jain-ruchi / neural-network

A simple neural network for classifying handwritten digits

A Simple Neural Network

Train and test a Neural Network on the MNIST Databse of handwritten digits with average accuracy of about 95%.

Requirements:

• GNU Octave or MATLAB 2016a

The Network:

• One input layer, one hidden layer, and one output layer.
• Hidden layer size = 200 units.
• Output layer size = 10 units (10 digits, i.e., 0, 1, 2, ..., 9).
• Applies nonlinear function tanh to input layer and sigmoid to hidden layer.
• Uses cross entropy for error function.

The Dataset:

• From the MNIST Database of handwritten digits.
• Training set: 60,000 examples
• Test set: 10,000 examples
• Contained in .mat files in /Dataset.

Source Files:

• trainNeuralNetwork.m: trains NN and returns two matrices, W1 and W2 containing weights.
• testNeuralNetowrk.m: tests NN on test and returns accuracy.
• sigmoid.m: applies the Sigmoid function on all elments of an array.

Instructions:

• cd to this directory.
• Run main from Octave or MATLAB.
• Optionally, uncomment line 20, line 47, and line 50 for visualization.