Any single galaxy reveals the composition of an entire universe | Hacker News Reader

Any single galaxy reveals the composition of an entire universe

quantamagazine.org

The title is misleading in an important way. What ML algorithm has found out that from properties of a single _simulated_ galaxy it was possible to deduce the properties of а _simulated_ universe.

So at this point it is not clear if this relation is an artifact of the simulation or there is a real dependency.

ML model made homogenous output.

The cosmological background radiation, which is an image of the universe shortly after the big bang and before inflation, looks similar to the current layout of matter across the universe. This would seem consistent with a universe being able to be extrapolated from a region of it if there's some underlying holographic relationships.

Based on this clip from David Tong. Starts here and goes on a few minutes: https://youtu.be/zNVQfWC_evg?t=2714

That's addressed in the article:

> The team spent half a year trying to understand how the neural network had gotten so wise. They checked to make sure the algorithm hadn’t just found some way to infer the density from the coding of the simulation rather than the galaxies themselves.

Except that they acknowledge its performance was dependent on the details of the simulation and failed with a different recipe for making universes, so in fact it was dependent on not just the code, but the settings.

You're right to say the title is misleading. It is touched in the very end of the article:

> But in practice, the technique would have to first overcome a major weakness. The CAMELS collaboration cooks up its universes using two different recipes. A neural network trained on one of the recipes makes bad density guesses when given galaxies that were baked according to the other. The cross-prediction failure indicates that the neural network is finding solutions unique to the rules of each recipe

The title tells us that there are results for multiple universes, which clearly implies that they're working with simulated data...

To see a world in a grain of sand

And a heaven in a wild flower,

Hold infinity in the palm of your hand,

And eternity in an hour.

— William Blake

“Buddhism uses a similar image to describe the interconnectedness of all phenomena. It is called Indra's Net. When Indra fashioned the world, he made it as a web, and at every knot in the web is tied a pearl. Everything that exists, or has ever existed, every idea that can be thought about, every datum that is true—every dharma, in the language of Indian philosophy—is a pearl in Indra's net. Not only is every pearl tied to every other pearl by virtue of the web on which they hang, but on the surface of every pearl is reflected every other jewel on the net. Everything that exists in Indra's web implies all else that exists.”

— Timothy Brook

Reminded me of the excellent documentary "Dangerous Knowledge".

That which is above is like to that which is below, and that which is below is like to that which is above, to accomplish the miracles of one thing

https://www.wikiwand.com/en/Emerald_Tablet

https://en.wikipedia.org/wiki/Emerald_Tablet

I, for one, am highly skeptical about the current wave of ML-enabled “discoveries” in physics and chemistry, where ML models replace simulations. ML can’t reliably extrapolate into never before seen territory in the way a simulation can. This has to backfire at some point.

Even simulations fail to push fundamental knowledge forward. Experiments are needed to create models. There are plasma regimes we simply have no empirical observations of, so have no sensical way to simulate.

A friend of mine finished their PhD and is now doing a PostDoc helping in the creation of the first matter-antimatter plasma (electrically neutral where both species have the same mass). It's really cool work.

https://www.ipp.mpg.de/4793936/nwg-19

This sounds like the neural net just learned how to identify an input parameter to the simulator based on some of its output.

It might be noteworthy if we could assume the simulator's behavior is equivalent to natural laws and starting conditions to within some absurdly small error bound, but there's no way that level of accuracy is possible (unless I'm mistaken).

The generator for a simulation may not be able to predict what the real Earth would look like correctly given infinite Earth-like planets, but that doesn't mean that this algorithm wouldn't be able to correctly deduce some input parameter for the real Earth after training in the sim (like the number of days in a year based on a single minute of simulated video from the surface).

It's a bit of a stretch to call it proof, but the article says they spent half a year of work on making sure it's not just an artifact of the simulation they're basing the prediction on. I wouldn't be surprised if this did actually work, if we could ever collect equivalent real-world data.

An important gotcha is buried at the end:

> The CAMELS collaboration cooks up its universes using two different recipes. A neural network trained on one of the recipes makes bad density guesses when given galaxies that were baked according to the other. The cross-prediction failure indicates that the neural network is finding solutions unique to the rules of each recipe. It certainly wouldn’t know what to do with the Milky Way, a galaxy shaped by the real laws of physics.

This sounds like it takes the cosmological principle to its extreme.

Such assumptions should be constantly challenged.

> The improbable find grew out of an exercise Villaescusa-Navarro gave to Jupiter Ding, a Princeton University undergraduate: Build a neural network that, knowing a galaxy’s properties, can estimate a couple of cosmological attributes.

Jupiter Ding... maybe nominative determinism has something to it [1].

[1] https://en.wikipedia.org/wiki/Nominative_determinism

Reply on news.ycombinator.com