DataTerminology.comNEW
An iterative routing-by-agreement mechanism in capsule networks that assigns lower-level outputs to higher-level capsules based on agreement.
DESCRIPTION
Dynamic routing is a core algorithm for capsule networks that updates coupling coefficients between capsules through an iterative process. The algorithm measures agreement between predicted outputs and actual capsule outputs to strengthen or weaken connections. This improves part-whole relationships and viewpoint invariance compared to max-pooling in CNNs. Variants include EM routing and routing via attention-like mechanisms. While influential, capsule networks are less common in production than transformers and CNNs.
ALTERNATIVE DEFINITIONS
arXiv
We introduce a routing-by-agreement mechanism that ensures that a capsule sends its output to appropriate parent capsules by iteratively updating coupling coefficients.
Defines routing-by-agreement and the iterative update of coupling coefficients.
NeurIPS Proceedings
Routing-by-agreement increases the coupling between two capsules when the prediction vector of the lower-level capsule agrees with the output of the higher-level capsule.
Focuses on the agreement principle driving coupling updates.
Wikipedia
Dynamic routing is an algorithm used in capsule neural networks to iteratively determine how much lower-level capsules contribute to higher-level capsules.
Summarizes dynamic routing’s role within capsule architectures.
RELATED TERMS
Capsule Networks Routing-by-Agreement Matrix Capsules Convolutional Neural Network Attention Mechanism Representation Learning Coupling Coefficients Hinton Sabour
COMPETING TERMS
EM Routing Self-Attention
RELATIONS
- Broader: Routing
KNOWLEDGE AREAS
TARGET AUDIENCE
EXTERNAL SOURCES
- Wikipedia - Capsule neural network
- arXiv - Original capsule network paper
- NeurIPS - Peer-reviewed proceedings PDF
TOP QUESTIONS
What problem does Dynamic Routing solve in capsule networks?
It assigns parts to wholes by strengthening connections where predictions agree, improving hierarchical representation over max-pooling.
Category: Topic
How does Dynamic Routing work?
It iteratively updates coupling coefficients based on agreement between predicted and actual capsule outputs.
Category: HowTo
What are alternatives to Dynamic Routing?
EM routing and attention-like mechanisms provide different routing strategies with varying computational and performance trade-offs.
Category: Alternatives
Is Dynamic Routing widely used in production?
Capsule networks are less common today; transformers and CNNs dominate production. Dynamic routing remains an influential research idea.
Category: Opinion
What are common pitfalls when experimenting with Dynamic Routing?
High computational cost, tuning iteration counts, and instability if agreement metrics are noisy.
Category: Mistakes
How does Dynamic Routing relate to attention?
Both allocate contributions among units, but routing uses agreement metrics iteratively; attention typically uses learned weights via softmax.
Category: Linkages
What datasets were used to showcase Dynamic Routing?
MNIST and variants demonstrated improved viewpoint invariance and part-whole modeling.
Category: Example
What is the risk of overfitting with Dynamic Routing?
Excessively tuned routing iterations can overfit source tasks, degrading generalization.
Category: Risk
Who introduced Dynamic Routing?
Sabour, Frosst, and Hinton introduced routing-by-agreement for capsules in 2017.
Category: Origin