Dynamic Routing is the mechanism in Capsule Networks used to determine the connections between layers — an iterative clustering process where lower-level capsules "vote" for higher-level capsules, and only the consistent votes are allowed to pass signal.

What Is Dynamic Routing?

• Problem: In a face, a "mouth" capsule should only activate the "face" capsule, not the "house" capsule.
• Algorithm:

1. Prediction: Low Capsule $i$ predicts High Capsule $j$. 2. Comparison: Check scalar product (similarity). 3. Update: Increase coupling coefficient $c_{ij}$ if prediction was good. 4. Repeat.

• Effect: Creates a dynamic computational graph specific to the image.

Why It Matters

• Parse Trees: Effectively builds a dynamic parse tree of the image (Eye + Nose + Mouth -> Face).
• Occlusion Handling: Robust to parts being missing or moved, as long as the remaining geometry is consistent.

Dynamic Routing is unsupervised clustering inside a network — grouping features into coherent objects on the fly.