Open Access
Table 2
ZINGULARITY training and application parameters. For Sgr A*, we found two equally viable models with different Nep.
| Category | Param. | Value | Description | |
|---|---|---|---|---|
| Common ZINGULARITY GRMHD -GRRT EHT parameters | f | Swish (Ramachandran et al. 2017) | Activation function for all hidden layers | |
| Ξ | RMSProp | Optimization algorithm | ||
| L | Negative log-likelihood | Loss function | ||
| ηval | 0.1 | Fraction of  used for validation |
||
| Nb | 256 | Training batch size | ||
| lr | 0.001 × n/Nep 0.0001 × (1 + cos(nπ/Nep))/2 | Learning rate warm-up for 0 ≤ n ≤ 0.1 x Nep Learning rate cosine decay for 0.1 × Nep ≤ n ≤ Nep | ||
| M87* | Sgr A* | |||
| Fiducial models | Nvis | 8 × 5489 | 8 × 13 840 | Number of data points in a training sample |
| Ntr | 600000 | 252 000 | Number of training samples | |
| Nep | 70 | 50, 60 | Number of training epochs | |
| ηdrop | 0.01 | 0 | Dropout rate for stochastic neuron deactivation | |
| ℒ1 | 0.01 | 0.01 | L1 (lasso) regularization hyperparameter | |
| ℒ2 | 0.01 | 0.01 | L2 (ridge) regularization hyperparameter | |
| kconv | 8 | 8 | Receptive field of CNN layers | |
| nCNNb | 16 | 8 | Baseline number of neurons for the ResNet CNN layers | |
| nCNNl | 128 | 2048 | Neurons in last ResNet CNN layer | |
| Ndense | 15 | 12 | Number of post-ResNet dense variational layers | |
| ndense | 128 | 1024 | Neurons in post-ResNet dense variational layers | |
| Nfree | 1376806 | 135 068 877 | Number of free parameters in the network | |
| Boot- strapping errors for the EHT | 𝒟 | 1–3% (EHT et al. 2021a) | Polarization leakage (𝒟-terms) | |
| 𝒢planet | 10% (Janssen et al. 2019a) | Primary calibrator model uncertainties | ||
| 𝒢scatter | 5–35% (Janssen et al. 2019a) | DPFU uncertainty due to measurement scatter | ||
| gcB | 3.6–10.4% (Janssen et al. 2019a) | Measurement error on gain curve curvature | ||
| gcE0 | 1–2% (Janssen et al. 2019a) | Measurement error on gain curve peak elevation | ||
| σth |  |
Thermal noise of EHT data used in this work | ||
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.