Table A.1
Accuracy–efficiency trade-off between the lowest-error and Pareto-selected configurations (see Fig. 2).
| Dataset | Arch | ∆Error ↓ [%] | ∆Time ↑ [%] | Ratio R |
|---|---|---|---|---|
| Primordial | MON | 24.9 | 155 | 6.2 |
| FCNN | 24.0 | 417 | 17.4 | |
| LP | 16.7 | 148 | 8.9 | |
| LNODE | 13.9 | 98 | 7.1 | |
| Primordial | MON | 29.6 | 109 | 3.7 |
| Parametric | FCNN | 37.3 | 265 | 7.1 |
| Cloud | LP | 9.5 | 143 | 15.1 |
| LNODE | 1.7 | 70 | 41.2 | |
| Cloud | MON | 4.6 | 36 | 7.8 |
| Parametric | LP | 5.4 | 411 | 76.1 |
| LNODE | 1.6 | 28 | 17.5 | |
| Averages | 15.4 | 171 | 18.9 | |
Notes. Architectures where both configurations coincide are omitted. Reported values indicate relative percentage changes of the accuracy-only configuration with respect to the Pareto-selected one. The final column reports the ratio R = ∆Time/∆Error, i.e., the percentage increase in inference time required to achieve a 1% reduction in error.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.