Today I want to briefly mention a couple of papers about AI in astronomy research. These tackle very different questions from the usual sort, which might examine how good LLMs can be at summarising documents or reading figures and the like. These, especially the second, are much more philosophical than that.
The first uses an LLM to construct a knowledge graph for astronomy, attempting to link different concepts together. The idea is to show how, at a very high level, astronomical thinking has shifted over time : what concepts were typically connected and how this has changed. Using distributional semantics, where the meanings of words in relation to other words are encoded as numerical vectors, they construct a very pretty diagram showing how different astronomical topics relate to each other. And it certainly does look very nice – you can even play with it online.
It's quite fun to see how different concepts like galaxy and stellar physics relate to each other, how connected they are and how closely (or at least it would be if the damn thing would load faster). It's also interesting to see how different techniques have become more widely-used over time, with machine learning having soared in popularity in the last ten years. But exactly what the point of this is I'm not sure. It's nice to be able to visualise these things for the sake of aesthetics, but does this offer anything truly new ? I get the feeling it's like Hubble's Tuning Fork : nice to show, but nobody actually does anything with it because the graphical version doesn't offer anything that couldn't be conveyed with text.
Perhaps I'm wrong. I'd be more interested to see if such an approach could indicate which fields have benefited from methods that other fields aren't currently using, or more generally, to highlight possible multi-disciplinary approaches that have been thus far overlooked.
The second paper is far more provocative and interesting. It asks, quite bluntly, whether machine learning is a good thing for the natural sciences : this is very general, though astronomy seems to be the main focus.
They begin by noting that machine learning is good for performance, not understanding. I agree, but once we do understand, then surely performance improvements are what we're after. Machine learning is good for quantification, not qualitative understanding and certainly not for proposing new concepts (LLMs might, and I stress might, be able to help with this). But it's a rather strange thing to examine, and possibly a bit of a straw man, since I've never heard of anyone thinking that ML could do this. And they admit that ML can be obviously beneficial in certain kinds of numerical problems, but this is still a bit strange : what, if any, qualitative problems is ML supposed to ever help with ?
Not that quantitative and qualitative are entirely separable. Sometimes once you obtain a number you can robustly exclude or confirm a particular model, so in that sense the qualitative requires the quantitative. But, as they rightly point out, as I have myself many times, interpretation is a human thing : machines know numbers but nothing else. More interestingly they note :
The things we care about are almost never directly observable... In physics, for example, not only do the data exist, but so do forces, energies, momenta, charges, spacetime, wave functions, virtual particles, and much more. These entities are judged to exist in part because they are involved in the latent structure of the successful theories; almost none of them are direct observables.
Well, this is something I've explored a lot on Decoherency (just go there and search for "triangles"). But I have to ask, what is the difference between an observation and a measurement ? For example we can see the effects of electrical charge by measuring, say, the deflection of a hair in the static field of a balloon, but we don't observe charge directly. But we also don't observe radio waves directly, yet we don't think they're less real than optical photons, which we do. Likewise some animals do appear to be able to sense charge and magnetic fields directly. In what sense, then, are these "real" and what sense are they just convenient labels we apply ?
I don't know. The extreme answer is that all we have are perceptions, i.e. labels, and no access to anything "real" at all, but this remains (in some ways) deeply unsatisfactory; again, see innumerable Decoherency posts on this, search for "neutral monism". Perhaps here it doesn't matter so much though. The point is that ML cannot extract any sort of qualitative parameters at all, whereas to humans these matter very much – regardless of their "realness" or otherwise. If you only quantify and never qualify, you aren't doing science, you're just constructing a mathematical model of the world : ultimately you might be able to interpolate perfectly but you'd have no extrapalatory power at all.
Tying in with this and perhaps less controversially are their statements regarding why some models are preferred over others :
When the expansion of the Universe was discovered, the discovery was important, but not because it permitted us to predict the values of the redshifts of new galaxies (though it did indeed permit that). The discovery was important because it told us previously unknown things about the age and evolution of the Universe, and it confirmed a prediction of general relativity, which is a theory of the latent structure of space and time. The discovery would not have been seen as important if Hubble and Humason had instead announced that they had trained a deep multilayer perceptron that could predict the Doppler shifts of held-out extragalactic nebulae.
Yes ! Hubble needed the numbers to formulate an interpretation, but the numbers themselves don't interpret anything. A device or mathematical model capable of predicting the redshifts from other data, without saying why the redshifts take the values that they do, without relating it to any other physical quantities at all, would be mathematical magic, and not really science.
For another example, consider the discovery that the paths of the planets are ellipses, with the Sun at one focus. This discovery led to extremely precise predictions for data. It was critical to this discovery that the data be well explained by the theory. But that was not the primary consideration that made the nascent scientific community prefer the Keplerian model. After all, the Ptolemaic model preceding Kepler made equally accurate predictions of held-out data. Kepler’s model was preferred because it fit in with other ideas being developed at the same time, most notably heliocentrism.
A theory or explanation has to do much more than just explain the data in order to be widely accepted as true. In physics for example, a model — which, as we note, is almost always a model of latent structure — is judged to be good or strongly confirmed not only if it explains observed data. It ought to explain data in multiple domains, and it must connect in natural ways to other theories or principles (such as conservation laws and invariances) that are strongly confirmed themselves.
General relativity was widely accepted by the community not primarily because it explained anomalous data (although it did explain some); it was adopted because, in addition to explaining (a tiny bit of new) data, it also had good structure, it resolved conceptual paradoxes in the pre-existing theory of gravity, and it was consistent with emerging ideas of field theory and geometry.
Which is a nice summary. Some time ago I'd almost finished a draft of a much longer post based on this this far more detailed paper which considers the same issues, but then blogger lost it all and I haven't gotten around to re-writing the bloody thing. I may yet try. Anyway the need for self-consistency is important, and doesn't throttle new theories in their infancy as you might expect : there are ways to overturn established findings independent of the models.
The rest of the paper is more-or-less in line with my initial expectations. ML is great, they say, when only quantification is needed : when a correlation is interesting regardless of causation, or when you want to find outliers. So long as the causative factors are well-understood (and sometimes they are !) it can be a powerful tool for rapidly finding trends in the data and points which don't match the rest.
If the trends are not well-understood ahead of time, it can reinforce biases, in particular confirmation bias by matching what was expected in advance. Similarly, if there are rival explanations possible, ML doesn't help you choose between them if they don't predict anything significantly different. But often, no understanding is necessary. To remove the background variations in a telescope's image it isn't necessary even to know where all the variations come from : it's usually obvious that they are artifacts, and all you need to is the mathematical description of them. Or more colourfully, "You do not have to understand your customers to make plenty of revenue off of them."
Wise words. Less wise, perhaps only intended as a joke, are the comments about "the unreasonable effectiveness of ML", that it's remarkable that these industrial-grade mathematical processes are any good for situations to which they were never designed. But I never even got around to blogging Wigner's famous "unreasonable effectiveness" essay because it seemed worryingly silly.
Finally, they note that it might be better if natural sciences were to shift their focus away from theories and more towards the data, and that the degeneracies in the sciences undermine the "realism" of the models. Well, you do you : it's provocative, but on this occasion, I shall allow myself not to be provoked. Shut up and calculate ? Nah. Shut up and contemplate.
No comments:
Post a Comment