COMP 551 — Machine Learning Projects

Contributors: Nicolas Smits, Bassem Mikhail
Projects completed for COMP 551: Introduction to Machine Learning (McGill University).

This repository contains four end-to-end machine learning assignments covering regression, neural networks, computer vision, and transformer-based NLP. Each project includes clean, reproducible implementations with supporting experiments and analysis.

---

🧮 A2 — Regularization & Model Evaluation

Topic: Linear Regression with Radial Basis Functions

This project explored the relationship between model complexity, bias–variance tradeoffs, and regularization. A linear model was enhanced with Gaussian RBF features to capture nonlinear structure in synthetic data.
Key components:

• Implementation of RBF feature generation
• L1 (Lasso) and L2 (Ridge) regularization
• Train/validation/test splitting
• k-fold cross‑validation for hyperparameter selection
• Analysis of underfitting vs. overfitting through learning curves

The assignment emphasized how regularization stabilizes training and improves generalization in high‑dimensional feature spaces.

---

🧠 A3 — Image Classification with Neural Networks

Datasets: OrganAMNIST (medical images)
Models: Scratch‑built MLP, CNN (PyTorch), ResNet (pretrained)

This project implemented and compared multiple neural architectures on grayscale medical image classification:

• Multilayer Perceptron (from scratch)
  Built using NumPy with manual backprop and SGD
• Convolutional Neural Network (PyTorch)
  Implemented using standard conv/pool blocks
• ResNet Transfer Learning
  Fine‑tuning a pretrained model to evaluate benefits of learned visual representations

Experiments examined:

• Architectural depth and width
• Data preprocessing and augmentation
• Regularization (dropout, weight decay)
• Optimization strategies and convergence behavior

---

💬 A4 — Emotion Classification with LLMs

Dataset: GoEmotions (27 emotion labels)
Models: Naive Bayes, Softmax, Random Forest, XGBoost, BERT

This NLP project investigated approaches to emotion classification, comparing traditional ML baselines with transformer‑based language models.

Highlights:

• Text cleaning, tokenization, and label preprocessing
• Baseline classifiers using TF‑IDF features
• Fine‑tuning BERT for multi‑label emotion detection
• Exploration of attention mechanisms and contextual embeddings
• Evaluation using macro/micro F1, precision, and recall

The results demonstrate the substantial performance advantage of pretrained transformers over classical bag‑of‑words methods for nuanced sentiment analysis.

---

📂 Repository Structure

COMP-551-ml-projects/
│
├── Assignment2/
│   └── ... code, figures, reports
├── Assignment3/
│   └── ... image models & analysis
├── Assignment4/
│   └── ... NLP models & BERT fine-tuning
└── README.md

---