Its not hard to find that there are legitimate academic criticism of this ‘photo’. For example here. The comparison you made is not correct, more like I gave a blurry photo to an AI trained on paintings of Donald Trump and asked it to make an image of him. Even if the original image was not of Trump, the chances are the output will be because that’s all the model was trained on.
This is the trouble with using this as ‘proof’ that the. Theory and the simulations are correct, because while that is still likely, there is a feedback loop causing confirmation bias here, especially when people refer to this image as a ‘photo’.
I think, at best, it shows that the observations are consistent with the model, or to take it back to the blurry low light photo… The photo wasn’t obviously not Trump.
I remember reading the original paper at the time and thinking, if I had been a reviewer I’d have wanted clear acknowledgement of the confirmation bias danger in the methodology. Ideally some sort of quantification of risk. It just seemed like too large a flaw to just be glossed over.
This is one team that disagrees out of many that agree.
To explain what you are seeing. The above image is the inverse Fourier transform (FT) of different frequencies of sinus waves that compose an image.
The very large baseline interferometer (VLBI) applied in the event horizon telescope (EHT) is using different telescopes all over the world, in a technique called interferometry, to achieve high enough resolutions to observe different frequencies in Fourier space that make up an image. If you observe all, you can recreate the full image perfectly. They did not, they observed for a long time and thus got a hefty amount of these “spatial” frequencies. Then they use techniques that limit the image to physical reality (e.g. no negative intensities/fluxes) and clean it from artefacts. Then transform it to image space (via the inverse FT)
Thereby, they get an actual image that approximates reality. There is no AI used at all. The researchers from Japan argued for different approach to the data, getting a slightly different inclination in that image. This may well be as the data is still too few to 100 % determine the shape, but looks more to me like they chose very different assumptions (which many other researchers do not agree with).
Edit: They did use ML for simulations to compare their sampling of the Fourier space to.
Most of what you said is correct but there is a final step you are missing, the image is not entirely constructed from raw data. The interferometry data is sparse and the ‘gaps’ are filled with mathematical solutions from theoretical models, and using statistical models trained on simulation data.
We recently developed PRIMO (Principal-component
Interferometric Modeling; Medeiros et al. 2023a) for in-
terferometric image reconstruction and used it to obtain
a high-fidelity image of the M87 black hole from the 2017
EHT data (Medeiros et al. 2023b). In this approach, we
decompose the image into a set of eigenimages, which
the algorithm “learned” using a very large suite of black-
hole images obtained from general relativistic magneto-
hydrodynamic (GRMHD) simulations
Thanks for sharing that paper. I was indeed missing that information and now agree with your earlier statement.
I think them using magnetohydrodynamical black hole models as a base for the ML is a better approach than standard CLEAN though that the Japanese team used. However, both “only” approach reality.
You’re welcome. I think calling it the output of an ‘AI model’ triggers thoughts of the current generative image models, i.e. entirely fictional which is not accurate, but it is important to recognise the difference between an image and a photo.
I also by no means want to downplay the achievement that the image represents, it’s an amazing result and deserves the praise. Defending criticism and confirming conclusions will always be vital parts of the scientific method.
True, ML and such fell under the umbrella term of AI before, but I feel that with most people using it mostly for LLMs (or things like diffusion models, etc.) right now, it has kinda lost that meaning to some extent…
Having triggered this conversation off, I’ll just congratulate you both on a quality discussion. I’ll admit I used loose terminology in my original post, but that was mainly to get my point across to a general audience. The specificity you both went to is laudable.
Its not hard to find that there are legitimate academic criticism of this ‘photo’. For example here. The comparison you made is not correct, more like I gave a blurry photo to an AI trained on paintings of Donald Trump and asked it to make an image of him. Even if the original image was not of Trump, the chances are the output will be because that’s all the model was trained on.
This is the trouble with using this as ‘proof’ that the. Theory and the simulations are correct, because while that is still likely, there is a feedback loop causing confirmation bias here, especially when people refer to this image as a ‘photo’.
I think, at best, it shows that the observations are consistent with the model, or to take it back to the blurry low light photo… The photo wasn’t obviously not Trump.
I remember reading the original paper at the time and thinking, if I had been a reviewer I’d have wanted clear acknowledgement of the confirmation bias danger in the methodology. Ideally some sort of quantification of risk. It just seemed like too large a flaw to just be glossed over.
This is one team that disagrees out of many that agree.
To explain what you are seeing. The above image is the inverse Fourier transform (FT) of different frequencies of sinus waves that compose an image.
The very large baseline interferometer (VLBI) applied in the event horizon telescope (EHT) is using different telescopes all over the world, in a technique called interferometry, to achieve high enough resolutions to observe different frequencies in Fourier space that make up an image. If you observe all, you can recreate the full image perfectly. They did not, they observed for a long time and thus got a hefty amount of these “spatial” frequencies. Then they use techniques that limit the image to physical reality (e.g. no negative intensities/fluxes) and clean it from artefacts. Then transform it to image space (via the inverse FT)
Thereby, they get an actual image that approximates reality. There is no AI used at all. The researchers from Japan argued for different approach to the data, getting a slightly different inclination in that image. This may well be as the data is still too few to 100 % determine the shape, but looks more to me like they chose very different assumptions (which many other researchers do not agree with).
Edit: They did use ML for simulations to compare their sampling of the Fourier space to.
Most of what you said is correct but there is a final step you are missing, the image is not entirely constructed from raw data. The interferometry data is sparse and the ‘gaps’ are filled with mathematical solutions from theoretical models, and using statistical models trained on simulation data.
Paper: https://arxiv.org/pdf/2408.10322
Thanks for sharing that paper. I was indeed missing that information and now agree with your earlier statement.
I think them using magnetohydrodynamical black hole models as a base for the ML is a better approach than standard CLEAN though that the Japanese team used. However, both “only” approach reality.
You’re welcome. I think calling it the output of an ‘AI model’ triggers thoughts of the current generative image models, i.e. entirely fictional which is not accurate, but it is important to recognise the difference between an image and a photo.
I also by no means want to downplay the achievement that the image represents, it’s an amazing result and deserves the praise. Defending criticism and confirming conclusions will always be vital parts of the scientific method.
True, ML and such fell under the umbrella term of AI before, but I feel that with most people using it mostly for LLMs (or things like diffusion models, etc.) right now, it has kinda lost that meaning to some extent…
@Tamo240@programming.dev and yourself.
Having triggered this conversation off, I’ll just congratulate you both on a quality discussion. I’ll admit I used loose terminology in my original post, but that was mainly to get my point across to a general audience. The specificity you both went to is laudable.