> The black hole “photo” was constructed using machine learning. I don’t know how it can reasonably be called a “photo.” And although I’m sure their work was reviewed, knowing the nature of computers and data, I wouldn’t be surprised if there were all sorts of errors in the construction of the model that produced that output. What makes the output correct anyway, the fact that it looks like what we expected?

The images of the EHT data were made using several independent techniques performed by different sub-teams who largely worked independently (e.g., [0] for M87). Most of the imaging techniques used are not machine learning approaches, but are rather standard imaging approaches, some of which have have been used for decades in radio astronomy (e.g., CLEAN). The fact that the images made from these different techniques look similar is what lends support to the idea that the "photo" is a reasonable representation. Sure, some images have been created using machine learning approaches, but other images have been done using well vetted approaches that are not machine learning.

But if one doesn't want to trust the imaging techniques, the size of the black hole shadow can be estimated directly from the radio telescope measurements. These are the "visibilities" (essentially measurements of components of the Fourier Transform of the distribution of radio emission on the sky) and the amplitudes of the visibilities (e.g., figure 2 of [1]) indicates that the brightness distribution of the source is approximately a thin ring, with the ring diameter setting the baseline length at which the amplitudes have their minimum.

So the basic result about the black hole shadow can be made from the visibilities themselves, without relying on the imaging. But the images of course make it easier to see some of the complexity (e.g., that the ring around M87 is not axisymmetric).

[0] https://ui.adsabs.harvard.edu/abs/2019ApJ...875L...4E/abstra...

[1] https://ui.adsabs.harvard.edu/abs/2019ApJ...875L...1E/abstra...