Table 2.
Comparison of the number of parameters generated by each model. The VMG-Routing model produced the least number of parameters with the ML producing the largest number of parameters. (∗) indicates the models that our device could not implement due to memory limitations. The values reported here were thus obtained from the literature. (−) indicates unavailable values. (#c) represents the number of classes in the dataset.
| Algorithm | Number of parameters | |||
|---|---|---|---|---|
| CIFAR-10 | Fashion-MNIST | MNIST | COVID-19 radiography | |
| VB-routing {64, 8, 16, 16, #c} | 145 K | 145 K | 145 K | 120 K |
| VB-routing∗ {64, 16, 32, 32, #c} | 323 K | 323 K | — | — |
| VB-routing∗ {64, 16, 16, 16, #c} | 172 K | 172 K | — | — |
| EM-routing∗ {64, 16, 16, 16, #c} | — | 323 K | — | — |
| EM-routing∗ {64, 16, 32, 32, #c} | 323 K | — | — | — |
| Multi-lane LBP-gabor capsule | 4.10 M | 4.10 M | 4.10 M | 3.70 M |
| Dynamic routing | 9.3 M | 8.2 M | 8.2 M | 9.8 M |
| VMG-routing {32, 4, 8, 8, #c} (ours) | 14 K | 15.5 K | 14 K | 10.2 K |