Parametrization of Convolutional Neural Network for Image Classification

Dasarathi, Srinivasan (2015) Parametrization of Convolutional Neural Network for Image Classification. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (766kB) | Preview
Preview	PDF (Configuration File) Download (917kB) | Preview

Abstract

In Artificial Intelligence, convolutional neural network has been the most widely used machine learning methodology of recent times for object recognition. The focus of this research is to identify a combination of key parameters that help improve the accuracy of image classification on this neural network. The network model used in this study comprises of an input layer for normalization and extraction of image data, three hidden layers for convolution, activation and pooling of the feature maps, one fully connected layer for extraction of consolidated image features, followed by an output layer where the image is classified. Sample images of size 32x32 pixels from the Kaggle’s CIFAR-10 image dataset belonging to 10 different classes has been used in this experiment.

The neural net is studied across Prototyping, Training, Validation and Testing phases, and the concept of Feed Forward and Backward Propagation has been applied in two stages – first in the hidden and fully connected layers, and later in the output layer – for different objectives as related to error convergence. The effectiveness of various parameterizations has been analysed in both these stages, including weights, bias, momentum, learning rate, regularization strength and iterative epochs. The application of different convolution, activation and pooling functions, key classifiers and novel concepts such as weight decay, weight dropouts and cross-entropy loss has been studied as part of this research project.

Item Type:	Thesis (Masters)
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Caoimhe Ní Mhaicín
Date Deposited:	19 Oct 2015 11:29
Last Modified:	20 Oct 2015 08:18
URI:	https://norma.ncirl.ie/id/eprint/2099

Actions (login required)

View Item