Hand Written Digit Recognition using Backpropagation Algorithm

The objective of the project is to generate an Artificial Neural Network (ANN) for hand written digit recognition application using two variations of BP algorithm variations of Back Propagation (BP) algorithm: simple BP and BP with momentum. Different patterns of handwritten digits are used to analyze the performance of BP algorithm. The criteria of performance being the accuracy and amount of time taken. Various parameters such as learning rate, number of hidden neurons in hidden layer, momentum term and number of training iterations are used during the analysis of the BP algorithm.

INTRODUCTION:-

Digit recognition is important to the overall system in terms of performance (accuracy and speed). In recent years, different algorithms and classification approaches have been presented to solve this problem. Digit recognition is an important component of handwritten character recognition system due to its wide application. From more than three decades can achieve high classification high recognition rates in the area of recognition of handwritten numerals Artificial Neural Networks (ANN) has been intensively used to solve the complex engineering problems. In ANN the basic element is the neuron which is from the nerves system of the human. Neuron can be a computational and communication function. Herein we have used "log sigmoid" function. The flow of information in ANN is in parallel manner while the knowledge is distributed among the processing units or neurons Some Parameters of artificial neural network while constructing considered are:

• Learning rate • Accuracy of results • Momentum Tern • Number of iterations used to train the Neural Network Model • Architecture of the Neural Network • Numer of hidden neurons in the hidden layer.

The main difference between in simple Back propagation and Back propagation with Momentum is the way by which weights are adjusted. The Back propagation with momentum uses an additional factor of momentum. In Back propagation neural network, the learning rate parameter can have a significant effect on generalization accuracy. The selection of small or large learning rate affects the generalization accuracy and training the neural network architecture directly

Ref:- https://www.tutorialspoint.com/artificial_intelligence/artificial_intelligence_neural_networks.htm

TRAINING AND TESTING :-

1. Data Sets:-

The neural network uses MNIST as the benchmark dataset of images of segmented handwritten digits. MNIST database includes a training set of 60,000 images and a test set of 10,000 images. The MNIST digit database contained fixed size images and digit image (foreground pixels) is center alignment with respect to the background pixels. The MNIST digit database is good database for applying learning techniques and patterns recognition methods

2. Training:-

Training sets are used to train and adjust the weights of ANN. The idea is that we do not want to make the ANN too specific, making it give precise results for the training data, but incorrect results for all other data (in case of new input). Otherwise ANN would be over-fitted.

3. Testing:-

The performance of neural network system is compared using different value to different learning rates for Backpropagation algorithm. The system also compares the efficiency when the momentum is changed.

PSEDO CODE AND IMPLEMENTATION PROCEDURE:-

Files :

1. train_single_example.m
2. backprop.m
3. feedforward.m
4. results.m
5. testing.m

backprop.m contains :- function [W1, W2, B1, B2] = backprop(num_of_hidden_neurons,eta,num_of_epochs,num_of_examples,momentum_const)

It takes training parameters in the function and returns the weight matrixes after training from MNIST database using backpropagation with stochastic gradient descent. Following is pseudocode of backprop.m

1. Load input images and its respective labels from database into 60000X784 matrix. and 60000X1 matrix and randomly initialize weight and bias matrices.
2. For each input image:- a)call the 'train_single_example' function to get new weight using back propagation algorithm without using momentum. b)Include momentum term using the previous ΔW to get new set of weights c) Store the new ΔW for subsequent calls.
3. Return the calculate weight matrix. train_single_example.m does the following for each input example: a) Feedforward --> Calculate hidden_layer_output and output_layer_output. b) Calculate δ for output layer neurons. c) Backpropagate δ for hidden layer neurons. d) Calculate ΔW using δ calculate above in b) and c). e) Update and return new weight matrix W using the above ΔW. testing.m --> uses the testing data to return the number of correctly matched images of digits. results.m --> is the wrapper for backprop.m, is used to find the relationship between testing result and parameters. It is the final result file.

Results:- Model was trained using the MNIST database. Maximum accuracy of 91.3% was observed for the testing data. Number of images in test database = 10000

PAPERS REFFERED :-

http://www.sciencedirect.com/science/article/pii/S0031320311004006 https://cs.nyu.edu/~wanli/dropc/dropc.pdf http://ce.sharif.edu/courses/85-86/2/ce667/resources/root/15%20- %20Convolutional%20N.%20N./fugu9.pdf http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=5625732