Table 5
Effects of reduced label quantity and mixed datasets on mean absolute errors for various labels.
| (a) Model descriptions | |||||
|---|---|---|---|---|---|
| Model type | Learning paradigm | Architecture | Dataset | Available labels (real data) | |
| Legend descriptor | |||||
| CNN 1.00 (real) | Encoder | Supervised | CNN | Real | 100% |
| CNN 0.01 (real) | Encoder | Supervised | CNN | Real | 1% |
| CNN 1.00 (mix) | Encoder | Supervised | CNN | Mixed | 100% |
| CNN 0.01 (mix) | Encoder | Supervised | CNN | Mixed | 1% |
| AE 1.00 (real) | AE | Semi-supervised | CNN-ResNet | Real | 100% |
| AE 0.01 (real) | AE | Semi-supervised | CNN-ResNet | Real | 1% |
| AE 1.00 (mix) | AE | Semi-supervised | CNN-ResNet | Mixed | 100% |
| AE 0.01 (mix) | AE | Semi-supervised | CNN-ResNet | Mixed | 1% |
| (b) Intrinsic labels | |||||
| Teff (K) | [M/H] (dex) | log(g) (dex) | All (−) | ||
| CNN 1.00 (real) | 51.27 ± 0.81 | 0.02109 ± 0.00064 | 0.04102 ± 0.00085 | 0.1837 ± 0.0022 | |
| CNN 0.01 (real) | 154.4 ± 5.8 | 0.049 ± 0.0012 | 0.0712 ± 0.0014 | 0.3954 ± 0.0061 | |
| CNN 1.00 (mix) | 50.21 ± 0.82 | 0.02118 ± 0.00064 | 0.03957 ± 0.00085 | 0.1793 ± 0.0022 | |
| CNN 0.01 (mix) | 108.6 ± 5.6 | 0.04468 ± 0.00081 | 0.05083 ± 0.00098 | 0.2984 ± 0.0051 | |
| AE 1.00 (real) | 50.32 ± 0.83 | 0.02478 ± 0.00056 | 0.03898 ± 0.00089 | 0.1844 ± 0.0022 | |
| AE 0.01 (real) | 100.2 ± 1.7 | 0.082 ± 0.0012 | 0.0848 ± 0.0013 | 0.4439 ± 0.0038 | |
| AE 1.00 (mix) | 49.74 ± 0.77 | 0.02292 ± 0.00055 | 0.04139 ± 0.00087 | 0.1867 ± 0.0021 | |
| AE 0.01 (mix) | 75.2 ± 2.8 | 0.02934 ± 0.00081 | 0.0567 ± 0.0011 | 0.2578 ± 0.0037 | |
| (c) Extrinsic labels | |||||
| Radvel (km s−1) | BERV (km s−1) | Airmass (−) | All (−) | ||
| CNN 1.00 (real) | 1.994 ± 0.058 | 0.1764 ± 0.0018 | 0.01163 ± 0.00022 | 0.07591 ± 0.00098 | |
| CNN 0.01 (real) | 4.4 ± 0.19 | 0.7122 ± 0.0071 | 0.06974 ± 0.00066 | 0.2839 ± 0.0031 | |
| CNN 1.00 (mix) | 2.375 ± 0.061 | 0.1671 ± 0.002 | 0.01069 ± 0.00022 | 0.0815 ± 0.001 | |
| CNN 0.01 (mix) | 3.91 ± 0.14 | 0.787 ± 0.011 | 0.04042 ± 0.00063 | 0.2047 ± 0.0026 | |
| AE 1.00 (real) | 1.94 ± 0.069 | 0.1214 ± 0.0021 | 0.00885 ± 0.00018 | 0.0664 ± 0.0011 | |
| AE 0.01 (real) | 5.19 ± 0.19 | 0.4659 ± 0.0063 | 0.07247 ± 0.00073 | 0.3002 ± 0.0032 | |
| AE 1.00 (mix) | 1.918 ± 0.063 | 0.1697 ± 0.0024 | 0.0103 ± 0.00022 | 0.0709 ± 0.0011 | |
| AE 0.01 (mix) | 3.53 ± 0.15 | 0.3719 ± 0.0062 | 0.05555 ± 0.00058 | 0.221 ± 0.0026 | |
Notes. The first table provides an overview of the model properties. The column titled “All” uses NMAE to summarize the mean absolute error across all normalized labels. The column named “Available labels (real data)” indicates the percentage of HARPS labels in the catalog that were used during the training phase. The notation “a ± b” represents the mean absolute error ± the standard deviation for each respective label and model.
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.