Table 1
Performance of various training data sets.
| CNN | ResNet | ||||||
|---|---|---|---|---|---|---|---|
| Data set | Summary | AUROC | TPR0 | TPR10 | AUROC | TPR0 | TPR10 |
| Baseline L1 + N1 | Bright arcs with μ ≥ 5, uniform θE over [0.75″–2.5″] 33% spirals, 27% LRGs, 6% compact groups, 33% random | 0.9557 | 0.0 | 44.4 | 0.9913 | 36.0 | 55.0 |
| N1 +… | |||||||
| L2 | Uniform θE over [1.0"–2.5"] | 0.9527 | 7.4 | 41.8 | 0.9860 | 26.5 | 41.3 |
| L3 | Uniform θE over [0.75"–2.0"] | 0.9581 | 0.0 | 49.8 | 0.9728 | 27.5 | 47.1 |
| L4 | Natural θE distribution | 0.9919 | 0.0 | 43.4 | 0.9941 | 42.9 | 70.9 |
| L5 | Decrease lensed arc brightness thresholds | 0.9686 | 8.5 | 49.7 | 0.9502 | 10.1 | 23.8 |
| L6 | No lower limit on μ, balanced number of doubles/quads | 0.9891 | 0.0 | 0.0 | 0.9948 | 0.0 | 0.0 |
| L7 | No boosting of fraction of red sources | 0.9624 | 0.0 | 50.3 | 0.9870 | 49.2 | 70.4 |
| L8 | Different lens/source pairs | 0.9531 | 0.0 | 50.3 | 0.9799 | 38.6 | 51.9 |
| L1 + … | |||||||
| N2 | Only random galaxies, no r-band magnitude cut | 0.9831 | 0.0 | 0.0 | … | … | … |
| N3 | Only spiral galaxies | 0.6141 | 0.0 | 0.0 | … | … | … |
| N4 | Only spiral galaxies and LRGs | 0.8984 | 0.0 | 20.6 | … | … | … |
| N5 | Classes from N1 plus 5000 FPs from previous searches | 0.9610 | 32.8 | 47.6 | … | … | … |
| N6 | Same as N5 with balanced number of objects per class | 0.9704 | 31.2 | 50.8 | 0.9762 | 19.1 | 37.0 |
| N7 | Same as N1 with two times more galaxies per class | 0.9763 | 23.8 | 57.7 | 0.9929 | 30.2 | 42.3 |
Notes. The second column gives a summary of the properties of baseline sets L1 and N1, as well as other ground-truth data sets considered in this section. Sets L2–L8 are simulated following the same method as set L1, apart from the modifications listed in the second column. See Section 5.2 for more details on the construction of the data sets. Architectures were kept fixed to the baseline CNN and ResNet. The ResNets trained on sets L1+N2, L1+N3, L1+N4, and L1+N5 are discarded since none of these networks reached acceptable performance (AUROC ≳ 0.8).
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.