Hierarchy‐guided neural network for species classification

Species classification is an important task which the foundation of industrial, commercial, ecological and scientific applications involving study species distributions, dynamics evolution. While conventional approaches for this use off-the-shelf machine learning (ML) methods such as existing Convolutional Neural Network (ConvNet) architectures, there opportunity to inform ConvNet architecture using our knowledge biological hierarchies among taxonomic classes. In work, we propose a new approach termed hierarchy-guided neural network (HGNN), infuses hierarchical information into network's training guide structure relationships extracted features. We perform extensive experiments on illustrative use-case classifying fish demonstrate that HGNN outperforms models in terms accuracy, especially under scarce data conditions. also observe shows better resilience adversarial occlusions, when some most informative patch regions image are intentionally blocked their effect accuracy studied.