. 2020 Jan 28;10:1631. doi: 10.3389/fphar.2019.01631

Table 4.

Algorithm and training procedure of restricted Boltzmann machine-capsule networks (RBM-CapsNet).

Algorithm: RBM-CapsNet training algorithm, using a mini-batch stochastic gradient descent (SGD) for simplicity.

Input: mini batch feature vector (x);
Number of RBM training epoch (S1);
Number of Capsule training epoch (S2);
Number of dynamic routing iterations (iter).
Output: Length of each capsules (Len).

1: For n=1 to S1 do
2:

h f_l a y e r \leftarrow ϕ 1 (x, θ 1)

………………………………………{RBM1 training}
3: End for
4: For n=1 to S1 do
5:

p c_l a y e r \leftarrow ϕ 2 (h f_l a y e r, θ 2)

…………………………………{RBM2 training}
6: End for

7: For n=1 to S2 do
8:

h f_l a y e r \leftarrow ϕ 1 (x, θ 1)

p c_l a y e r \leftarrow ϕ 2 (h 1_l a y e r, θ 2)

10:

u \leftarrow Encapule (p c_l a y e r)

11: For all capsule i in PrimaryCaps layer:

{\hat{u}}_{j | i} \leftarrow W_{i j} u_{i}

…………{contribution computes Eq. 1}
12: For all capsule i in PrimaryCaps layer and capsule j in DigitCaps layer:

b_{i j} \leftarrow 0

13: For m=1 to iter do
14: For all capsule i in PrimaryCaps layer: ……{softmax computes Eq. 2-1}
15: For all capsule j in DigitCaps layer:

s_{j} \leftarrow \sum_{i} c_{i j} {\hat{u}}_{j | i}

………{dynamic computes Eq. 2-2}
16: For all capsule j in DigitCaps layer:

v_{j} \leftarrow s q u a s h (s_{j})

…………{squash computes Eq. 2-3}
17: For all capsule i in PrimaryCaps layer and capsule j in DigitCaps layer:

b_{i j} \leftarrow b_{i j} + {\hat{u}}_{j | i} \cdot v_{j}

18: End for

19:

L e n \leftarrow L e n g t h o f v

20:

L \leftarrow l o s s o f v

………………………………………{loss computes Eq. 3}
21:

W \leftarrow W - \partial L / \partial W

22:

θ 1 \leftarrow θ 1 - \partial L / \partial θ 1

23:

θ 2 \leftarrow θ 2 - \partial L / \partial θ 2

24: End for