An Investigation of Convolutional Neural Networks

Abstract

This report documents the process of creating a Convolutional Neural Network from basis in Matlab. Convolution, Fully Connected, Average Pooling, Max Pooling, Rectifying, Softmax and Cross Entropy Layers were implemented as Object Classes. This allows construction of Convolutional Neural Networks with arbitrary topology comprised of such layers. Experiments verifying that the features learnt by the Convolutional Neural Networks are what allows them to improve upon Neural Network Classifiers were carried out and detailed. The redundancy in the output of the Convolutional Neural Network was examined, and revealed that reducing such a Networks output to as few as eight of its principle components still contains enough information to correctly classify handwritten digits with a precision of 92.09%. Finally the effect of choosing between Average and Max pooling was explored, revealing Max Pooling tends to lead to more precise classifiers.