Table 5.
Performances of the proposed model before and after using sparse learning.
| # of Classes | Sparse Learing only FC Layer |
Proposed DepthConv-LSTM Accuracy (%) |
Proposed DepthConv-GRU Accuracy (%) |
Node Selection Method |
|---|---|---|---|---|
| 10 | Initial Network | 97.86 | 98.72 | - |
| 10 | 80% re-train, 20% Freeze | 98.72 | 98.72 | Ranking Mg. of nodes |
| 10 | 60% re-train, 40% Freeze | 98.08 | 98.5 | Ranking Mg. of nodes |
| 7 | Initial Network | 99.11 | 99.4 | - |
| 7+3 | 80% re-train, 20% Freeze | 97.86 | 99.15 | Ranking Mg. of nodes |
| 7+3 | 60% re-train, 40% Freeze | 98.29 | 99.15 | Ranking Mg. of nodes |
| 4 | Initial Network | 99.51 | 99.51 | - |
| 4+6 | 80% re-train, 20% Freeze | 96.37 | 98.08 | Ranking Mg. of nodes |
| 4+6 | 60% re-train, 40% Freeze | 97.86 | 97.44 | Ranking Mg. of nodes |
| 4+6 | re_train 0=<Av<Eq_value | 96.79 | 98.08 | Avg Activation Method |