Number Recognition Neural Net

A neural network classifier for recognizing numbers from the MNIST database.

To Run

Using Python 3.7 or greater:

• Open the data directory
• Run the mnist_csv3.py script (courtesy of USC CSCI 561) to convert the mnist images into csv data

Four csv files should then be present in the data directory. Then,

• Change directory back to the root of this repository
• Run NeuralNetwork3.py

Details

Structure

• Expects 784 inputs (28x28 pixels in each image)
• 2 hidden layers with 32 neurons each
• 10 output neurons representing the NN's choice of 0 through 9
• Sigmoid activation used on the inner neurons
• Softmax activation used on the output neurons

Training

• Weights initialized using Xavier Initialization
• Uses Stochastic Gradient Descent with momentum
• Uses cross-entropy loss