Since there is no other suitable field for me to insert the following, I use this box as a general purpose "further review comments" box. The following suggestions are intended to assist the authors when conducting their re-write. The AC does not need to read this section.
Minor Corrections / Editorial Suggestions
C1 On lines 46/47, "or a variational learning only recently made to train them" does not make sense. Changing the phrasing to something like "or bespoke variational learning strategies" would be more readable.
C2 On line 49, "We start from..." is an odd way to begin this paragraph because the authors have yet to specify what their goal is. If they wish to keep things general, this phrase would be improved by writing "We start by investigating...". Otherwise, the authors should be explicit about their goal of designing a new approximate posterior, which then motivates "starting" somewhere in the eyes of the reader. An easy way to achieve this would be "Motivated by the need to design ever better approximate posteriors for BNNs, we start by investigating...".
C3 On line 56, "Section 4 shows what it does to the fit" reads ambiguously; Section 4 of the current work or of Farquhar et al. (2024)? I would alter this sentence to something like "and we demonstrate the negative impact this has on approximate posterior quality in Section 4". "Fit" is also a rather ambiguous term in the context of BNNs; do the authors mean fit-to-data (e.g. expected log-likelihood term in VI, or predictive mean RMSE etc) or degree to which the approximate posterior matches the true posterior?
C4 On line 67, "is the uniform prior" would read better as "corresponds to the uniform prior".
C5 On lines 81 and 82, the sentence starting with "The network's output..." reads poorly. There is no need to compare the case for weights w with weights cw for any c>0 since the former is already a special case of the latter where c=1. It would be simpler to write something like "Since the network output is then invariant to a scaling of the weights by multiplicative scaling factor c>0, the likelihood also exhibits scale invariance such that dcdp(D∣cw)=0".
C6 On line 95, the authors write "almost all of the Gaussian’s KL is spent on the radius". What does it mean to "spend" KL divergence? This language needs to be formalised.
C7 On line 102, "the network's weight" should be removed. Also, the authors should clarify what they mean by "per unit or per layer"; what determines which is the case? Furthermore, the authors never introduce the meaning of CD(κ); line 102 would be the perfect place to do that.
C8 On line 104, what does I⋅(κ) represent? The authors should introduce this notation before using it.
C9 On line 105, "in both limits" needs to be clarified; in both limits of what?
C10 On line 105, the authors need to define σu before using it.
C11 On lines 105/106, the authors introduce σeff as the "noise relative to the signal". Why don't they refer to it as the noise-to-signal ratio (NSR), the coefficient of variation, or any other common term?
C12 On lines 106 and 136, the authors write "the layer computes [an equation] and learns logσeff. How does the layer itself learn the noise parameter? If the authors instead mean that the noise parameter is learned/optimised during training, they should write something to that effect instead.
C13 On line 137, I suggest the authors swap out "absorbs" in favour of "eliminates the effect of".
C14 The acronym "VD" is never introduced, so readers have no way to understand what the authors mean by it. This should be fixed.
C15 In Figure 1's caption, what do the authors mean CAP's posterior is a cap on the unit circle? The authors should remove that ambiguous phrase in favour of something more precise, and hopefully something that makes it clearer to readers that the red density curve in the leftmost plot is intended to come out-of-the-page rather than veering off the unit circle within the plane of the page (which is how I first interpreted it and I expect other readers might too). Also, the unconventional phrase "spends mass" should be swapped out in favour of "places mass".
C16 In Figure 2's caption and line 139, the authors refer to "the family", but they never describe what family they mean. If they mean something like the von Mises-Fisher family of approximate posteriors for neuron weights, they should write that somewhere.
C17 Everything in Figure 2 is far too small. Axis labels, legend labels, tick labels, and so on, are all practically illegible. The authors should fix this.
C18 On line 141, "The dictionary between noise and angle" is a meaningless phrase. Perhaps the authors mean something like "The relationship between noise and angle"? The rest of this paragraph is also difficult to understand; the authors should work to improve the clarity.
C19 The authors need to explain where equation 3 comes from. Have they derived it or adopted it from prior work?
C20 On line 158, the authors need to formalise their definition and exposition of the objective function. They never introduce the acronym NLL (even though many readers might guess it refers to negative log-likelihood). The authors should justify where this objective function comes from, as it appears similar but different to the tempered ELBO (which would have an expected log-likelihood term instead of an NLL term). Furthermore, while readers might guess that the KL terms are additive across layers, and therefore that the subscript ⋅l indexes network layers, the authors need to mention this somewhere explicitly!
C21 The first sentence of section 3 is clunky and difficult to understand. The authors should explain what is meant by the "tangent covariance of a von Mises-Fisher". Furthermore, "; across units q factorises" should be replaced with something like ", and the posterior q factorises between units" (to make it clear that this is a separate point to the within-unit covariance structure".
C22 The second sentence of section 3 is also both clunky and difficult to understand. Which two assumptions are the authors referring to? What is meant by a "scalable" BNN? What is meant by "the posterior is too hard to sample to test them"\textemdashI am not aware of any BNN architecture whose posterior can be sampled, except for a single hidden-layer architecture?
C23 The third sentence of Section 3 is difficult to understand. What do the authors mean by "On the directions of a small normalised network HMC converges"?
C24 In the paragraph below Proposition 2, what do the authors mean that the inequality in Proposition 2 "is" Cauchy-Schwarz? On lines 190-191, what "predictions" are the authors referring to? What do they mean by an "erf teacher" and "ReLU teacher"? What do they mean by "The error of a single concentration per unit is thus a known function of the posterior’s shape"? This whole paragraph needs to be re-written in clearer terms.
C25 On line 198, the authors need to introduce the acronyms AM and HM somewhere before using them.
C26 On line 199, the authors should replace "crossed" with something like "investigated".
C27 On line 203, it makes no sense to "affect different assumptions". Something can affect the validity of an assumption, but not an assumption itself.
C28 The plots in Figure 3 need to be larger. Also, r needs to be defined before it is used.
C29 On lines 231-232, the authors claim that normalised ReLU networks develop sparse activations as they train (in general). The authors need to support this with suitable a argumentation, demonstration, or citation.
C30 Section 4 is more clear than Section 3, but there is still insufficient explanation of the student-teacher setup. This should be addressed.
C31 Figure 5's results should be displayed in some kind of figure (table or plot). The current method of reporting adds to the paper's immense lack of clarity.
C32 On line 323, the authors state the AM/HM ratio vanishes. By this, do they mean it converges to 1.0? If so, they should make that clear.
C33 On line 346, what does an "arm" refer to? The authors should clarify this.
Thank you for the review and for two technical points that improved the paper. We are glad that you found the paper well motivated, and that the HMC comparisons and the controlled experiments on normalisation and sparsity were useful.
"…standard layer normalization does not generally satisfy this assumption…" Indeed, layer normalisation does not normalise each unit, as weight and batch normalisation do, and we have made this precise in the paper (Section 2 and Appendix A). Our results are unaffected, because none of our networks relies on layer normalisation for the invariance.
"Which tested layers satisfy independent per-unit scale invariance?" All of them. We weight-normalise every trained layer with a fixed gain and follow it with batch normalisation without affine parameters (Appendix K), which gives the invariance for each unit. In the teacher-student networks the weight vectors have unit norm, so each unit is also normalised. We also apply layer normalisation to the pre-activations of some of these networks, to measure its effect on the posterior (Section 4.1, Appendix I).
"For a unit input aligned with the mean direction … their leading orders are different." Our closed form ‖x‖²/(κ + D) assumes (µᵀx)² ≈ ‖x‖²/D, an input with no preferred alignment to the mean direction, which is the typical case in high dimension. Appendix A of the original manuscript stated this assumption, and we now state it in the main text as well (Section 3). For an input aligned with µ the leading order is different, and we derive it in Appendix B ("For an input along µ the variance is instead C∥‖x‖², which for large κ is ‖x‖²(D − 1)/(2κ²)"), in agreement with your expression.
"Under what assumptions on dimension and input direction is the variance approximation accurate?" It is accurate for inputs whose normalised pre-activation z is of order one, and the more so the larger the fan-in D. The exact variance is C∥(µᵀx)² + C⊥(‖x‖² − (µᵀx)²) and (µᵀx)² = z²‖x‖²/D, so the variance differs from the unaligned case (z² = 1) by a relative amount of at most |z² − 1|/D. This is below 4% for |z| ≤ 3 at fan-ins of 256 and above. The aligned input corresponds to z² = D. In the unaligned case the closed form is within 20% of the exact variance at every κ (4% at κ/D = 10).
"Does correcting its directional dependence change the posterior and calibration results?" We tested this by retraining with the exact dependence, in two ways. We first replaced the Gaussian noise by a vMF sampler of the weights, which has the full dependence on the direction of the input, and the ECE stays at 3.1% (Table 6). We then kept the Gaussian noise and used the exact variance in place of the closed form, and the learned noise scale is 2% smaller while the accuracy, ECE and NLL change by at most one standard error (Table 3). The results therefore do not change.