Author: Maciej Kuchciak
Date: January 2024
This project explores the t-SNE (t-Distributed Stochastic Neighbor Embedding) dimensionality reduction technique, applied to the MNIST dataset. t-SNE is a powerful method for visualizing high-dimensional data in a low-dimensional space, preserving local data relationships and revealing clusters.
The t-SNE algorithm is applied to the MNIST dataset. This dataset was originally created by Yann LeCun, Corinna Cortes, and Chris Burges, which contains 60,000 training and 10,000 test grayscale images of handwritten digits. These images are 28x28 grayscale, representing the digits 0 to 9. By flattening and normalizing the data, we visualize complex patterns in a 2D space.
This project highlights the use of t-SNE for understanding high-dimensional data structures while discussing important parameters like perplexity and the number of iterations used in the optimization process.
Below is an animation of the MNIST dataset elements:
t-SNEExploration/
├── README.md # Project overview and instructions
├── LICENSE # License file for the project
├── .gitignore # List of files/folders ignored in version control
├── .gitattributes # Git settings for line endings and other attributes
├── t-SNEExploration.Rproj # R project file for easy setup in RStudio
├── src/ # Source files for analysis
│ ├── Dimension_reduction.Rmd # Main RMarkdown file for t-SNE analysis
│ ├── Dimension_reduction.md # Generated Markdown output (without adjustments for GitHub Pages)
│ ├── Dimension_reduction_files/ # Auto-generated figures and images from RMarkdown
│ ├── HTML_version/ # (Optional) HTML outputs from RMarkdown
│ └── R.File/ # Folder for R script outputs from purling RMarkdown
│ ├── main.R # Script to purl Dimension_reduction.Rmd into R code
│ └── Dimension_reduction.R # Auto-generated R code extracted from Dimension_reduction.Rmd
├── docs/ # GitHub Pages website files
│ ├── index.md # Main page for GitHub Pages site (based on Dimension_reduction.md)
│ ├── _config.yml # Jekyll configuration for GitHub Pages
│ ├── Dimension_reduction_files/ # Auto-generated figures and images for GitHub Pages
│ └── images/ # Additional images for the website
├── images/ # Extra project images (plots, figures, etc.)
│ └── mnist_animation_enlarged.gif # Animation illustrating the dataset
To replicate this analysis, you need to install the required R packages and set up your environment:
Prerequisites:
- R (>= 4.0)
- RStudio (Optional but recommended)
- TensorFlow backend for Keras
# Install necessary packages
if (!requireNamespace("keras", quietly = TRUE)) install.packages("keras")
if (!requireNamespace("magick", quietly = TRUE)) install.packages("magick")
if (!requireNamespace("ggplot2", quietly = TRUE)) install.packages("ggplot2")
if (!requireNamespace("Rtsne", quietly = TRUE)) install.packages("Rtsne")# Install TensorFlow
keras::install_keras()This project is licensed under the GNU General Public License v3.0 (GPL-3.0). You are free to modify, share, and use this project under the terms of the GPL-3.0 license. If you distribute modified versions of this work, you must release them under the same license.
Contributions are welcome! If you have suggestions or improvements, feel free to open an issue or submit a pull request.