MNIST Digit Recognition

Overview

In this project, we perform hand-written digit recognition on the famous MNIST dataset. The dataset contains 60,000 training samples and 10,000 testing samples. Two experiments were conducted on this dataset. Moreover, an SVM classifier with various kernels and a CNN classifier with 2 convolutional layers was designed to solve the classification problem.

Dependencies

• MATLAB (2019b version used)
• Deep Learning Toolbox
• Parallel Computing Toolbox (if you have GPU with 3.0 or higher compute power and want to run GPU accelerated training)

Run

• Open the project folder in MATLAB and run digit_recognition.mlx
• digit_recognition.mlx provides the options for running the SVM and the CNN classifiers
• For SVM, you get the options to specify the dimensionality reduction (DR) method as well as the choice of kernel
• Note that the CNN takes about 2 hours (on GPU) to train and give out an accuracy
• Make sure you have the Parallel Computing Toolbox installed on your system for MATLAB to access your GPU
• Note that the SVM classifier can take from 5 minutes to an hour depending on your choice of DR method and kernel