Extended Data Fig. 1: The schematic diagram of the dynamic routing algorithm.
The PrimaryCaps layer outputs 32 × 73 = 2336 vectors ui(i = 1, . . . , 2336), which are then transformed into \({\hat{{{{\bf{u}}}}}}_{j| i}\)(j = 1, 2) by premultiplying pose matrices Wij. Note that pose matrices are learned during training. The right half of this figure shows the iterative dynamic routing procedures. The coupling coefficients cij indicate the probability that primary capsule i agrees with phenotype capsule j. The input sj to one of the phenotype capsules is subsequently processed by using the squashing operation. The output vj is used to update the bij and cij until the model converges.
