Table 2.
Training stage |
Pretext task |
Downstream ImageNet classification |
||
---|---|---|---|---|
Threshold | Training loss ↓ | Training loss ↓ | Test loss ↓ | Test accuracy (%) ↑ |
(0.8, 12) | 3.92 ± 0.003 | 4.62 ± 0.001 | 4.67 ± 0.004 | 18.39 ± 0.082 |
(0.4, 6) | 3.77 ± 0.004 | 4.98 ± 0.326 | 4.67 ± 0.024 | 18.03 ± 0.114 |
(1.6, 24) | 4.26 ± 0.003 | 4.72 ± 0.011 | 4.82 ± 0.013 | 17.61 ± 0.135 |