kaarelh AT gmail DOT com
Kaarel
I’d be very interested in a concrete construction of a (mathematical) universe in which, in some reasonable sense that remains to be made precise, two ‘orthogonal pattern-universes’ (preferably each containing ‘agents’ or ‘sophisticated computational systems’) live on ‘the same fundamental substrate’. One of the many reasons I’m struggling to make this precise is that I want there to be some condition which meaningfully rules out trivial constructions in which the low-level specification of such a universe can be decomposed into a pair such that and are ‘independent’, everything in the first pattern-universe is a function only of , and everything in the second pattern-universe is a function only of . (Of course, I’d also be happy with an explanation why this is a bad question :).)
I started writing this but lost faith in it halfway through, and realized I was spending too much time on it for today. I figured it’s probably a net positive to post this mess anyway although I have now updated to believe somewhat less in it than the first paragraph indicates. Also I recommend updating your expected payoff from reading the rest of this somewhat lower than it was before reading this sentence. Okay, here goes:
{I think people here might be attributing too much of the explanatory weight on noise. I don’t have a strong argument for why the explanation definitely isn’t noise, but here is a different potential explanation that seems promising to me. (There is a sense in which this explanation is still also saying that noise dominates over any relation between the two variables—well, there is a formal sense in which that has to be the case since the correlation is small—so if this formal thing is what you mean by “noise”, I’m not really disagreeing with you here. In this case, interpret my comment as just trying to specify another sense in which the process might not be noisy at all.) This might be seen as an attempt to write down the “sigmoids spiking up in different parameter ranges” idea in a bit more detail.
First, note that if the performance on every task is a perfectly deterministic logistic function with midpoint x_0 and logistic growth rate k, i.e. there is “no noise”, with k and x_0 being the same across tasks, then these correlations would be exactly 0. (Okay, we need to be adding an epsilon of noise here so that we are not dividing by zero when calculating the correlation, but let’s just do that and ignore this point from now on.) Now as a slightly more complicated “noiseless” model, we might suppose that performance on each task is still given by a “deterministic” logistic function, but with the parameters k and x_0 being chosen at random according to some distribution. It would be cool to compute some integrals / program some sampling to check what correlation one gets when k and x_0 are both normally distributed with reasonable means and variances for this particular problem, with no noise beyond that.}
This is the point where I lost faith in this for now. I think there are parameter ranges for how k and x_0 are distributed where one gets a significant positive correlation and ranges where one gets a significant negative correlation in the % case. Negative correlations seem more likely for this particular problem. But more importantly, I no longer think I have a good explanation why this would be so close to 0. I think in logit space, the analysis (which I’m omitting here) becomes kind of easy to do by hand (essentially because the logit and logistic function are inverses), and the outcome I’m getting is that the correlation should be positive, if anything. Maybe it becomes negative if one assumes the logistic functions in our model are some other sigmoids instead, I’m not sure. It seems possible that the outcome would be sensitive to such details. One idea is that maybe if one assumes there is always eps of noise and bounds the sigmoid away from 1 by like 1%, it would change the verdict.
Anyway, the conclusion I was planning to reach here is that there is a plausible way in which all the underlying performance curves would be super nice, not noisy at all, but the correlations we are looking at would still be zero, and that I could also explain the negative correlations without noisy reversion to the mean (instead this being like a growth range somewhere decreasing the chance there is a growth range somewhere else) but the argument ended up being much less convincing than I anticipated. In general, I’m now thinking that most such simple models should have negative or positive correlation in the % case depending on the parameter range, and could be anything for logit. Maybe it’s just that these correlations are swamped by noise after all. I’ll think more about it.
I think this comment is incorrect (in the stated generality). Here is a simple counterexample. Suppose you have a starting endowment of $1, and that you can bet any amount at 0.50001 probability of doubling your bet and 0.49999 probability of losing everything you bet. You can bet whatever amount of your money you want a total of n times. (If you lost everything in some round, we can think of this as you still being allowed to bet 0 in remaining future rounds.) The strategy that maximizes expected linear utility is the one where you bet everything every time.
I don’t understand why 1 is true – in general, couldn’t the variable $W$ be defined on a more refined sample space? Also, I think all $4$ conditions are technically satisfied if you set $W=X$ (or well, maybe it’s better to think of it as a copy of $X$).I think the following argument works though. Note that the distribution of $X$ given $(Z,Y,W)$ is just the deterministic distribution $X=Y \xor Z$ (this follows from the definition of Z). By the structure of the causal graph, the distribution of $X$ given $(Z,Y,W)$ must be the same as the distribution of $X$ given just $W$. Therefore, the distribution of $X$ given $W$ is deterministic. I strongly guess that a deterministic connection is directly ruled out by one of Pearl’s inference rules.
The same argument also rules out graphs 2 and 4.
I took the main point of the post to be that there are fairly general conditions (on the utility function and on the bets you are offered) in which you should place each bet like your utility is linear, and fairly general conditions in which you should place each bet like your utility is logarithmic. In particular, the conditions are much weaker than your utility actually being linear, or than your utility actually being logarithmic, respectively, and I think this is a cool point. I don’t see the post as saying anything beyond what’s implied by this about Kelly betting vs max-linear-EV betting in general.
I’d be quite interested in elaboration on getting faster alignment researchers not being alignment-hard — it currently seems likely to me that a research community of unupgraded alignment researchers with a hundred years is capable of solving alignment (conditional on alignment being solvable). (And having faster general researchers, a goal that seems roughly equivalent, is surely alignment-hard (again, conditional on alignment being solvable), because we can then get the researchers to quickly do whatever it is that we could do — e.g., upgrading?)
I think does not have to be a variable which we can observe, i.e. it is not necessarily the case that we can deterministically infer the value of from the values of and . For example, let’s say the two binary variables we observe are and . We’d intuitively want to consider a causal model where is causing both, but in a way that makes all triples of variable values have nonzero probability (which is true for these variables in practice). This is impossible if we require to be deterministic once is known.
more on 4: Suppose you have horribly cyclic preferences and you go to a rationality coach to fix this. In particular, your ice cream preferences are vanilla>chocolate>mint>vanilla. Roughly speaking, Hodge is the rationality coach that will tell you to consider the three types of ice cream equally good from now on, whereas Mr. Max Correct Pairs will tell you to switch one of the three preferences. Which coach is better? If you dislike breaking cycles arbitrarily, you should go with Hodge. If you think losing your preferences is worse than that, go with Max. Also, Hodge has the huge advantage of actually being done in a reasonable amount of time :)
I liked the post; here are some thoughts, mostly on the “The futility of computing utility” section:
1 )
If we’re not incredibly unlucky, we can hope to sort N-many outcomes with comparisons.
I don’t understand why you need to not be incredibly unlucky here. There are plenty of deterministic algorithms with this runtime, no?
2) I think that in step 2, once you know the worst and the best outcome, you can skip to step 3 (i.e. the full ordering does not seem to be needed to enter step 3. So instead of sorting in n log n time, you could find min and max in linear time, and then skip to the psychophysics.
3) Could you elaborate on what you mean by a basis of possibility-space? It is not obvious to me that possibility-space has a linear structure (i.e. that it is a vector space), or that utility respects this linear structure (e.g. the utility I get from having chocolate ice cream and having vanilla ice cream is not in general approximated well by the sum of the utilities I get from having only one of these, and similarly for multiplication by scalars). Perhaps you were using these terms metaphorically, but then I currently have doubts about this being a great metaphor / would appreciate having the metaphor spelled out more explicitly. I could imagine doing something like picking some random subset of the possibilities, doing something to figure out the utilities of this subset, and then doing some linear regression (or something more complicated) on various parameters to predict the utilities of all possibilities. It seems like a more symmetric way to think about this might be to consider the subset of outcomes (with each pair, or some of the pairs, being labeled according to which one is preferred) to be the training data, and then training a neural network (or whatever) that predict utilities of outcomes so as to minimize loss on this training data. And then to go from training data to any possibility, just unleash the same neural network on that possibility. (Under this interpretation, I guess “elementary outcomes” <-> training data, and there does not seem to be a need to assume linear structure of possibility-space.)
4) I think I have something cool to say about a specific and very related problem. Input: a set of outcomes and some pairwise preferences between them. Desired output: a total order on these outcomes such that the number of pairs which are ordered incorrectly is minimal. It turns out that this is NP-hard: https://epubs.siam.org/doi/pdf/10.1137/050623905. (The problem considered in this paper is the special case of the problem I stated where all the pairwise preferences are given, but well, if a special case is NP-hard, then the problem itself is also NP-hard.)
You can prove e.g. by (backwards) induction that you should bet everything every time. With the odds being p>0.5 and 1-p, if the expectation of whatever strategy you are using after n-1 steps is E, then the maximal expectation over all things you could do on the n’th step is p2E (you can see this by writing the expectation as a conditional sum over the outcomes after n-1 steps), which corresponds uniquely to the strategy where you bet everything in any situation on the n’th step. It then follows that the best you can do on the (n-1)th step is also to maximize the expectation after it, and the same argument gives that you should bet everything, and so on.
(Where did you get n=10^5 from? If it came from some computer computation, then I would wager that there was some overflow/numerical issues.)
A thread into which I’ll occasionally post notes on some ML(?) papers I’m reading
I think the world would probably be much better if everyone made a bunch more of their notes public. I intend to occasionally copy some personal notes on ML(?) papers into this thread. While I hope that the notes which I’ll end up selecting for being posted here will be of interest to some people, and that people will sometimes comment with their thoughts on the same paper and on my thoughts (please do tell me how I’m wrong, etc.), I expect that the notes here will not be significantly more polished than typical notes I write for myself and my reasoning will be suboptimal; also, I expect most of these notes won’t really make sense unless you’re also familiar with the paper — the notes will typically be companions to the paper, not substitutes.
I expect I’ll sometimes be meaner than some norm somewhere in these notes (in fact, I expect I’ll sometimes be simultaneously mean and wrong/confused — exciting!), but I should just say to clarify that I think almost all ML papers/posts/notes are trash, so me being mean to a particular paper might not be evidence that I think it’s worse than some average. If anything, the papers I post notes about had something worth thinking/writing about at all, which seems like a good thing! In particular, they probably contained at least one interesting idea!
So, anyway: I’m warning you that the notes in this thread will be messy and not self-contained, and telling you that reading them might not be a good use of your time :)
A few notes/questions about things that seem like errors in the paper (or maybe I’m confused — anyway, none of this invalidates any conclusions of the paper, but if I’m right or at least justifiably confused, then these do probably significantly hinder reading the paper; I’m partly posting this comment to possibly prevent some readers in the future from wasting a lot of time on the same issues):
1) The formula for here seems incorrect:
This is because W_i is a feature corresponding to the i’th coordinate of x (this is not evident from the screenshot, but it is evident from the rest of the paper), so surely what shows up in this formula should not be W_i, but instead the i’th row of the matrix which has columns W_i (this matrix is called W later). (If one believes that W_i is a feature, then one can see this is wrong already from the dimensions in the dot product not matching.)
2) Even though you say in the text at the beginning of Section 3 that the input features are independent, the first sentence below made me make a pragmatic inference that you are not assuming that the coordinates are independent for this particular claim about how the loss simplifies (in part because if you were assuming independence, you could replace the covariance claim with a weaker variance claim, since the 0 covariance part is implied by independence):However, I think you do use the fact that the input features are independent in the proof of the claim (at least you say “because the x’s are independent”):
Additionally, if you are in fact just using independence in the argument here and I’m not missing something, then I think that instead of saying you are using the moment-cumulants formula here, it would be much much better to say that independence implies that any term with an unmatched index is . If you mean the moment-cumulants formula here https://en.wikipedia.org/wiki/Cumulant#Joint_cumulants , then (while I understand how to derive every equation of your argument in case the inputs are independent), I’m currently confused about how that’s helpful at all, because one then still needs to analyze which terms of each cumulant are 0 (and how the various terms cancel for various choices of the matching pattern of indices), and this seems strictly more complicated than problem before translating to cumulants, unless I’m missing something obvious.
3) I’m pretty sure this should say x_i^2 instead of x_i x_j, and as far as I can tell the LHS has nothing to do with the RHS:(I think it should instead say sth like that the loss term is proportional to the squared difference between the true and predictor covariance.)
An attempt at a specification of virtue ethics
I will be appropriating terminology from the Waluigi post. I hereby put forward the hypothesis that virtue ethics endorses an action iff it is what the better one of Luigi and Waluigi would do, where Luigi and Waluigi are the ones given by the posterior semiotic measure in the given situation, and “better” is defined according to what some [possibly vaguely specified] consequentialist theory thinks about the long-term expected effects of this particular Luigi vs the long-term effects of this particular Waluigi. One intuition here is that a vague specification could be more fine if we are not optimizing for it very hard, instead just obtaining a small amount of information from it per decision.
In this sense, virtue ethics literally equals continuously choosing actions as if coming from a good character. Furthermore, considering the new posterior semiotic measure after a decision, in this sense, virtue ethics is about cultivating a virtuous character in oneself. Virtue ethics is about rising to the occasion (i.e. the situation, the context). It’s about constantly choosing the Luigi in oneself over the Waluigi in oneself (or maybe the Waluigi over the Luigi if we define “Luigi” as the more likely of the two and one has previously acted badly in similar cases or if the posterior semiotic measure is otherwise malign). I currently find this very funny, and, if even approximately correct, also quite cool.
Here are some issues/considerations/questions that I intend to think more about:
What’s a situation? For instance, does it encompass the agent’s entire life history, or are we to make it more local?
Are we to use the agent’s own semiotic measure, or some objective semiotic measure?
This grounds virtue ethics in consequentialism. Can we get rid of that? Even if not, I think this might be useful for designing safe agents though.
Does this collapse into cultivating a vanilla consequentialist over many choices? Can we think of examples of prompting regimes such that collapse does not occur? The vague motivating hope I have here is that in the trolley problem case with the massive man, the Waluigi pushing the man is a corrupt psycho, and not a conflicted utilitarian.
Even if this doesn’t collapse into consequentialism from these kinds of decisions, I’m worried about it being stable under reflection, I guess because I’m worried about the likelihood of virtue ethics being part of an agent in reflective equilibrium. It would be sad if the only way to make this work would be to only ever give high semiotic measure to agents that don’t reflect much on values.
Wait, how exactly do we get Luigi and Waluigi from the posterior semiotic measure? Can we just replace this with picking the best character from the most probable few options according to the semiotic measure? Wait, is this just quantilization but funnier? I think there might be some crucial differences. And regardless, it’s interesting if virtue ethics turns out to be quantilization-but-funnier.
More generally, has all this been said already?
Is there a nice restatement of this in shard theory language?
I agree with you regarding 0 lebesgue. My impression is that the Pearl paradigm has some [statistics → causal graph] inference rules which basically do the job of ruling out causal graphs for which having certain properties seen in the data has 0 lebesgue measure. (The inference from two variables being independent to them having no common ancestors in the underlying causal graph, stated earlier in the post, is also of this kind.) So I think it’s correct to say “X has to cause Y”, where this is understood as a valid inference inside the Pearl (or Garrabrant) paradigm. (But also, updating pretty close to “X has to cause Y” is correct for a Bayesian with reasonable priors about the underlying causal graphs.)
(epistemic position: I haven’t read most of the relevant material in much detail)
In this comment, I will be assuming that you intended to talk of “pivotal acts” in the standard (distribution of) sense(s) people use the term — if your comment is better described as using a different definition of “pivotal act”, including when “pivotal act” is used by the people in the dialogue you present, then my present comment applies less.
I think that this is a significant mischaracterization of what most (? or definitely at least a substantial fraction of) pivotal activists mean by “pivotal act” (in particular, I think this is a significant mischaracterization of what Yudkowsky has in mind). (I think the original post also uses the term “pivotal act” in a somewhat non-standard way in a similar direction, but to a much lesser degree.) Specifically, I think it is false that the primary kinds of plans this fraction of people have in mind when talking about pivotal acts involve creating a superintelligent nigh-omnipotent infallible FOOMed properly aligned ASI. Instead, the kind of person I have in mind is very interested in coming up with pivotal acts that do not use a general superintelligence, often looking for pivotal acts that use a narrow superintelligence (for instance, a narrow nanoengineer) (though this is also often considered very difficult by such people (which is one of the reasons they’re often so doomy)). See, for instance, the discussion of pivotal acts in https://www.lesswrong.com/posts/7im8at9PmhbT4JHsW/ngo-and-yudkowsky-on-alignment-difficulty.
A small observation about the AI arms race in conditions of good infosec and collaboration
Suppose we are in a world where most top AI capabilities organizations are refraining from publishing their work (this could be the case because of safety concerns, or because of profit motives) + have strong infosec which prevents them from leaking insights about capabilities in other ways. In this world, it seems sort of plausible that the union of the capabilities insights of people at top labs would allow one to train significantly more capable models than the insights possessed by any single lab alone would allow one to train. In such a world, if the labs decide to cooperate once AGI is nigh, this could lead to a significantly faster increase in capabilities than one might have expected otherwise.
(I doubt this is a novel thought. I did not perform an extensive search of the AI strategy/governance literature before writing this.)
A big chunk of my uncertainty about whether at least 95% of the future’s potential value is realized comes from uncertainty about “the order of magnitude at which utility is bounded”. That is, if unbounded total utilitarianism is roughly true, I think there is a <1% chance in any of these scenarios that >95% of the future’s potential value would be realized. If decreasing marginal returns in the [amount of hedonium → utility] conversion kick in fast enough for 10^20 slightly conscious humans on heroin for a million years to yield 95% of max utility, then I’d probably give >10% of strong utopia even conditional on building the default superintelligent AI. Both options seem significantly probable to me, causing my odds to vary much less between the scenarios.
This is assuming that “the future’s potential value” is referring to something like the (expected) utility that would be attained by the action sequence recommended by an oracle giving humanity optimal advice according to our CEV. If that’s a misinterpretation or a bad framing more generally, I’d enjoy thinking again about the better question. I would guess that my disagreement with the probabilities is greatly reduced on the level of the underlying empirical outcome distribution.
Something that confuses me about your example’s relevance is that it’s like almost the unique case where it’s [[really directly] impossible] to succumb to optimization pressure, at least conditional on what’s good = something like coherent extrapolated volition. That is, under (my understanding of) a view of metaethics common in these corners, what’s good just is what a smarter version of you would extrapolate your intuitions/[basic principles] to, or something along these lines. And so this is almost definitionally almost the unique situation that we’d expect could only move you closer to better fulfilling your values, i.e. nothing could break for any reason, and in particular not break under optimization pressure (where breaking is measured w.r.t. what’s good). And being straightforwardly tautologically true would make it a not very interesting example.
editorial remark: I realized after writing the two paragraphs below that they probably do not move one much on the main thesis of your post, at least conditional on already having read Ege Erdil’s doubts about your example (except insofar as someone wants to defer to opinions of others or my opinion in particular), but I decided to post anyway in large part since these family matters might be a topic of independent interest for some:
I would bet that at least 25% of people would stop loving their (current) family in <5 years (i.e. not love them much beyond how much they presently love a generic acquaintance) if they got +30 IQ. That said, I don’t claim the main case of this happening is because of applying too much optimization pressure to one’s values, at least not in a way that’s unaligned with what’s good—I just think it’s likely to be the good thing to do (or like, part of all the close-to-optimal packages of actions, or etc.). So I’m not explicitly disagreeing with the last sentence of your comment, but I’m disagreeing with the possible implicit justification of the sentence that goes through [“I would stop loving my family” being false].
The argument for it being good to stop loving your family in such circumstances is just that it’s suboptimal for having an interesting life, or for [the sum over humans of interestingness of their lives] if you are altruistic, or whatever, for post-IQ-boost-you to spend a lot of time with people much dumber than you, which your family is now likely to be. (Here are 3 reasons to find a new family: you will have discussions which are more fun → higher personal interestingness; you will learn more from these discussions → increased productivity; and something like productivity being a convex function of IQ—this comes in via IQs of future kids, at least assuming the change in your IQ would be such as to partially carry over to kids. I admit there is more to consider here, e.g. some stuff with good incentives, breaking norms of keeping promises—my guess is that these considerations have smaller contributions.)
The Deep Neural Feature Ansatz
@misc{radhakrishnan2023mechanism, title={Mechanism of feature learning in deep fully connected networks and kernel machines that recursively learn features}, author={Adityanarayanan Radhakrishnan and Daniel Beaglehole and Parthe Pandit and Mikhail Belkin}, year={2023}, url = { https://arxiv.org/pdf/2212.13881.pdf } }
The ansatz from the paper
Let denote the activation vector in layer on input , with the input layer being at index , so . Let be the weight matrix after activation layer . Let be the function that maps from the th activation layer to the output. Then their Deep Neural Feature Ansatz says that (I’m somewhat confused here about them not mentioning the loss function at all — are they claiming this is reasonable for any reasonable loss function? Maybe just MSE? MSE seems to be the only loss function mentioned in the paper; I think they leave the loss unspecified in a bunch of places though.)
A singular vector version of the ansatz
Letting be a SVD of , we note that this is equivalent to i.e., that the eigenvectors of the matrix on the RHS are the right singular vectors. By the variational characterization of eigenvectors and eigenvalues (Courant-Fischer or whatever), this is the same as saying that right singular vectors of are the highest orthonormal directions for the matrix on the RHS. Plugging in the definition of , this is equivalent to saying that the right singular vectors are the sequence of highest-variance directions of the data set of gradients .
(I have assumed here that the linearity is precise, whereas really it is approximate. It’s probably true though that with some assumptions, the approximate initial statement implies an approximate conclusion too? Getting approx the same vecs out probably requires some assumption about gaps in singular values being big enough, because the vecs are unstable around equality. But if we’re happy getting a sequence of orthogonal vectors that gets variances which are nearly optimal, we should also be fine without this kind of assumption. (This is guessing atm.))
Getting rid of the dependence on the RHS?
Assuming there isn’t an off-by-one error in the paper, we can pull some term out of the RHS maybe? This is because applying the chain rule to the Jacobians of the transitions gives , so
Wait, so the claim is just which, assuming is invertible, should be the same as . But also, they claim that it is ? Are they secretly approximating everything with identity matrices?? This doesn’t seem to be the case from their Figure 2 though.
Oh oops I guess I forgot about activation functions here! There should be extra diagonal terms for jacobians of preactivations->activations in , i.e., it should really say We now instead get This should be the same as which, with denoting preactivations in layer and denoting the function from these preactivations to the output, is the same as This last thing also totally works with activation functions other than ReLU — one can get this directly from the Jacobian calculation. I made the ReLU assumption earlier because I thought for a bit that one can get something further in that case; I no longer think this, but I won’t go back and clean up the presentation atm.
Anyway, a takeaway is that the Deep Neural Feature Ansatz is equivalent to the (imo cleaner) ansatz that the set of gradients of the output wrt the pre-activations of any layer is close to being a tight frame (in other words, the gradients are in isotropic position; in other words still, the data matrix of the gradients is a constant times a semi-orthogonal matrix). (Note that the closeness one immediately gets isn’t in to a tight frame, it’s just in the quantity defining the tightness of a frame, but I’d guess that if it matters, one can also conclude some kind of closeness in from this (related).) This seems like a nicer fundamental condition because (1) we’ve intuitively canceled terms and (2) it now looks like a generic-ish condition, looks less mysterious, though idk how to argue for this beyond some handwaving about genericness, about other stuff being independent, sth like that.
proof of the tight frame claim from the previous condition: Note that clearly implies that the mass in any direction is the same, but also the mass being the same in any direction implies the above (because then, letting the SVD of the matrix with these gradients in its columns be , the above is , where we used the fact that ).
Some questions
Can one come up with some similar ansatz identity for the left singular vectors of ? One point of tension/interest here is that an ansatz identity for would constrain the left singular vectors of together with its singular values, but the singular values are constrained already by the deep neural feature ansatz. So if there were another identity for in terms of some gradients, we’d get a derived identity from equality between the singular values defined in terms of those gradients and the singular values defined in terms of the Deep Neural Feature Ansatz. Or actually, there probably won’t be an interesting identity here since given the cancellation above, it now feels like nothing about is really pinned down by ‘gradients independent of ’ by the DNFA? Of course, some -dependence remains even in the gradients because the preactivations at which further gradients get evaluated are somewhat -dependent, so I guess it’s not ruled out that the DNFA constrains something interesting about ? But anyway, all this seems to undermine the interestingness of the DNFA, as well as the chance of there being an interesting similar ansatz for the left singular vectors of .
Can one heuristically motivate that the preactivation gradients above should indeed be close to being in isotropic position? Can one use this reduction to provide simpler proofs of some of the propositions in the paper which say that the DNFA is exactly true in certain very toy cases?
The authors claim that the DNFA is supposed to somehow elucidate feature learning (indeed, they claim it is a mechanism of feature learning?). I take ‘feature learning’ to mean something like which neuronal functions (from the input) are created or which functions are computed in a layer in some broader sense (maybe which things are made linearly readable?) or which directions in an activation space to amplify or maybe less precisely just the process of some internal functions (from the input to internal activations) being learned of something like that, which happens in finite networks apparently in contrast to infinitely wide networks or NTK models or something like that which I haven’t yet understood? I understand that their heuristic identity on the surface connects something about a weight matrix to something about gradients, but assuming I’ve not made some index-off-by-one error or something, it seems to probably not really be about that at all, since the weight matrix sorta cancels out — if it’s true for one , it would maybe also be true with any other replacing it, so it doesn’t really pin down ? (This might turn out to be false if the isotropy of preactivation gradients is only true for a very particular choice of .) But like, ignoring that counter, I guess their point is that the directions which get stretched most by the weight matrix in a layer are the directions along which it would be the best to move locally in that activation space to affect the output? (They don’t explain it this way though — maybe I’m ignorant of some other meaning having been attributed to in previous literature or something.) But they say “Informally, this mechanism corresponds to the approach of progressively re-weighting features in proportion to the influence they have on the predictions.”. I guess maybe this is an appropriate description of the math if they are talking about reweighting in the purely linear sense, and they take features in the input layer to be scaleless objects or something? (Like, if we take features in the input activation space to each have some associated scale, then the right singular vector identity no longer says that most influential features get stretched the most.) I wish they were much more precise here, or if there isn’t a precise interesting philosophical thing to be deduced from their math, much more honest about that, much less PR-y.
So, in brief, instead of “informally, this mechanism corresponds to the approach of progressively re-weighting features in proportion to the influence they have on the predictions,” it seems to me that what the math warrants would be sth more like “The weight matrix reweights stuff; after reweighting, the activation space is roughly isotropic wrt affecting the prediction (ansatz); so, the stuff that got the highest weight has most effect on the prediction now.” I’m not that happy with this last statement either, but atm it seems much more appropriate than their claim.
I guess if I’m not confused about something major here (plausibly I am), one could probably add 1000 experiments (e.g. checking that the isotropic version of the ansatz indeed equally holds in a bunch of models) and write a paper responding to them. If you’re reading this and this seems interesting to you, feel free to do that — I’m also probably happy to talk to you about the paper.
typos in the paper
indexing error in the first displaymath in Sec 2: it probably should say ″, not ″
I find [the use of square brackets to show the merge structure of [a linguistic entity that might otherwise be confusing to parse]] delightful :)