The effect of feature selection methods on deep convolutional neural networks

Since there has been large increase in the amount of data in recent years, and these data have become increasingly higher dimension, it has become even more important to extract discriminative high-quality features for this data. Specifically, in image classification, Deep Convolutional Neural Networks are frequently used because of their high discrimination power. However, the dimension of the features learned by Deep Convolutional Neural Networks are very high and may include irrelevant or unnecessary features for classification tasks. Removing these irrelevant and unnecessary features may lead to increased classification performance. Although there are a lot of publications about the Convolutional Neural Networks in the literature, there are not many publications on the selection of the features extracted by them. In this thesis, various feature selection methods have been applied to the features obtained from two fully connected layers of well-known deep neural network architectures such as AlexNet, VGG16 and VGG19, and the classification performances of these selected features have been compared.