Title: Efficient Inference and Update via WordNet-Embedded Network Tree
Abstract: Deeper and deeper convolutional neural network achieve very precise accuracy. For scenarios with very large amount of classes, the very deep structures benefit from their huge parameter space, while sacrifices computation efficiency. However, in some scenarios, the number of target classes is relatively small, the deep structure is no longer economic. Thus, we propose a novel tree-structure convolutional neural network, called Network Tree. The architecture consists of Resblocks as building blocks, and is organized according to the hierarchal structure of WordNet, which is much shallower and describable. Based on this architecture, we further propose the associated efficient inference and update algorithms. In the inference phase, using SVM-based predictors to prune sub-trees dynamically, our algorithm can reduce computation load substantially. In the update phase, according to tree structure's class-specific branches, update along branch can reduce impact on other branches. Extensive experiments on CIFAR-10 and VID-2017 demonstrate that the Network Tree is able to gain 42% acceleration, at expense of slight accuracy reduction compare to ResNet-152.
Publication Year: 2018
Publication Date: 2018-10-01
Language: en
Type: article
Indexed In: ['crossref']
Access and Citation
AI Researcher Chatbot
Get quick answers to your questions about the article from our AI researcher chatbot