CoDiNet: Path Distribution Modeling With Consistency and Diversity for Dynamic Routing

Dynamic routing networks, aimed at finding the best paths in have achieved significant improvements to neural networks terms of accuracy and efficiency. In this paper, we see dynamic a fresh light, formulating method as mapping from sample space space. From perspective mapping, prevalent methods did not take into account how would be distributed Thus, propose novel method, termed CoDiNet, model relationship between by regularizing distribution with properties consistency diversity. In principle, for self-supervised similar samples should closely space. Moreover, design customizable module, which can strike balance When deployed upon ResNet models, our achieves higher performance effectively reduces average computational cost on four widely used datasets.