[SiempreViernes]

Actually, Hossenfelders great fight has been with the concept of "naturalness", a fight that has now been won by the LHC killing off all the theories based on that concept.

And it wasn't ever so simple as "chasing mathematical elegance instead of trying to explain observations", the problem has been that there was a theory that could explain almost perfectly everything within a certain region of physics, but can't easily be extended.

Thus you work on crazy schemes to extend the existing theory (all the sensible ones already having failed), or you are forced to make an entirely new framework, and that takes a lot of work before it's finished enough to even reproduce the results of the limited theory.

If you take the second route, you are very vulnerable to the "chasing mathematical elegance" slander, but it's not like the other guys are doing any better: there aren't actually any unexpected observations that need explaining within the reach of the existing theory.

[nonbel]

>"the problem has been that there was a theory that could explain almost perfectly everything within a certain region of physics, but can't easily be extended."

Sounds a lot like overfitting.

[kosievdmerwe]

Sounds more like a formula that needs more terms.

Compare the classical physics formula for momentum, p=mv, vs relativistic formula, p=γmv. γ is almost 1 for most low velocities, it only starts jumping up to infinity when we get close to c.

The point being that the classical formula is pretty good in it's zone of low velocities, but as soon as you get too far out of the implicit term's "constraints" the formula breaks down and you need to add more to it to get it working for both low and high velocities. Which doesn't sound easy.

[nonbel]

Adding more terms sounds like more overfitting.

[jfengel]

It sounds like that only because scientists reformulate their models in terms that people are familiar with. Actual physicists don't work in terms of the gamma correction factor; they work in tensor fields that don't look anything like conventional arithmetic.

But they can pare all that down to something expressed in terms people are familiar with. And that has the bonus purpose of helping them understand why the familiar terms were familiar: the "correction factor" is small under circumstances we encounter, and only becomes large under circumstances we rarely do.

If that intrigues somebody enough to learn the actual physics, they'll encounter a completely different and more-encompassing formulation which looks not at all like overfitting. One that turns out to be more elegant, in fact, cramming more information into less notation. But it's information nobody needs until they're doing fairly advanced physics, so we're not going to be teaching it in elementary school any time soon.

[SiempreViernes]

Don't worry about it, I think you're talking to a NN stuck in a local optima around the term "overfit".

[nonbel]

I don't see how what you wrote could possibly address concerns with adding parameters to the model until it explains current data so perfectly that it cannot generalize to future or other data.

This is not a "theory" problem, it has to do with matching the theory to observations.

[SiempreViernes]

If that’s what it sounds like then that is a failure of my description: the problem is nothing at all like overfitting, and casting it in that light would be pointless.

[nonbel]

How many free parameters do "the foundations of physics" allow?

I was concerned a few years back when a "blip" at CERN resulted in theoretical physicists publishing 300+ different theories to explain it in a short period of time. All of these theories were presumably consistent with "the foundations of physics". And I guess that "blip" got rejected as a not something worth explaining anyway.

Sounds like post-hoc overfitting to me.

[SiempreViernes]

Calling it ”overfitting” is actually you overfitting your model of how theory works on this one concept from machine learning.

The actual events of the 750 GeV peak are much closer to neural networks hallucinating, and the diversity of models created doesen’t strike me as evidence of overfitting...

Anyway, you clearly didn’t trust my summary, and I no longer trust you have an honest interest in learning more, so I’ll stop here.

[nonbel]

If you can come up with 300 different explanations for a random blip, your theory is obviously very flexible. You are probably overfitting.

If your solution to the problem of not being able to generalize your model to new/other data is to add more parameters. You are probably overfitting.

These are the hallmarks of overfitting.

Sorry that you cannot learn from others and only expect them to learn from you.

[SiempreViernes]

Look at this: theorists come up with something like 300 different models, and you declare they all used the same theory to do it? Of course they didn't! Why would you assume they all did the same thing?!

Further, of the various frameworks used, many will have been created by imposing some further symmetry on the standard theory, in effect decreasing the number of free parameters!