Table 3
Correction performance dependence on training data size.
| Ntrain | Order 54 | Order 60 | Order 64 |
|---|---|---|---|
| (10−3 RMSE) | (10−3 RMSE) | (10−3 RMSE) | |
| 838 | 4.0 ± 0.5 | 5.0 ± 0.6 | 6.7 ± 1.4 |
| 400 | 4.0 ± 0.5 | 5.0 ± 0.7 | 6.9 ± 1.5 |
| 200 | 4.0 ± 0.6 | 5.2 ± 0.9 | 7.0 ± 1.8 |
Notes. Corrections are performed with tellurics extracted from a network trained either on the full 838 observations, or with tellurics extracted from networks trained on random samples of the training observations. We retrain the network for random samples of size either 400 or 200 observations. We perform the random sampling 5 times for both cases. Performance is shown in terms of the RMSE of the residual from corrections made on 400 random observations between 2015 and 2018 in the same spectral ranges shown in Table 2. The performance for “838” is the mean and standard deviation of the 400 corrections, while the performance for “400” and “200” is the mean and standard deviation for 2000 corrections each (5 × 400). Variance increases and performance decreases slightly for less training data.
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.