EADR: Entropy Adjusted Dynamic Routing Capsule Networks

Reza Babaei, Huong Pham, Samuel Cheng, Shangqing Zhao

Abstract: Capsule networks offer a useful approach for modeling part-whole hierarchies in visual data, yet they remain limited in effectively handling uncertainty in part-object relationships. This work introduces an entropy-adjusted dynamic routing (EADR) algorithm that leverages information-theoretic principles to enhance both the performance and interpretability of capsule networks. By incorporating an entropy-based regularization term into the final iteration of the dynamic routing process, our approach refines routing decisions, reducing reliance on uncertain capsule connections. Experimental evaluations on the CIFAR-10 dataset demonstrate that our method achieves a mean accuracy of 85.73%, surpassing the baseline accuracy of 84.22% achieved by standard dynamic routing. Comparative analysis with recent entropy-based routing methods highlights our approach’s balance of computational efficiency and routing flexibility, achieved without much additional model complexity. These results suggest that EADR routing can be a useful tool for enhancing both the interpretability and efficacy of capsule networks in uncertain environments.