Table 6.
The best and mean error rates over 10 runs for the proposed method, resPsoCnn and the best baseline method, sosCNN [14], where a (−)/(+) symbol indicates that resPsoCnn performed better/worse than sosCNN.
| Model | MNIST | MNIST-RD | MNIST-RB | MNIST-BI | MNIST-RD+BI | Rectangles-I |
|---|---|---|---|---|---|---|
| resPsoCnn (best) | 0.31% | 2.67% | 1.70% | 1.74% | 8.76% | 1.19% |
| resPsoCnn (mean) | 0.33% | 3.02% | 1.76% | 1.90% | 9.27% | 1.47% |
| sosCNN (best) [14] | 0.30% | 3.01% | 1.49% | 1.68% | 10.65% | 1.57% |
| sosCNN (mean) [14] | 0.40% | 3.78% | 1.89% | 1.98% | 13.61% | 2.37% |
| error difference (best) | 0.01%(+) | −0.34%(−) | 0.21%(+) | 0.06%(+) | −1.89%(−) | −0.38%(−) |
| error difference (mean) | −0.07%(−) | −0.76%(−) | −0.13%(−) | −0.08%(−) | −4.34%(−) | −0.90%(−) |