Capsule Routing with Dynamic Programming refers to a routing method in capsule networks where dynamic programming algorithms are used to efficiently compute optimal assignments between lower and higher capsule layers.
Formally, the routing consists of determining weights \( c_{ij} \) that represent the agreement between the output of the lower capsule \( i \) and the higher capsule \( j \).
Dynamic programming optimizes these weights by formulating the routing process as an optimization problem with sequential intermediate steps, enabling an efficient calculation of the best paths and distributions.
This method addresses the high complexity of classical routing-by-agreement and improves the stability as well as convergence of training.
Definition:
“Capsule Routing with Dynamic Programming is an optimization method that uses dynamic programming algorithms to efficiently and robustly determine the assignments between capsule layers, thereby improving the recognition of complex patterns.”
Source:
Wang, Y., Zhang, Y., & Liu, J. (2021). Dynamic Programming Routing for Capsule Networks. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW).