ML-Foundry

A collection of machine learning exercises implementing core algorithms and computer vision techniques from scratch. Developed as part of postgraduate study in machine learning.

Overview

This repository contains hands-on exercises that build foundational understanding through implementation rather than relying on high-level libraries. Each notebook focuses on understanding the underlying mathematics and algorithms that power modern ML systems.

Exercises

Exercise 1: Facial Landmark Detection in Video

Analyses video sequences to detect the frame where a person's mouth is open widest using facial landmark detection.

Key Concepts:

• Video frame extraction and processing with OpenCV
• 68-point facial landmark detection using deep learning models
• Feature-based distance calculations between facial points
• Temporal analysis across video sequences

Applications: Emotion recognition, speech analysis, human-computer interaction

Exercise 2: Linear Regression with Gradient Descent

Implements linear regression from first principles to model the relationship between years of experience and salary.

Key Concepts:

• Gradient descent optimisation
• Manual implementation of parameter updates
• Train-test split methodology
• Model evaluation and visualisation

Critical Analysis: Includes identifying flaws in sequential data splitting versus random splitting for representative train-test sets.

Exercise 3: Logistic Regression for Binary Classification

Implements binary classification using logistic regression on synthetic data with 1,000 samples.

Key Concepts:

• Sigmoid activation function
• Binary cross-entropy loss
• Training for 2,000 epochs tracking accuracy and loss
• Visualisation of model convergence

Exercise 4: Image Denoising Filters

Compares mean filtering and median filtering techniques for image denoising on grayscale images.

Key Concepts:

• Spatial filtering with 3×3 kernels
• Mean filter: averaging pixel values
• Median filter: replacing pixels with median values
• Comparative analysis of denoising effectiveness

Exercise 5: Multi-Layer Neural Network for XOR

Solves the classic XOR problem using a multi-layer perceptron, demonstrating how neural networks handle non-linearly separable data.

Key Concepts:

• Forward and backward propagation implementation
• Two-layer architecture (3 input, 2 hidden, 1 output)
• Backpropagation algorithm from scratch
• Non-linear decision boundaries

Why It Matters: XOR is not linearly separable—demonstrating the necessity of hidden layers for complex problems.

Prerequisites

pip install -r requirements.txt

Getting Started

git clone https://github.com/vl-pixel00/ML-Foundry.git
cd ML-Foundry
jupyter notebook

Recommended Order

1. Exercise 2 - Linear regression fundamentals
2. Exercise 3 - Classification with logistic regression
3. Exercise 5 - Neural networks and non-linear problems
4. Exercise 4 - Image processing techniques
5. Exercise 1 - Computer vision application

Educational Approach

• From-scratch implementation for deeper understanding
• Mathematical foundations with underlying equations
• Visualisation for intuitive concept grasp
• Critical analysis questions throughout