Learn class hierarchy using convolutional neural networks

Abstract A large amount of research on Convolutional Neural Networks (CNN) has focused flat Classification in the multi-class domain. In real world, many problems are naturally expressed as hierarchical classification problems, which classes to be predicted organized a hierarchy classes. this paper, we propose new architecture for classification, introducing stack deep linear layers using cross-entropy loss functions combined center function. The proposed can extend any neural network model and simultaneously optimizes discover local class relationships function global information from whole while penalizing violations. We experimentally show that our classifier presents advantages traditional approaches finding application computer vision tasks. same approach also applied some CNN text classification.