Open Access

Table 2.

Statistics for various CNN trials, with the B16 results in parenthesis where the comparison is relevant.

Trial Training sample size Size of 1 test sample Bias σMAD η ⟨CRPS⟩
Training with 80% of the dataset 393 219
Full test sample 103 306 0.00010 0.00912 0.31 0.00674
(B16) (103 306) (0.00062) (0.01350) (1.34)
Suspect zone (SZ) removed 101 499 0.00004 0.00908 0.31 0.00672
Widest 10% of PDFs removed 91 543 0.00006 0.00848 0.09 0.00606
Widest 20% of PDFs removed 79 897 0.00005 0.00789 0.06 0.00556
Stripe 82 only 3943 −0.00009 0.00727 0.34 0.00574
Stripe 82 with widest 20% of PDFs removed 3131 0.00004 0.00635 0.09 0.00467
Training with 50% of the dataset 250 000 252 500 0.00007 0.00910 0.29 0.00672
Training with 20% of the dataset 99 001 385 970 −0.00001 0.00914 0.30 0.00677
Training with 2% of the dataset 10 100 434 228 −0.00017 0.01433 1.26 0.01009
Training on Stripe 82 15 771
Stripe 82 removed 478 274 0.00194 0.01341 1.15 0.00988
Stripe 82 only 3942 −0.00002 0.00795 0.38 0.00622
Training w/o Stripe 82 486 560
Stripe 82 removed 97 607 0.00000 0.00914 0.33 0.00680
Stripe 82 only 19 714 −0.00077 0.00760 0.41 0.00606

Notes.

The bias, σMAD and fraction of outliers η are defined in Sect. 4.1. The values are averaged over five test samples, except in the cases marked with a ⋆ where there is only one. The CRPS and PDF width are defined in Sect. 4.3. The SZ was identified as a small region of the SDSS with above average bias (see Sect. 6.4). It has been removed by default in all other cases below the “SZ removed” line.

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.