Table 9.
The test accuracies of the MobileNet-v2 trained by the optimization methods for the CIFAR-10 image dataset classification task. “HyAdamC-Basic −Other Method” (or “HyAdamC-Scale− Other Method”) indicates a difference between the test accuracies of HyAdamC-Basic (or HyAdamC-Scale) and the compared method. “W (Win)”, “T (Tie)”, and “L (Loss)” refer to the number of the compared methods for which HyAdamC-Basic (or HyAdamC-Scale) achieved better, equivalent, and worse test accuracies, respectively.
| Methods | Test Accuracy | HyAdamC-Basic − Other Method |
HyAdamC-Scale − Other Method |
|---|---|---|---|
| SGD | 0.8 | 0.116 | 0.118 |
| RMSProp | 0.915 | 0.001 | 0.003 |
| Adam | 0.92 | −0.004 | −0.002 |
| AdamW | 0.914 | 0.002 | 0.004 |
| Adagrad | 0.885 | 0.031 | 0.033 |
| AdaDelta | 0.92 | −0.004 | −0.002 |
| Rprop | 0.341 | 0.575 | 0.577 |
| Yogi | 0.921 | −0.005 | −0.003 |
| Fromage | 0.891 | 0.025 | 0.027 |
| TAdam | 0.918 | −0.002 | 0 |
| diffGrad | 0.911 | 0.005 | 0.007 |
| HyAdamC-Basic | 0.916 | - | - |
| HyAdamC-Scale | 0.918 | - | - |
| Win/Tie/Lose (HyAdamC-Basic) | 7/0/4 | ||
| Win/Tie/Lose (HyAdamC-Scale) | 7/1/3 | ||