Additional Feature Layers from Ordered Aggregations for Deep Neural Networks

A Study of OWA Operators Learned in Convolutional Neural Networks

“…Regarding the selection of the operators, they manually established fixed operators. In contrast, our OWA layer proposal in [13] is also composed of several OWA operators that generate new fused feature maps, but with important differences. With respect to the usage of the new feature maps, in our proposal, the fused feature maps are used to augment the information of the network, preserving the original feature maps.…”

“…The OWA layer that we explore in this work was previously defined in [13,38], where we presented some preliminary results about the impact of the layer in CNNs. In the previous works, we found an increase in the accuracy of a medium-sized VGG13 network on a large image dataset, but we did not study the learned OWA operators and the way they codify information.…”

“…We propose the integration of OWA operators in a general CNN via the use of a special OWA layer [13,38]. This layer will be placed inside the convolutional region of the network and will modify the input data, a 3D matrix of activations, by adding newly generated feature maps by means of OWA operators.…”

“…Aiming at integrating OWA operators in CNNs to improve the performance of CNNs in image classification problems, in a previous work [13], we introduced the OWA layer. The OWA layer modifies the usual CNN architecture by using OWA operators capable of aggregating the feature maps of a CNN based on global information (feature map metrics), even when inserted in the early layers of the network.…”

“…In our previous work [13], we already showed how the OWA layer added new knowledge to a mid-size CNN (based on the VGG13 architecture) on a relatively large dataset. In this work, instead, we are interested in the specific operators learned, and we decide to choose a reduced architecture to facilitate the study.…”