Fig. 3From: Accurate classification of membrane protein types based on sequence and evolutionary information using deep learningComputation between the PrimaryCaps and ProteinCaps layers.There are 2976 16D capsules (each ui is an 16D vector) in PrimaryCaps. Each \(\hat {u_{j|i}}\) is produced by multiplying ui by a weight matrix Wi,j (16 ×16). Capsule Vj (16D vector and j∈ [1, 8]) in ProteinCaps is produced by a weighted sum over all \(\hat {u_{j|i}}\) and the squashing nonlinear activation function. The parameter ci,j is determined by the iterative dynamic routing processBack to article page