CS7300: Practical Midterm

Jaryt Salvo

Date: 10/26/24

Fall 2024 | CS7300 Unsupervised Learning

This project focuses on implementing and comparing different unsupervised learning techniques, specifically autoencoders and dimensionality reduction methods. The project is divided into two main questions, each addressing a specific aspect of unsupervised learning.

Question 1: Super-resolution with Autoencoders (50 Points)

In this part, we implement a super-resolution task using autoencoders on the MNIST dataset. The main steps include:

  1. Data Preprocessing: Loading the MNIST dataset and adding noise to the original images.
  2. Autoencoder Design: Implementing a convolutional autoencoder using PyTorch.
  3. Model Training: Training the autoencoder to reconstruct clean images from noisy inputs.
  4. Evaluation: Plotting training curves and visualizing results on test data.

Key features of the implementation:

  • Custom MNIST dataset loader
  • Convolutional autoencoder architecture
  • Noise addition function for data augmentation
  • Training and evaluation loops
  • Visualization of original, noisy, and reconstructed images

Question 2: Dimensionality Reduction (50 Points)

This section compares two dimensionality reduction techniques: Principal Component Analysis (PCA) and Autoencoders, using the Fashion MNIST dataset. The main steps include:

  1. Data Preprocessing: Loading the Fashion MNIST dataset.
  2. PCA Implementation: Using scikit-learn’s PCA for baseline dimensionality reduction.
  3. Autoencoder Implementation: Designing and training an autoencoder for dimensionality reduction.
  4. Comparison: Evaluating both methods for reduced dimensionalities of 1, 2, and 3.
  5. Visualization: Displaying reconstructed images and training curves.

Key features of the implementation:

  • Custom Fashion MNIST dataset loader
  • Flexible autoencoder architecture for different encoded dimensions
  • PCA implementation using scikit-learn
  • Comparative analysis of PCA and autoencoder performance
  • Visualization of reconstructed images for both methods

Technologies and Libraries Used:

  • Python as the primary programming language
  • PyTorch for neural network implementation
  • NumPy for numerical computations
  • Matplotlib for data visualization
  • scikit-learn for PCA implementation

This project demonstrates the application of autoencoders in both super-resolution tasks and dimensionality reduction, showcasing their versatility in unsupervised learning. It also provides a comparative analysis between traditional (PCA) and modern (autoencoder) dimensionality reduction techniques, offering insights into their respective strengths and use cases.

The code for both questions is implemented in separate Python scripts (question_1.py and question_2.py), allowing for modular development and easy comparison of results.