Machine Learning Coursework

SMU MSDS — Applied Machine Learning

Overview

Four-part coursework progressing from Python data structure mastery through applied deep learning. The final assignment explored AI-assisted learning via Cursor’s Model Context Protocol (MCP) — turning the ML assignments themselves into an object of study for understanding code with AI.

Assignments

HW1 — Python Data Structures

Comprehensive review of Python built-in data structures: lists (append, extend, index, insert, remove, pop, sort, reverse), dictionaries, sets, and their associated operations. Demonstrated through annotated examples covering time complexity tradeoffs.

HW2 — Hyperparameter Optimization

Designed a multi-classifier evaluation framework comparing Logistic Regression, Random Forest, and SVC across three different hyperparameter configurations each. Used KFold cross-validation for reliable performance estimates and GridSearchCV for systematic search. Extended to Optuna (Bayesian optimization) for neural network hyperparameter tuning — significantly faster than grid search for continuous parameter spaces.

Key results: Evaluated accuracy, precision, recall, and F1 across all classifiers. Generated matplotlib comparison plots for model selection visualization.

HW3 — Transfer Learning with PyTorch

Built a SimpleCNN architecture from scratch:

Conv2d(1, 32, 3) → ReLU → Conv2d(32, 64, 3) → ReLU → MaxPool2d → Dropout → FC(9216, 128) → FC(128, 10)
Trained on MNIST digit recognition for 5 epochs with Adam optimizer
Froze convolutional layers (which learned general edge/shape detectors)
Replaced final FC layer for custom handwritten character classes
Fine-tuned only the new classification head on a smaller custom dataset

This achieved faster convergence and better accuracy than training from scratch, demonstrating the practical value of transfer learning for limited-data scenarios.

HW4 — AI-Assisted Learning with Cursor + MCP

Explored how AI tools (Cursor with MCP) transform the learning experience for ML coursework. Loaded HW2 into a Cursor workspace and used the AI to explain each code block, visualize what each algorithm is doing conceptually, and answer targeted questions without needing to search documentation.

Key insight: MCP allows the AI to have direct context about the running codebase, making it qualitatively different from pasting code into a chat interface. The debugging and explanation loop is orders of magnitude faster.

Technology Stack

Category	Tools
Deep Learning	PyTorch, torchvision, Conv2d, DataLoader
ML Framework	scikit-learn, GridSearchCV, KFold
Optimization	Optuna (TPE sampler)
Data	numpy, pandas, matplotlib
Computer Vision	OpenCV, MNIST dataset
Dev Tools	Cursor IDE, MCP

Key Takeaways

Transfer learning is most valuable when your target dataset is small relative to the pre-training dataset
Optuna’s Bayesian search outperforms grid search by ~3x when parameter space exceeds 3 dimensions
KFold cross-validation gives more reliable estimates than a single train/test split, especially with n < 10,000
AI-assisted code exploration (via Cursor MCP) dramatically accelerates comprehension of unfamiliar ML code

GitHub Repository

← Back to Projects