“PRIMO is a new approach to the difficult task of constructing images from EHT observations.”
See full article...
See full article...
Iconic image of M87 black hole just got a machine learning makeover
- Thread starter JournalBot
- Start date
It's cool, though I gotta say that the Primo version looks flipped (left right) compared to the original
Upvote
32
(32
/
0)
So using models to generate training data to clean up an image to test the models. Seems circular to me.
Upvote
137
(139
/
-2)
I'm just an idiot with a keyboard, but "we trained the AI on a bunch of simulated images; look how much the output looks like what we predicted" seems to be heading towards circular logic?
I tried to get an accretion disc joke in there. I failed.
edit: aaand ninja'd
I tried to get an accretion disc joke in there. I failed.
edit: aaand ninja'd
Upvote
147
(150
/
-3)
... consistent with the set of models we fed into the AI training which we hope have some relation to the physics of the actual black hole.
Upvote
36
(38
/
-2)
It's not working in Chrome either. It seems to be a problem with the video file itself.
I found this video on Youtube that has the same description of "Overview of simulations that were generated for the training set of the PRIMO algorithm. "
View: https://www.youtube.com/watch?v=k53nOYAApE0
Upvote
22
(22
/
0)
Upvote
35
(35
/
0)
Without more info on the details it is hard to say, but it feels like a bad idea. The training set are simulations, so the network is a priori biased on what we expect to see and not what is actually observed.
it embodies The definition of “ confirmation bias”
it embodies The definition of “ confirmation bias”
Upvote
79
(81
/
-2)
It kind of does, yes.
Modelling is always part of physics, from the very smallest bits of subatomic debris we find in particle accelerators to the very largest cosmological situations.
The usual, and common, approach is to say "IF this particular model is correct, THEN the observations should be as follows." Then you compare the actual observations to the predictions made by many different variants of the model. There's a lot of back-and-forth iteration here, but the general idea is that the correct model and the actual observations should eventually converge on agreement, while the incorrect models will increasingly diverge from observations as you whittle down the uncertainties.
Your observations, though, also depend on other modelled results. If, for example, you are using a gas-filled ionization chamber to measure gamma ray intensity, then most of what you know about the detailed behaviour of that specific ionization chamber and the matter surrounding it comes not from direct measurements, but from Monte Carlo simulations in the EGSnrc program. EGSnrc is very well tested and validated, but you still have to trust its modelled outputs in order to obtain a valid characterization for your physical detector. Most other detectors similarly depend on modelling and simulation in other programs. The original EHT image discussed here, for example, relied on a whole lot of modelling and calculation to piece together very sparse data, and some features in that image — the ring width, for example — depend rather significantly on terms such as the strength of the regularizers in the regularized maximum likelihood algorithms. Parts of the reconstruction are actually mathematically unbounded unless you fill in this data somehow.
Now we look at using ML to "fill in the gaps" of incomplete data. This is, I think, a riskier stretch, because the data that gets made up depends on the model not only of the detection hardware, but also of the thing being observed. You could do it well — iterating on the model of the physical object, and therefore on the made-up data, until you get the best possible agreement with the actual observed data. Or you could do it badly — iterating on the made-up data until the made-up data and the observed data combine to make the best fit to the model.
The average Joe on the street won't be able to tell those two approaches, or anything in between, apart. And I'd hazard that many journal editors and reviewers would also have a hard time verifying exactly which approach is at play.
To their credit, Medeiros et al seem to have made a decent attempt at sussing out which characteristics of their reconstruction are robust and consistent with actual observed data over a wide range of model parameters (and are therefore probably real) versus which ones vary from model run to model run but still remain consistent with actual observed data (and therefore could be artefacts of the ML and modelling process).
Last edited:
Upvote
114
(115
/
-1)
Seems like bad science to use a bunch of simulated images to alter a real world image of something we know exceedingly little about.
Upvote
14
(19
/
-5)
Same thought.
Adding to that: PRIMO didn't train on real images. PRIMO trained on existing simulations and then attempted to create one that matched the observational data.
This feels like an "artists conception" here. I fully realize that it's possible to compute data that's present but functionally obscured (see: reading license plates off of blurry video); but to this layman, that video looks speculative.
As a "which model(s) match this image", it seems useful/interesting; but it feels less like I'm looking at an actual image of the black hole and more like I'm looking at a simulation of a black hole matching the specifics of this one that we could glean".
Upvote
17
(19
/
-2)
They did use observed data.
The model is sampled from the observations - blurring (to reduced detail), then PCA.
MCMC(Markov Chain Monte Carlo) and sampling the posterior distribution is essential to show they didn't "add" anything that wasn't already there.
It's a model of physically observed components (dictionary lookup) that are combined into images that are consistent with the original observations.
Upvote
-5
(7
/
-12)
Honest question- Why does this look so round? are we perfectly aligned with it or would the "ring" around a black hole appear round no matter where one observes it from?
Upvote
15
(15
/
0)
Upvote
9
(9
/
0)
I believe it's a bit of both. The original analyses attempted to determine the angle of the disk and it was in the 80's of degrees as I recall. And the EH of any BH will appear to have a ring very near the body but the accretion disk is typically much further out than that.
Upvote
12
(12
/
0)
Yes, anything near the BH is strongly lensed. The frightening bit is that the photons leaving the vicinity of the EH (those that barely escape) might have made several trips (or several thousand) around the BH before escaping. That's what a straight line in spacetime looks like near a BH.
Upvote
28
(28
/
0)
Sounds like you went around it a few times but eventually fell flat.
Upvote
51
(51
/
0)
"The EHT captured photons trapped in orbit around the black hole, swirling around at near the speed of light, creating a bright ring around it."
Photons travel at the speed of light, not near it. I am guessing this should be photons emitted from other particles (fermions?) traveling at near the speed of light. Also if the photons were trapped in orbit, they wouldn't be hitting our detectors.
Photons travel at the speed of light, not near it. I am guessing this should be photons emitted from other particles (fermions?) traveling at near the speed of light. Also if the photons were trapped in orbit, they wouldn't be hitting our detectors.
Upvote
29
(29
/
0)
It does, yes.
I wonder if this could explain it:
I suppose the bright spots in the central region might differ from the bright spots in the surrounding gas, if that's what the quoted sentence means – but I see no particular reason why it would be mirrored.
Upvote
2
(2
/
0)
How does monte carlo simulation (effectively, building a probability distribution of results for the model) do anything fundamentally different to "AI"? In both cases you make an inference from the model and the inputs. It's not a comparison to a complete observation, because they don't have one.
Upvote
7
(8
/
-1)
As a scientist, I find this largely unvalidated application of "AI" to massage hard data extremely troubling. Suffice it to say I won't be making a similar mistake any time soon.
Upvote
11
(17
/
-6)
The difference is that we know exactly what a monte carlo simulation does and how it gets to its result. That cannot generally be said of an "AI-based" simulation.
Upvote
16
(18
/
-2)
Upvote
16
(16
/
0)
Also those are the slowest moving photons I've ever heard of - shouldn't they be traveling at the speed of light?
-d
Upvote
-2
(2
/
-4)
I would like the uncertainties in this technique to be described (and enumerated) more thoroughly.
Upvote
3
(3
/
0)
What they're describing there is the photon sphere, but the actual bright part in the image is the accretion disk with a gravitational lensing effect. I'm pretty sure you can't see the photon sphere on this picture?
Upvote
3
(3
/
0)
I'd like someone who's familiar with the EHT process to explain how far apart from a "picture" this really is. Don't get me wrong, I think it's super cool and amazing, but looking at the amount of computation required to get this out of the data, I'm given the (perhaps false) impression that this is sort of like reconstructing a picture of a face by capturing a few pixels of the lips, a few pixels on the nose, and a few pixels of the hair. If that's the case then there would be a lot of guessing as to what the final face looks like. And we guess that with these color combinations of lips, nose and hair the face should looks "sort of like this", i.e. a general idea of what it could look like but it's more like an artist's interpretation than a "photo".
Or is it more like they have a very very faint picture of a lot of sample areas of the face, and they have a very very blurry picture of the face, but they know the "unblurring" algorithm so well that they can state with 99% certainty that the face must look like this?
Or is it more like they have a very very faint picture of a lot of sample areas of the face, and they have a very very blurry picture of the face, but they know the "unblurring" algorithm so well that they can state with 99% certainty that the face must look like this?
Upvote
6
(6
/
0)
It does for me with default settings. Do you have an adblocker enabled?
Upvote
0
(1
/
-1)
Interesting, but what they've essentially done here is create a new model. Worthwhile to be tested against future observations, but not a replacement for those observations. The thing about all our "enhance" methods is they are done via guesswork. In this case, very sophisticated guesswork that may even turn out to be accurate, but as they say a drop of evidence is worth an ocean of speculation.
Upvote
3
(5
/
-2)
As someone who went several years with uncorrected myopia as a child I can say that reading blurry text is easier than recognizing faces because there are only 26 letters to choose from, but pretty much every person has a unique face. Do we even know whether black holes are more like letters than faces? To me stars are like letters, at least from far away, but galaxies, planets, comets, and nebulae are like faces. I guess in technical terms I would say the models probably won't reproduce the individual gestalt properly. In biology we frequently encounter species differentiations that aren't easily diagnosed with hard measurements but if you know them well enough you can see the differing gestalt. Can the ml represent that property in a quantifiable form, and if so, how many examples would it need to be trained on, and if it can be trained on models how can you know the models are accurate with only one observed system to compare against? To the commenter speaking about an ionization chamber and detector (I'm having trouble recalling the exact post), yes, you may be modeling the detectors and the sample material to find your limit conditions but you also have the luxury of testing those models against your observations over hundreds (thousands? I really don't know much about this) of runs of varied parameters if need be, don't you? AFAIK there's only this one set of observed data to confirm the model, yet they think a best fit can be generalized to then reiterate an improved best fit? I'm thinking that's not what's going on with your example? What do I know, I didn't specialize in physics or machine learning, I've just been taught generally that nothing can reproduce unobserved data.
Upvote
4
(5
/
-1)
A Monte Carlo simulation is not as deterministic as you’re implying, neither is a ML model so distant from a Monte Carlo sulimulation.
Both are statistical techniques that use random sampling to try to converge on an answer, for instance.
Upvote
4
(5
/
-1)