====== Ndung'u 2023: Advances on the morphological classification of radio galaxies ======
https://doi.org/10.1016/j.newar.2023.101685

  - SKA will generate datasets on the scale of exabytes. SKA-LOW data rate 10 TB/s, SKA-MID 19 TB/s.
  - MeerKAT raw data rate 2.2 TB/s. MWA 300 GB/s, LOFAR 13 TB/s.
  - EMU will map 70M radio sources, SKA 500M.
  - Need for classification: science, serendipitous discovery (Ray 2016).
  - Yatawatta 2021 gives a smart calibration package based on deep reinforcement learning.
  - LOFAR sensitivity 100 $\mu$Jy, resolution $6''$.
  - FR0 are 5 times more numerous than FR-I-II combined. The gradient in Fig.3 is nice.
  - Is it right that WAT-NAT come from both FR I-II???
  - IVOA (int virt obs alliance) gives FAIR (findable accessible interoperable reusable) data.
  - Besides classification, need to focus on source extraction and anomaly detection. Create AI-alternatives to PyBDSF.
  - Table 2 dos not have a dataset created from Sasmal 23.
  - Most works depend on deep and shallow CNN (ConvNet, mimicks human visual cortex). The deeper (more layers) the better.
  - Regularization techniques (dropout of weakly connected neurons) used to solve overfitting.
  - Bowles 21 uses attention-gate layers to suppress irrelevant info.
  - Tang 22 uses multidomain multibranch CNN to take multiple inputs.
  - Brand 23 aligned the PC of all galaxies with the axis of the coordinate system via PCA.
  - Sadeghi 21 uses Zernike polynomials (SVM) to extract image moments. Ntwaetsile 21 uses Haralick features for texture.
  - Tang 19 successfully used FIRST-trained model on NVSS, but not the reverse.
  - Wang 21 uses SKnet module for attention. Zhang 22 uses SE Net.
  - Tiramisu (Pino 21) works best for source extraction, detection, localization and then classification.
  - Main drawback of AI is the lack of explainability. 
  - Another drawback is that we are reducing 4D data cubes into 2D.