Taxonomy-Regularized Semantic Deep Convolutional Neural Networks

Abstract

We propose a novel convolutional network architecture that abstracts and dierentiates the categories based on a given class hier- archy. We exploit grouped and discriminative information provided by the taxonomy, by focusing on the general and specic components that comprise each category, through the min- and dierence-pooling operations. Without using any additional parameters or substantial increase in time complexity, our model is able to learn the features that are discriminative for classifying often confused sub-classes belonging to the same superclass, and thus improve the overall classication performance. We validate our method on CIFAR-100, Places-205, and ImageNet Animal datasets, on which our model obtains signicant improvements over the base convolutional networks.