a few thoughts on hyperparams for a better learning theory (for understanding what happens when a neural net is trained with gradient descent)

Having found myself repeating the same points/claims in various conversations about what NN learning is like (especially around singular learning theory), I figured it's worth writing some of them down. My typical confidence in a claim below is like 95%^[1]. I'm not claiming anything here is significantly novel. The claims/points:

• local learning (eg gradient descent) strongly does not find global optima. insofar as running a local learning process from many seeds produces outputs with 'similar' (train or test) losses, that's a law of large numbers phenomenon^[2], not a consequence of always finding the optimal neural net weights.^[3][4]
  • if your method can't produce better weights: were you trying to produce better weights by running gradient descent from a bunch of different starting points? getting similar losses this way is a LLN phenomenon
  • maybe this is a crisp way to see a counterexample instead: train, then identify a 'lottery ticket' subnetwork after training like done in that literature. now get rid of all other edges in the network, and retrain that subnetwork either from the previous initialization or from a new initialization — i think this literature says that you get a much worse loss in the latter case. so training from a random initialization here gives a much worse loss than possible
• dynamics (kinetics) matter(s). the probability of getting to a particular training endpoint is highly dependent not just on stuff that is evident from the neighborhood of that point, but on there being a way to make those structures incrementally, ie by a sequence of local moves each of which is individually useful.^[5][6][7] i think that this is not an academic correction, but a major one — the structures found in practice are very massively those with sensible paths into them and not other (naively) similarly complex structures. some stuff to consider:
  • the human eye evolving via a bunch of individually sensible steps, https://en.wikipedia.org/wiki/Evolution_of_the_eye
  • (given a toy setup and in a certain limit,) the hardness of learning a boolean function being characterized by its leap complexity, ie the size of the 'largest step' between its fourier terms, https://arxiv.org/pdf/2302.11055
  • imagine a loss function on a plane which has a crater somewhere and another crater with a valley descending into it somewhere else. the local neighborhoods of the deepest points of the two craters can look the same, but the crater with a valley descending into it will have a massively larger drainage basin. to say more: the crater with a valley is a case where it is first loss-decreasing to build one simple thing, (ie in this case to fix the value of one parameter), and once you've done that loss-decreasing to build another simple thing (ie in this case to fix the value of another parameter); getting to the isolated crater is more like having to build two things at once. i think that with a reasonable way to make things precise, the drainage basin of a 'k-parameter structure' with no valley descending into it will be exponentially smaller than that of eg a 'k-parameter structure' with 'a k/2-parameter valley' descending into it, which will be exponentially smaller still than a 'k-parameter structure' with a sequence of valleys of slowly increasing dimension descending into it
  • it seems plausible to me that the right way to think about stuff will end up revealing that in practice there are basically only systems of steps where a single [very small thing]/parameter gets developed/fixed at a time
  • i'm further guessing that most structures basically have 'one way' to descend into them (tho if you consider sufficiently different structures to be the same, then this can be false, like in examples of convergent evolution) and that it's nice to think of the probability of finding the structure as the product over steps of the probability of making the right choice on that step (of falling in the right part of a partition determining which next thing gets built)
  • one correction/addition to the above is that it's probably good to see things in terms of there being many 'independent' structures/circuits being formed in parallel, creating some kind of ecology of different structures/circuits. maybe it makes sense to track the 'effective loss' created for a structure/circuit by the global loss (typically including weight norm) together with the other structures present at a time? (or can other structures do sufficiently orthogonal things that it's fine to ignore this correction in some cases?) maybe it's possible to have structures which were initially independent be combined into larger structures?^[8]
• everything is a loss phenomenon. if something is ever a something-else phenomenon, that's logically downstream of a relation between that other thing and loss (but this isn't to say you shouldn't be trying to find these other nice things related to loss)
  • grokking happens basically only in the presence of weight regularization, and it has to do with there being slower structures to form which are eventually more efficient at making logits high (ie more logit bang for weight norm buck)
  • in the usual case that generalization starts to happen immediately, this has to do with generalizing structures being stronger attractors even at initialization. one consideration at play here is that
  • nothing interesting ever happens during a random walk on a loss min surface
• it's not clear that i'm conceiving of structures/circuits correctly/well in the above. i think it would help a library of like >10 well-understood toy models (as opposed to like the maybe 1.3 we have now), and to be very closely guided by them when developing an understanding of neural net learning

some related (more meta) thoughts

• to do interesting/useful work in learning theory (as of 2024), imo it matters a lot that you think hard about phenomena of interest and try to build theory which lets you make sense of them, as opposed to holding fast to an existing formalism and trying to develop it further / articulate it better / see phenomena in terms of it
  • this is somewhat downstream of current formalisms imo being bad, it imo being appropriate to think of them more as capturing preliminary toy cases, not as revealing profound things about the phenomena of interest, and imo it being feasible to do better
  • but what makes sense to do can depend on the person, and it's also fine to just want to do math lol
  • and it's certainly very helpful to know a bunch of math, because that gives you a library in terms of which to build an understanding of phenomena
• it's imo especially great if you're picking phenomena to be interested in with the future going well around ai in mind

(* but it looks to me like learning theory is unfortunately hard to make relevant to ai alignment^[9])

acknowledgments

these thoughts are sorta joint with Jake Mendel and Dmitry Vaintrob (though i'm making no claim about whether they'd endorse the claims). also thank u for discussions: Sam Eisenstat, Clem von Stengel, Lucius Bushnaq, Zach Furman, Alexander Gietelink Oldenziel, Kirke Joamets

a thing i think is probably happening and significant in such cases: developing good 'concepts/ideas' to handle a problem, 'getting a feel for what's going on in a (conceptual) situation'

a plausibly analogous thing in humanity(-seen-as-a-single-thinker): humanity states a conjecture in mathematics, spends centuries playing around with related things (tho paying some attention to that conjecture), building up mathematical machinery/understanding, until a proof of the conjecture almost just falls out of the machinery/understanding

I find it surprising/confusing/confused/jarring that you speak of models-in-the-sense-of-mathematical-logic=:L-models as the same thing as (or as a precise version of) models-as-conceptions-of-situations=:C-models. To explain why these look to me like two pretty much entirely distinct meanings of the word 'model', let me start by giving some first brushes of a picture of C-models. When one employs a C-model, one likens a situation/object/etc of interest to a situation/object/etc that is already understood (perhaps a mathematical/abstract one), that one expects to be better able to work/play with. For example, when one has data about sun angles at a location throughout the day and one is tasked with figuring out the distance from that location to the north pole, one translates the question to a question about 3d space with a stationary point sun and a rotating sphere and an unknown point on the sphere and so on. (I'm not claiming a thinker is aware of making such a translation when they make it.) Employing a C-model $\approx$ making an analogy. From inside a thinker, the objects/situations on each side of the analogy look like... well, things/situations; from outside a thinker, both sides are thinking-elements.^[1] (I think there's a large GOFAI subliterature trying to make this kind of picture precise but I'm not that familiar with it; here are two papers that I've only skimmed: https://www.qrg.northwestern.edu/papers/Files/smeff2(searchable).pdf , https://api.lib.kyushu-u.ac.jp/opac_download_md/3070/76.ps.tar.pdf .)

I'm not that happy with the above picture of C-models, but I think that it seeming like an even sorta reasonable candidate picture might be sufficient to see how C-models and L-models are very different, so I'll continue in that hope. I'll assume we're already on the same page about what an L-model is ( https://en.wikipedia.org/wiki/Model_theory ). Here are some ways in which C-models and L-models differ that imo together make them very different things:

• An L-model is an assignment of meaning to a language, a 'mathematical universe' together with a mapping from symbols in a language to stuff in that universe — it's a semantic thing one attaches to a syntax. The two sides of a C-modeling-act are both things/situations which are roughly equally syntactic/semantic (more precisely: each side is more like a syntactic thing when we try to look at a thinker from the outside, and just not well-placed on this axis from the thinker's internal point of view, but if anything, the already-understood side of the analogy might look more like a mechanical/syntactic game than the less-understood side, eg when you are aware that you are taking something as a C-model).
• Both sides of a C-model are things/situations one can reason about/with/in. An L-model takes from a kind of reasoning (proving, stating) system to an external universe which that system could talk about.
• An L-model is an assignment of a static world to a dynamic thing; the two sides of a C-model are roughly equally dynamic.
• A C-model might 'allow you to make certain moves without necessarily explicitly concerning itself much with any coherent mathematical object that these might be tracking'. Of course, if you are employing a C-model and you ask yourself whether you are thinking about some thing, you will probably answer that you are, but in general it won't be anywhere close to 'fully developed' in your mind, and even if it were (whatever that means), that wouldn't be all there is to the C-model. For an extreme example, we could maybe even imagine a case where a C-model is given with some 'axioms and inference rules' such that if one tried to construct a mathematical object 'wrt which all these axioms and inference rules would be valid', one would not be able to construct anything — one would find that one has been 'talking about a logically impossible object'. Maybe physicists handling infinities gracefully when calculating integrals in QFT is a fun example of this? This is in contrast with an L-model which doesn't involve anything like axioms or inference rules at all and which is 'fully developed' — all terms in the syntax have been given fixed referents and so on.
• (this point and the ones after are in some tension with the main picture of C-models provided above but:) A C-model could be like a mental context/arena where certain moves are made available/salient, like a game. It seems difficult to see an L-model this way.
• A C-model could also be like a program that can be run with inputs from a given situation. It seems difficult to think of an L-model this way.
• A C-model can provide a way to talk about a situation, a conceptual lens through which to see a situation, without which one wouldn't really be able to [talk about]/see the situation at all. It seems difficult to see an L-model as ever doing this. (Relatedly, I also find it surprising/confusing/confused/jarring that you speak of reasoning using C-models as a semantic kind of reasoning.)

(But maybe I'm grouping like a thousand different things together unnaturally under C-models and you have some single thing or a subset in mind that is in fact closer to L-models?)

All this said, I don't want to claim that no helpful analogy could be made between C-models and L-models. Indeed, I think there is the following important analogy between C-models and L-models:

• When we look for a C-model to apply to a situation of interest, perhaps we often look for a mathematical object/situation that satisfies certain key properties satisfied by the situation. Likewise, an L-model of a set of sentences is (roughly speaking) a mathematical object which satisfies those sentences.

(Acknowledgments. I'd like to thank Dmitry Vaintrob and Sam Eisenstat for related conversations.)

1. ^{^}

   This is complicated a bit by a thinker also commonly looking at the C-model partly as if from the outside — in particular, when a thinker critiques the C-model to come up with a better one. For example, you might notice that the situation of interest has some property that the toy situation you are analogizing it to lacks, and then try to fix that. For example, to guess the density of twin primes, you might start from a naive analogy to a probabilistic situation where each 'prime' p has probability (p-1)/p of not dividing each 'number' independently at random, but then realize that your analogy is lacking because really p not dividing n makes it a bit less likely that p doesn't divide n+2, and adjust your analogy. This involves a mental move that also looks at the analogy 'from the outside' a bit.

That said, the hypothetical you give is cool and I agree the two principles decouple there! (I intuitively want to save that case by saying the COM is only stationary in a covering space where the train has in fact moved a bunch by the time it stops, but idk how to make this make sense for a different arrangement of portals.) I guess another thing that seems a bit compelling for the two decoupling is that conservation of angular momentum is analogous to conservation of momentum but there's no angular analogue to the center of mass (that's rotating uniformly, anyway). I guess another thing that's a bit compelling is that there's no nice notion of a center of energy once we view spacetime as being curved ( https://physics.stackexchange.com/a/269273 ). I think I've become convinced that conservation of momentum is a significantly bigger principle :). But still, the two seem equivalent to me before one gets to general relativity. (I guess this actually depends a bit on what the proof of 12.72 is like — in particular, if that proof basically uses the conservation of momentum, then I'd be more happy to say that the two aren't equivalent already for relativity/fields.)

here's a picture from https://hansandcassady.org/David%20J.%20Griffiths-Introduction%20to%20Electrodynamics-Addison-Wesley%20(2012).pdf :

Given 12.72, uniform motion of the center of energy is equivalent to conservation of momentum, right? P is const <=> dR_e/dt is const.

(I'm guessing 12.72 is in fact correct here, but I guess we can doubt it — I haven't thought much about how to prove it when fields and relativistic and quantum things are involved. From a cursory look at his comment, Lubos Motl seems to consider it invalid lol ( in https://physics.stackexchange.com/a/3200 ).)

The microscopic picture that Mark Mitchison gives in the comments to this answer seems pretty: https://physics.stackexchange.com/a/44533 — though idk if I trust it. The picture seems to be to think of glass as being sparse, with the photon mostly just moving with its vacuum velocity and momentum, but with a sorta-collision between the photon and an electron happening every once in a while. I guess each collision somehow takes a certain amount of time but leaves the photon unchanged otherwise, and presumably bumps that single electron a tiny bit to the right. (Idk why the collisions happen this way. I'm guessing maybe one needs to think of the photon as some electromagnetic field thing or maybe as a quantum thing to understand that part.)

And the loss mechanism I was imagining was more like something linear in the distance traveled, like causing electrons to oscillate but not completely elastically wrt the 'photon' inside the material.

Anyway, in your argument for the redshift as the photon enters the block, I worry about the following:

1. can we really think of 1 photon entering the block becoming 1 photon inside the block, as opposed to needing to think about some wave thing that might translate to photons in some other way or maybe not translate to ordinary photons at all inside the material (this is also my second worry from earlier)?
2. do we know that this photon-inside-the-material has energy $\hslash \omega$?

re redshift: Sorry, I should have been clearer, but I meant to talk about redshift (or another kind of energy loss) of the light that comes out of the block on the right compared to the light that went in from the left, which would cause issues with going from there being a uniformly-moving stationary center of mass to the conclusion about the location of the block. (I'm guessing you were right when you assumed in your argument that redshift is 0 for our purposes, but I don't understand light in materials well enough atm to see this at a glance atm.)

Note however, that the principle being broken (uniform motion of centre of mass) is not at all one of the "big principles" of physics, especially not with the extra step of converting the photon energy to mass. I had not previously heard of the principle, and don't think it is anywhere near the weight class of things like momentum conservation.

I found these sentences surprising. To me, the COM moving at constant velocity (in an inertial frame) is Newton's first law, which is one of the big principles (and I also have a mental equality between that and conservation of momentum).

I guess we can also reach your conclusion in that thought experiment arguing from conservation of momentum directly (though I guess the argument I'll give just contains a proof of one direction of the equivalence to the conservation of momentum as a step). Ignoring relativity for a second, we could go into the center of mass frame as the particle approaches the piece of glass from the left, then note that the momentum in this frame needs to zero forever (by conservation of momentum), then note $\int p \text{d}t=m\delta(x)$, where $\delta(x)$ is the distance moved by the center of mass, from which $\delta(x)=0$. I would guess that essentially the same argument also works when relativistic things like photons are involved (and when fields or quantum stuff is involved), as long as one replaces the center of mass by the center of energy ( https://physics.stackexchange.com/questions/742770/centre-of-energy-in-special-relativity ).

One thing that worries me about that thought experiment more than [whether Newton's first law carries over to this context] is the assumption that (in ideal conditions) photons do not lose any energy to the material — that they don't end up redshifted or something. (If photons got redshifted as they go through, then the photons would lose some energy and the block would end up with some momentum and heat, obviously causing issues with the broader argument.) Still, I guess it's probably fine to say that frequency/energy of the light is indeed conserved ( https://physics.stackexchange.com/questions/810869/why-does-the-energy-and-thus-frequency-of-a-photon-entering-glass-stay-constan ), but I unfortunately don't atm understand how to think about a light packet (or something) going through a (potentially moving) material well enough to decide for myself atm. (ChatGPT tells me of some standard argument involving the displacement field, but I haven't decided if I'll trust that argument in this context yet. I also tried to see whether such an effect would be higher-order in some parameter even if it existed but I didn't see a good reason why that would be the case.)

A second thing that worries me about this argument even more is whether it even makes sense to talk about individual photons passing through materials — I think the argument doesn't make sense if photon number is not conserved before vs after a light pulse enters a material (here I'm thinking of the light pulse having small horizontal extent compared to the material). But I really haven't thought very carefully about this. (Also, I'd like to point out that if some kind of light packet number were conserved and we are operating with a notion of momentum such that all of it can be attributed to wave packets, then momentum conservation implies the momentum attributed to a given packet stays constant. But I guess some of it might be more naturally attributed to stuff in the block at some point. I'd need to think more about what kind of partition would be most natural.)

It additionally seems likely to me that we are presently missing major parts of a decent language for talking about minds/models, and developing such a language requires (and would constitute) significant philosophical progress. There are ways to 'understand the algorithm a model is' that are highly insufficient/inadequate for doing what we want to do in alignment — for instance, even if one gets from where interpretability is currently to being able to replace a neural net by a somewhat smaller boolean (or whatever) circuit and is thus able to translate various NNs to such circuits and proceed to stare at them, one probably won't thereby be more than $\frac{1}{10}$ of the way to the kind of strong understanding that would let one modify a NN-based AGI to be aligned or build another aligned AI (in case alignment doesn't happen by default) (much like how knowing the weights doesn't deliver that kind of understanding). To even get to the point where we can usefully understand the 'algorithms' models implement, I feel like we might need to have answered sth like (1) what kind of syntax should we see thinking as having — for example, should we think of a model/mind as a library of small programs/concepts that are combined and updated and created according to certain rules (Minsky's frames?), or as having a certain kind of probabilistic world model that supports planning in a certain way, or as reasoning in a certain internal logical language, or in terms of having certain propositional attitudes; (2) what kind of semantics should we see thinking as having — what kind of correspondence between internals of the model/mind and the external world should we see a model as maintaining(; also, wtf are values). I think that trying to find answers to these questions by 'just looking' at models in some ML-brained, non-philosophical way is unlikely to be competitive with trying to answer these questions with an attitude of taking philosophy (agent foundations) seriously, because one will only have any hope of seeing the cognitive/computational structure in a mind/model by staring at it if one stares at it already having some right ideas about what kind of structure to look for. For example, it'd be very tough to try to discover [first-order logic]/ZFC/[type theory] by staring at the weights/activations/whatever of the brain of a human mathematician doing mathematical reasoning, from a standpoint where one hasn't already invented [first-order logic]/ZFC/[type theory] via some other route — if one starts from the low-level structure of a brain, then first-order logic will only appear as being implemented in the brain in some 'highly encrypted' way.

There's really a spectrum of claims here that would all support the claim that agent foundations is good for understanding the 'algorithm' a model/mind is to various degrees. A stronger one than what I've been arguing for is that once one has these ideas, one needn't stare at models at all, and that staring at models is unlikely to help one get the right ideas (e.g. because it's better to stare at one's own thinking instead, and to think about how one could/should think, sort of like how [first-order logic]/ZFC/[type theory] was invented), so one's best strategy does not involve starting at models; a weaker one than what I've been arguing is that having more and better ideas about the structure of minds would be helpful when staring at models. I like TsviBT's koan on this topic.