Implemented Multilayer Feedforward Neural Network (MLFNN) with Backpropagation (BP) learning.The aim was to code a complete handwritten digit recognizer and test it on the MNIST dataset.
DATASET: The MNIST database of handwritten digits, available from this page, has a training set of 60,000 examples, and a test set of 10,000 examples. It is a subset of a larger set available from NIST. The digits have been size-normalized and centered in a fixed-size image. The dataset and its description can be downloaded from:http://yann.lecun.com/exdb/mnist.
IMPLEMENTATION: Recognizer first read the image data, extract features from it and use a multilayer feedforward neural network classifier to recognize any test image. Number of hidden layer nodes are 90.This was obtained by trial-error and cross validation method. File "inputweights90.txt" Contains weights of nodes from input to hidden layers,whereas file "hiddenweights90.txt" weights of nodes from hidden to output layers. Also implemented K-NN to compare the results of the two methods RESULTS: Neural NW: 94.12 1NN: 96.33 5-fold: 95.35
Files Description: first.m: Training model, NN algorithm is applied. test.m : Testing the trained model hiddenweights90.txt: Weights for hidden to output layer. inputweights90.txt: Weights for input to hidden layer. knn.m: Applied KNN classifier for comparision purpose. graph.m : To plot graph B/W Error Rate vs No. Of Features.