A layered capsule network where lower-level part capsules (rendered as small oriented vectors) send weighted predictions upward to higher-level object capsules, with connecting lines that thicken and brighten over routing iterations as agreeing predictions cluster together and disagreeing ones fade out.
Choose an input pose or scramble the part arrangement, then step or auto-play the routing iterations to watch coupling coefficients shift toward whichever higher-level capsule the lower-level votes agree on, and rotate the viewpoint to see how pose vectors update.
Input pose/scramble selector, routing iteration step/auto-play control, viewpoint rotation, and a reset button to rebuild the capsule layer connections.
In Hinton's original 2017 capsule network paper, just three routing iterations were typically enough for the coupling coefficients between lower and higher-level capsules to settle into a stable agreement.
A layered capsule network where lower-level part capsules (rendered as small oriented vectors) send weighted predictions upward to higher-level object capsules, with connecting lines that thicken and brighten over routing iterations as agreeing predictions cluster together and disagreeing ones fade out.
A layered capsule network where lower-level part capsules (rendered as small oriented vectors) send weighted predictions upward to higher-level object capsules, with connecting lines that thicken and brighten over routing iterations as agreeing predictions cluster together and disagreeing ones fade out.
Choose an input pose or scramble the part arrangement, then step or auto-play the routing iterations to watch coupling coefficients shift toward whichever higher-level capsule the lower-level votes agree on, and rotate the viewpoint to see how pose vectors update.
In Hinton's original 2017 capsule network paper, just three routing iterations were typically enough for the coupling coefficients between lower and higher-level capsules to settle into a stable agreement.