Aprendizado de operadores de agregação do tipo média ponderada ordenada em redes neurais convolucionais

Abstract

In convolutional neural networks, aggregation operations are performed in the convolution, pooling and fully connected dense layers. Promising results have been obtained in recent years when using ordered weighted averaging operators, better known as OWA operators, to aggregate data within convolutional neural networks. There are recent works demonstrating that there is a performance gain when using OWA operators, training their weights, to perform the pooling operation, when compared with the most usual operators (maximum and average). Other studies have shown that OWA operators can be used to learn additional order-based information from the feature maps of a certain layer, and the newly generated information is used to complement or replace the input data for the next layer. The purpose of this dissertation is to analyze and combine the two mentioned ideas. Several tests were done to evaluate the performance change when applying OWA operators to classify images, using the VGG13, Network in Network and AlexNet models and the CIFAR-10 and CIFAR-100 datasets.