Table 2.
Results from a single round of UDC applied to various incorrect label levels.
| Error rate (%) | Before UDC accuracy (%) | After UDC accuracy (%) | Enhancement (%) |
|---|---|---|---|
| 0 | 98.8 ± 0.1 | 98.8 ± 0.1 | 0.0 ± 0.1 |
| 10 | 97.5 ± 0.2 | 98.6 ± 0.1 | 1.1 ± 0.2 |
| 20 | 96.7 ± 0.3 | 98.5 ± 0.1 | 1.8 ± 0.3 |
| 30 | 94.7 ± 0.8 | 98.2 ± 0.1 | 3.5 ± 0.8 |
| 40 | 92.8 ± 0.5 | 98.3 ± 0.1 | 5.5 ± 0.5 |
| 50 | 87.7 ± 1.9 | 98.1 ± 0.1 | 10.3 ± 1.9 |
| 60 | 74.7 ± 2.6 | 97.0 ± 0.2 | 22.3 ± 2.6 |
| 70 | 47.4 ± 2.5 | 92.2 ± 0.6 | 44.8 ± 2.6 |
| 80 | 15.3 ± 2.0 | 3.7 ± 0.5 | -11.7 ± 2.0 |
| 90 | 2.9 ± 0.6 | 1.2 ± 0.2 | -1.7 ± 0.7 |
The bold font draws attention to whether the 'Before UDC' value or the 'After UDC' value is the greater, for each scenario, so that it is easy for the reader to judge which rows there was an improvement, and which rows there was a decrease.
Incorrect labels are distributed evenly across all four classes: airplanes, boats, motorcycles, and trucks. UDC is robust to label error rates up to 70%, where a 44.8% gain in AI performance is achieved after a single round of UDC.