Handwritten Digits Recognition

Implementation of a Neural Network with three layers for handwritten digits recognition.

The neural network has 3 layers - an input layer, a hidden layer and an output layer. Images are pixel values of digit images and of size 20 x 20, which gives 400 input layer units. The cost function is computed by using Feedforward and gradiant is computed by doing a Backpropagation.

Note that the cost function is also regularized.

The training data set - training_data.mat has 5000 examples.

You can change the number of iterations in the file 'training.m' at

options = optimset('MaxIter', 50);

and change '50' to the number you want.

Some experiments done before:

Iterations       Accuracy on the training set

----------------------------------------------

10                          76.3%

50                          95.3%

100                         98.7%

This project is part of Stanford Machine Learning course on Coursera.

Tutorial

1.Starting the Octave and move to the folder which contains all the source files.

2.Type

training

in the Octave to train a Neural Network. It also saves weights(Theta_1, Theta_2) into wieghts.mat which will be used by prediction.m .

3.Type

prediction

in the Octave to predict a random picture of a digit.