — left panel shows the gradient descent path on the loss contour map; right panel shows the evolving regression line in data space; both update in sync
  • Loss curve (convergence monitoring)
  • 3D trajectory of gradient descent overlaid on the loss surface
  • Side-by-side learning rate comparison (too small / just right / too large)
• Key insight conveyed: Gradient descent moves through parameter space (slope, intercept) — the regression line improves only because the parameters improve. The synchronized dual-panel visualization makes this abstract idea concrete and observable.

---

📊 Quick Overview

Project	Learning Type	Key Techniques	Best Result
Seoul Bikes	Regression	Linear Reg, TF/Keras NN, Feature Selection	R² = 0.528
MAGIC Telescope	Binary Classification	KNN, SVM, LR, NB, NN + SMOTE	88% accuracy
Seeds Clustering	Unsupervised	K-Means, PCA	3 clean clusters
MNIST Classifier	Image Classification	CNN (PyTorch), 2 architectures	99.09% test accuracy
Add Numbers	Intro / Regression	PyTorch linear net	Weights ≈ [1.0, 1.0]
LR & GD Visual	Educational / Optimization	Gradient Descent, Interactive Plotly, MSE Loss Surface	Recovers true m=4, b=7

---

🛠️ Technologies & Libraries Used

Deep Learning: PyTorch, TensorFlow / Keras
Machine Learning: Scikit-learn, imbalanced-learn (SMOTE)
Data & Analysis: Pandas, NumPy
Visualization: Matplotlib, Seaborn, Plotly
Data Sources: UCI ML Repository, Kaggle Hub
Environment: Jupyter Notebook, Python 3

---

Feel free to explore the notebooks and follow along as I add more projects!