Abstract
In this work, we present a new residual network family based on a tree structure. We present three types of tree modules that can be employed in different kinds of convolutional networks as a replacement for one or more convolutional layers. The new architecture exposes two new hyper-parameters - tree height and branching factor - that grant more control on the model size and codependency between the maps. Tree modules provide the flexibility to merge two important techniques: branching technique, which provides better feature representation, and group convolution, which reduces the number of parameters. Most previous studies focused on accuracy regardless of the model complexity. However, we focus on information density metric to design a model that effectively utilizes its parametric space. We conducted numerous experiments on the dataset of the Canadian Institute for Advanced Research-10 (CIFAR-10) and demonstrated that the proposed networks are superior to many famous networks in terms of information density and accuracy. (C) 2019 Elsevier B.V. All rights reserved.