acceptodds
Under review as a conference paper at ICLR 2027

When Do Spikes Matter? Timing Beyond Rates

Abstract

When does a spiking neural network (SNN) depend only on how many spikes each neuron fires, and when does it also depend on when they fire? This matters: an SNN that ignores timing is only a time-stepped rate network, and one that uses timing loses information when spikes are reduced to counts. We study this question in wide networks of leaky integrate-and-fire (LIF) neurons with Gaussian weights, which pass on only the mean and the temporal second moment of each layer's spike trains. In wide networks this yields a kernel over spike trains, the spiking analogue of the neural-network Gaussian-process (NNGP) kernel, and rate coding keeps only its time-averaged part. It suffices for exact rate computation that synaptic filtering removes every temporal component except the average, and for approximate rate computation that little passes through. Conversely, we construct two spike populations with identical spike counts and firing rates for which one spiking layer produces different outputs, even when only their rates are read out. In this model timing matters only when structure survives filtering and reaches the readout; for trained networks these quantities are diagnostics.

Primary area:
applications to neuroscience & cognitive science
Keywords:
spiking neural networks, spike timing, rate coding, temporal coding, neural network Gaussian process, wide networks, leaky integrate-and-fire, ANN-to-SNN conversionv

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Ru8BOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors port the analysis of wide networks with Gaussian weights to SNNs. This requires a new setup due to sequence output, synaptic operators and LIF activations. They establish a descriptor of a layer (the mean spike train and second moment) that is sufficient to understand the behaviour of the next layer. They then look at the descriptor when spike trains are decomposed into the constant-in-time direction and its orthogonal complement and use this to describe under what conditions temporal (rather than just rate) information affects network outputs. They follow up this theoretical treatment with a number of experiments that test theory-inspired predictions.

Critical strengths

The paper appears to make mathematically sound progress that establishes a useful and well-motivated way to analyse SNNs. Theorems such as 3.1's exact finite-width sufficiency seem quite important.

A caveat is that I did not cover all of the theoretical work that extends into the appendix and seems important for the paper as a whole. Furthermore, I am relatively unfamiliar with this area and cannot comment confidently on whether the work is novel and well-placed in the literature.

Critical weaknesses

I suggest a weak reject as the main section (without appendices) does not present a self-contained coherent account of the work. I imagine this issue is relatively straightforward to address.

While I would guess that the concepts and results are likely interesting to many people who study SNNs, I think the paper does a poor job of communicating its ideas. Three issues are:

  1. It is not possible to read the main text alone and understand the paper.
  2. The consequences of the assumptions needed to create the theory should be made clearer.
  3. Although more of a minor issue, an improvement to the mathematical presentation could make reading the paper far easier.

I spent a reasonable amount of time trying to read this paper, yet I do not think I have been able to properly review the whole paper. This is partly due to the subject matter being relatively unfamiliar to me, and partly because the main text is so inextricable from the appendix. I wonder if the material would not be better split into two: wide-SNN kernel theory plus convergence experiments and a "when timing matters" investigation.

Below, are the three points above in more detail.

1. Main vs appendix

$ n_l, Q_{ab}^{l}, c_n^l, V_{\Lambda}, Q_{\text{out}} $ are examples of important objects that appear in the main text without full definition or reference to the appendix where many are defined. Such objects without sufficient definition appear throughout the main text and are later defined in the appendix. Other objects such as $ b_{l+1} $ and $ \mathcal{F}_l $ get terse and informal definitions in the main text, with much clearer definitions in the appendix. This makes it difficult to read the main text without reading the appendix.

Not just symbols, but I would argue that conceptual material needed to understand the main text is missing. For example, the Experiments section (L305-) investigates a list of predictions listed only in the appendix.

2. Consequences of assumptions

The antecedent (If $ B_l\Pi_{\perp} = 0 $ of Theorem 3.3's implication seems to preclude most causal synaptic operators including the Toeplitz filter used as the motivating example (L128). If so, this significantly undermines the usefulness of the theorem. Consider the equivalent non-block case: $ A_lP_{\perp} = 0$. We can say:

& A_lP_{\perp} = 0 \\\\ &\implies A_l(I - P_{0}) = 0 \\\\ &\implies A_l = A_lP_0 \\\\ &\implies A_l = A_lv_0 v_0^T \\\\ &\implies A_l = uv_0^T & \text{(for some vector } u \text{ )} \\\\ \end{align*}$$ But as all entries of \$ v_0 \$ are equal, \$ A_l \$ must have all equal columns. I think this rules out all causal operators except maybe columns of [0, 0, 0, ... 1]^T which accumulates all input into the final time. The causality of \$ A_l \$ is strongly suggested by the example given on L128. If I haven't made a mistake here, I think the authors should explicitly state which nontrivial class of temporal operators Theorem 3.3 covers. While I have not noticed any other such issue, I do think that assumptions that underpin the theory should be explored in more detail in order to convince readers that the theory isn't restricted to applying to a trivial situation. ### 3. Improve mathematical presentation Some suggestions: - Define types of objects. For example, \$ x_1, x_2, ..., X^{M} \$ are of unknown type. - Do not overload \$ S_i^l $. This symbol is used as two distinct vectors ($ S_i^l \in \mathbb{R}^T, S_i^l \in \mathbb{R}^D \$) and two different functions outputting those vectors. - The notation is lacking a stable semantic convention for superscripts and subscripts due to mixing indexing with tagging and superscripts and subscripts. The use of subscript \$ n \$ not as an index or coordinate, but as a marker of the finite case is confusing and not described well. Compare \$ Q_n^l \$ which is the \$ Q \$ for finite layer \$ l \$ versus \$ Q_{ab}^l \$ which is an \$ (a,b) \$ index into the \$ Q \$ for the infinite width layer, \$ l \$. Consider \$A_l, b_l, B_l, W_{ij}^l, S_i^l, X^l, \mathcal{F}_l, M_l, c_n^l \$, where \$ l \$ indexes the network layer. Similarly, block extension is inconsistent, with \$ A_l \to B_l \$ and \$ \Psi \to \Psi^{(M)} \$ being the block extensions. Addressing these issues would make the paper easier to read.
Questions for authors
  1. Were there no other uses of AI that fall under the required to disclose list? (https://iclr.cc/Conferences/2027/AIPolicyForAuthors)

  2. Do you have thoughts on how this theoretical foundation could (or could not) be used in theoretical neuroscience, where there is a similar rate vs. timing question?

7 comments

We thank the reviewer for the careful reading and for working through the condition of Theorem 3.3. Each point below led to a change in the paper.

1. "It is not possible to read the main text alone and understand the paper."

We agree, and in the revision the main text stands on its own. The objects it uses, with their types, in the notation of point 3 and with the seven that the reviewer names first:

ObjectTypeMeaning
nℓn_\ellpositive integernumber of neurons in layer ℓ\ell
Qℓ,ab\mathbf Q_{\ell,ab}RT×T\mathbb R^{T\times T}the block of the second moment Qℓ\mathbf Q_\ell for inputs xax_a and xbx_b, which is the temporal kernel Kℓ(xa,xb)\mathbf K_\ell(x_a,x_b), written Qabℓ\mathbf Q_{ab}^\ell in the submission
c^ℓ=(mℓ^,Qℓ^)\hat c_\ell=(\widehat{\mathbf m_\ell},\widehat{\mathbf Q_\ell})RD×RD×D\mathbb R^D\times\mathbb R^{D\times D} with D=MTD=MTdescriptor of layer ℓ\ell at finite width, the first and second moments of its stacked spike trains over its nℓn_\ell neurons, written cnℓc_n^\ell in the submission, with limit cℓ=(mℓ,Qℓ)c_\ell=(\mathbf m_\ell,\mathbf Q_\ell)
VΛV_{\boldsymbol\Lambda}closed subspace of the square-integrable functions of the inputthe functions that a readout with weighting Λ\boldsymbol\Lambda can represent, the closure of the range of the integral operator of its kernel KΛK_{\boldsymbol\Lambda}
Qout\mathbf Q_{\rm out}RD×D\mathbb R^{D\times D}second moment of the spike trains that a layer emits, the second component of Tℓ(c)\mathcal T_\ell(c)
bℓ+1(x)\mathbf b_{\ell+1}(x)RT\mathbb R^Tbias path of layer ℓ+1\ell+1 for input xx, deterministic and the same for every neuron
Fℓ\mathcal F_\ellσ\sigma-fieldthe information in the spike trains of layer ℓ\ell, generated by Sℓ,1,…,Sℓ,nℓ\mathbf S_{\ell,1},\ldots,\mathbf S_{\ell,n_\ell}
xx and XM\mathcal X_Man element of an arbitrary set, and a set of MM of theman input, and the inputs x1,…,xMx_1,\ldots,x_M on which descriptors and kernels are evaluated
sℓ,i(x)\mathbf s_{\ell,i}(x){0,1}T\lbrace0,1\rbrace^Tspike train of neuron ii in layer ℓ\ell for input xx
Sℓ,i\mathbf S_{\ell,i}{0,1}D\lbrace0,1\rbrace^Dthe spike trains of neuron ii for the MM inputs, stacked
CD\mathfrak C_Dset of descriptorsthe descriptors that some law on {0,1}D\lbrace0,1\rbrace^D realizes
ΨLIF\Psi_{\rm LIF} and Ψ(M)\Psi^{(M)}RT→{0,1}T\mathbb R^T\to\lbrace0,1\rbrace^T and RD→{0,1}D\mathbb R^D\to\lbrace0,1\rbrace^Dthe neuron, from a drive path to a spike train, and the same map applied to each of the MM inputs
Aℓ(M)\mathbf A_\ell^{(M)} and P0(M)\mathbf P_0^{(M)}D×DD\times D matricesthe synaptic operator Aℓ\mathbf A_\ell and the time average P0\mathbf P_0, each applied to every one of the MM inputs, written Bℓ\mathbf B_\ell and Π\boldsymbol\Pi in the submission
Tℓ\mathcal T_\ellCD→CD\mathfrak C_D\to\mathfrak C_Dlayer map of the wide limit, from the descriptor of layer ℓ\ell to that of layer ℓ+1\ell+1
R\mathcal RCD→RD×RD×D\mathfrak C_D\to\mathbb R^D\times\mathbb R^{D\times D}rate quotient, the time-averaged part (P0(M)m,P0(M)QP0(M))(\mathbf P_0^{(M)}\mathbf m,\mathbf P_0^{(M)}\mathbf Q\mathbf P_0^{(M)}) of a descriptor
Tℓ,dc\mathcal T_{\ell,{\rm dc}}CD→CD\mathfrak C_D\to\mathfrak C_Dthe layer map when the synapse sees only the rate quotient
Etemp(c;Aℓ(M))\mathcal E_{\rm temp}(c;\mathbf A_\ell^{(M)})nonnegative numbermean squared part of the drive that comes from the non-constant temporal directions
ϑt(s)\vartheta_t(\mathbf s) and Θt\Theta_ta number and a finite setthe threshold at time tt once the resets of the earlier spikes s\mathbf s are counted in, and the values it can take
Δ∗\Delta^\astinteger between 1 and TTthe coarsest window over which spikes can be averaged without changing the output kernel beyond a tolerance
KSNNK_{\rm SNN}kernel on pairs of inputscovariance of the Gaussian readout in the wide limit

The predictions that Section 4 tests:

PredictionTested in
P1finite networks approach the limit recursion with widthFigures 2 and 3a
P2the gap between a network and its rate projection vanishes with the transmitted temporal energyFigures 3b, 3c, 4a and 4b
P3inputs with equal spike counts can give different outputsFigure 4c
P4the timing difference after the synapse, the transmitted difference, predicts the output change better than the raw oneFigure 5b
P5the benefit of timing follows its alignment with the task, not its amountFigures 5a and 5c
P6converted and timing-trained networks separate when spike times are permuted at fixed countsFigure 6

We thank the reviewer for this point and for naming the objects one by one, which gave the main text the definitions it lacked. The revision adds them where each object first appears.

2. "The consequences of the assumptions needed to create the theory should be made clearer." and "…the authors should explicitly state which nontrivial class of temporal operators Theorem 3.3 covers."

We partially agree. The derivation is correct, but it does not restrict the theory to a trivial situation, since Theorem 3.3 is only its exact case and three other statements cover the operators it leaves out, which we may not have emphasized and clarify here.

The class of Theorem 3.3.

  • What it is. Exactly the last line of the derivation, A=γv0⊤\mathbf A=\boldsymbol\gamma\mathbf v_0^\top with a free vector γ\boldsymbol\gamma of TT gains (the reviewer's uu, renamed because utu_t is the membrane potential in the paper). It is a spike counter, which hands every time step a fixed gain times the number of spikes in the window.
  • Its only causal member. The one that the reviewer identifies, the terminal accumulator γTeTv0⊤\gamma_T\mathbf e_T\mathbf v_0^\top with eT\mathbf e_T the last unit vector, which releases the count at the last step. Theorem 3.3 therefore does not cover the Toeplitz filter of line 128.

The other operators. The table lists the statement that covers each kind, and the Toeplitz filter of line 128 falls under Theorem 3.4. In the windowed counter 1Wk\mathbf 1_{W_k} is the indicator vector of the kk-th window of Δ\Delta steps, and in conversion the neuron, and not the synapse, does the counting. What each kind of operator gets from the theory:

Temporal operator A\mathbf ACausalExact rate factorizationStatement that covers it
spike counter γv0⊤\boldsymbol\gamma\mathbf v_0^\top, for example the time average P0=v0v0⊤\mathbf P_0=\mathbf v_0\mathbf v_0^\topnoyesTheorem 3.3
terminal accumulator γTeTv0⊤\gamma_T\mathbf e_T\mathbf v_0^\topyesyes, the only causal caseTheorem 3.3
windowed counter ∑kγk1Wk⊤\sum_k\boldsymbol\gamma_k\mathbf 1_{W_k}^\top, each γk\boldsymbol\gamma_k supported at or after the last step of its window WkW_kyesyes, at resolution Δ\DeltaTheorem 3.3 at resolution Δ\Delta (Appendix A.2.12)
exponential filteryesnoTheorem 3.4 at resolution Δ\Delta (Appendix A.2.12)
identity, with the integrate-and-fire neuron of conversionyesnoProposition A.17, the rate readout closes up to O(1/T)O(1/T) when the terminal potential stays bounded

The other assumptions. Theorems 3.1, 3.2 and 3.4 hold for every temporal operator, and each assumption excludes an explicit set. What each assumption is needed for and what it excludes:

AssumptionNeeded forWhat it excludesSize of the excluded set
Gaussian weights, independent, with mean μ/n\mu/n and variance σ2/n\sigma^2/nTheorem 3.1, exact at any widthother weight lawsthe modelling assumption of the paper
width to infinity at fixed MM, TT and depthTheorem 3.2finite widthmeasured distance to the limit 0.173 at width 32 and 0.009 at width 8,192
boundary-null (Assumption 1)continuity of the layer map despite the hard thresholda membrane coordinate of zero variance that sits exactly on a thresholdat most M(2T−1)M(2^T-1) threshold values for a fixed descriptor and operator (Lemma A.31)
positive conditional variances of the rate-projected membrane potentialsthe main form of the sharpened Theorem 3.4a potential that is an exact affine function of its temporal part, or deterministiccovered by the two further forms of the theorem (Theorem A.11a)
anti-concentration (Assumption 2)the earlier form of the bound (Theorem A.11)the same exceptional thresholdsholds with explicit constants elsewhere (Lemma A.32) and in 75 of 75 sampled configurations (Appendix A.5)
constants uniform along the trajectorythe finite-depth version of Theorem 3.4 (Theorem A.12)families of descriptors that approach the exceptional setan explicit hypothesis of that theorem
AP⊥=0\mathbf A\mathbf P_\perp=\mathbf 0Theorem 3.3 onlyevery continuously updated causal filtercharacterized above

The reviewer's derivation gave the paper an exact statement of what Theorem 3.3 covers, for which we thank the reviewer. The revision states it as Proposition 3.3a and adds the table of assumptions as Appendix Table 10.

3. "…an improvement to the mathematical presentation could make reading the paper far easier."

We agree with each of the suggestions, and the revision applies them throughout. The rule is that indices are subscripts in the order layer, neuron, input, time, that a hat marks finite width, and that the superscript (M)(M) marks the extension to MM inputs.

These suggestions give the paper a single rule for its indices and make Section 3 easier to read, and we thank the reviewer for them. The revision adds a notation table to the appendix (Table 4).

4. "I wonder if the material would not be better split into two: wide-SNN kernel theory plus convergence experiments and a 'when timing matters' investigation."

We considered this and kept one paper, because the kernel theory is what makes the timing question decidable. The rate criterion is a statement about the recursion of Theorem 3.2, namely whether it factors through the time-averaged part of the descriptor, so the second part has no object to test without the first.

We thank the reviewer for the suggestion, which led to a clearer division. The revision marks the two parts in Sections 3 and 4.

5. "Were there no other uses of AI that fall under the required to disclose list?"

There is one more use, and the revision adds it to the AI use statement. Beyond polishing the text, searching for related work and drawing the figures, which the submission declared, we used generative AI tools to audit the proofs for correctness. This is laborious checking work that does not change the contributions of the paper, and as the statement says we have reviewed all AI-assisted work and take responsibility for the final content.

We thank the reviewer for the question, which made the statement more complete.

6. "Do you have thoughts on how this theoretical foundation could (or could not) be used in theoretical neuroscience, where there is a similar rate vs. timing question?"

We think it gives that question a testable form for feedforward pathways, and that it stops at recurrent circuits.

  • What a population sees. In the theory a population with Gaussian weights sees the population before it only through the mean spike train over neurons, which is the population peri-stimulus time histogram, and the joint second moment across time bins (Theorem 3.1).
  • Link to mean-field theory. The need for a second-order variable beside the rate matches the exact mean-field reductions of spiking populations, which close only when a synchrony variable is carried along (Montbrió et al., 2015).

What the framework offers and where it stops:

Question in neuroscienceWhat the framework offersWhere it stops
does a downstream area read spike timing or only ratesthe question becomes whether the non-constant part of the second moment survives the synaptic filter, moves threshold crossings and reaches the readouta feedforward pathway with independent weights, since in a recurrent circuit the weights that act on a spike train also helped to produce it
at what temporal resolution is a code a timing codethe index Δ∗\Delta^\ast, the coarsest window that leaves the readout kernel unchanged within a tolerance, which gives the encoding time window of Theunissen & Miller (1995) a definition in terms of the networka fixed finite horizon
can equal firing rates hide a difference that matters downstreamTheorem 3.5, where populations with equal counts and equal histograms but different pairwise timing change the second moment of the firing rates one layer downstreamproved for the random-weight prior, a diagnostic for trained or biological circuits
how to test it in datashuffles of spike times that preserve counts, with the prediction that their effect follows the transmitted difference and the alignment with the task and not the amount of timing structurestructured connectivity and plasticity are outside the theory

We thank the reviewer for the question, which gave the paper a statement of what it offers to neuroscience and where it stops. The revision adds it to the discussion of neural coding in Section 2.

We are here for an in-depth discussion and further feedback. We would be glad to answer and address any issue you raise, and to keep improving the paper together with you.

Reviewer_mzAQOpenReview ↗
Rating
4(+2)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates when spiking neural networks rely solely on spike counts and when they truly exploit spike timing. For feedforward hard-LIF networks with independent Gaussian weights, the authors prove that the conditional distribution of inputs to the next layer depends only on the mean and temporal second moment of the spike trajectories from the preceding layer, and derive corresponding infinite-width temporal kernel recursions under fixed time duration and depth. The paper then maps count-based representations onto a time-constant subspace and presents sufficient conditions for exact and approximate rate simplification. Experiments examine finite-width convergence, rate approximation error, temporal separation, and readout effects. By means of transformed models, temporally encoded networks, and count-preserving temporal interventions, this work explores the diagnostic implications of these analyses in trained finite-size networks.

Critical strengths

The separation construction with equal spike counts and identical marginal distributions cleanly isolates the effect of temporal dependence structures. Theorem 3.5 preserves identical marginal probabilities at every time step, so the separation indeed arises from temporal dependencies. Building upon this, the authors prove that such differences can manifest in the second moments of downstream rate readouts. This conclusion is more definitive than merely showing distinct spike rasters or accuracy degradation after temporal shuffling.

Critical weaknesses
  1. There exists a gap between the rate-simplification conditions in the theory and actionable criteria for common causal networks. Specifically, the condition AP⊥=0AP_\perp=0 in Theorem 3.3 is clear yet strong: it requires the synaptic operator to completely discard all zero-mean temporal variations. For lower-triangular causal operators defined over a finite window, this heavily constrains outputs at early time steps and thus does not directly cover typical continuously updated causal filters.

  2. The approximation theorem better matches practical scenarios, but the bounds reported in the main text are roughly 397 to 13000 times larger than the measured error. The appendix further notes that the anti-concentration constant may grow rapidly with the time horizon. Therefore, while the current theory well explains why rate simplification can occur, it does not yet supply sufficiently practical error diagnostics. This limits the utility of these results as criteria for real networks.

Questions for authors

Regarding the rate simplification conditions in causal networks: Could the main text elaborate on what AP⊥=0AP_\perp=0 specifically means for lower-triangular causal operators? For the commonly used exponential synaptic filter, under what time durations and parameter ranges can the approximation bound provide non-trivial control over readout error? If the result primarily supports only qualitative interpretation, this should also be explicitly stated.

5 comments

We thank the reviewer for the precise reading of Theorems 3.3 to 3.5. Each point below led to a sharper statement in the paper.

1. "…does not directly cover typical continuously updated causal filters." and "Could the main text elaborate on what AP⊥=0AP_\perp=0 specifically means for lower-triangular causal operators?"

We partially agree. Theorem 3.3 indeed covers a single lower-triangular operator, the terminal accumulator, but continuously updated causal filters are the case of Theorem 3.4, which we may not have emphasized and clarify here.

  • The whole window. AP⊥=0\mathbf A\mathbf P_\perp=\mathbf 0 holds if and only if A=γv0⊤\mathbf A=\boldsymbol\gamma\mathbf v_0^\top with γ=Av0\boldsymbol\gamma=\mathbf A\mathbf v_0, where v0=T−1/21T\mathbf v_0=T^{-1/2}\mathbf 1_T is the constant unit vector of the paper, so all columns of A\mathbf A are equal. A lower-triangular A\mathbf A then reduces to γTeTv0⊤\gamma_T\mathbf e_T\mathbf v_0^\top, with eT\mathbf e_T the last unit vector, the synapse that delivers the window's spike count at its last step.
  • Windows of Δ\Delta steps (Appendix A.2.2). Let PΔ\mathbf P_\Delta project onto signals that are constant on windows of Δ\Delta steps and PΔ,⊥=I−PΔ\mathbf P_{\Delta,\perp}=\mathbf I-\mathbf P_\Delta. Then APΔ,⊥=0\mathbf A\mathbf P_{\Delta,\perp}=\mathbf 0 holds if and only if A=∑kγk1Wk⊤\mathbf A=\sum_k\boldsymbol\gamma_k\mathbf 1_{W_k}^\top, a counter over the windows WkW_k with indicator vectors 1Wk\mathbf 1_{W_k}, which is causal exactly when each count is delivered at or after the last step of its window.

What each kind of operator gets from the theory:

Temporal operator A\mathbf ACausalExact rate factorizationStatement that covers it
spike counter γv0⊤\boldsymbol\gamma\mathbf v_0^\top, for example the time average P0=v0v0⊤\mathbf P_0=\mathbf v_0\mathbf v_0^\topnoyesTheorem 3.3
terminal accumulator γTeTv0⊤\gamma_T\mathbf e_T\mathbf v_0^\topyesyes, the only causal caseTheorem 3.3
windowed counter ∑kγk1Wk⊤\sum_k\boldsymbol\gamma_k\mathbf 1_{W_k}^\top, each γk\boldsymbol\gamma_k supported at or after the last step of its window WkW_kyesyes, at resolution Δ\DeltaTheorem 3.3 at resolution Δ\Delta (Appendix A.2.12)
exponential filteryesnoTheorem 3.4 at resolution Δ\Delta (Appendix A.2.12)
identity, with the integrate-and-fire neuron of conversionyesnoProposition A.17, the rate readout closes up to O(1/T)O(1/T) when the terminal potential stays bounded

We thank the reviewer for this point, which gave Theorem 3.3 an exact statement of the operators it covers. The revision states it as Proposition 3.3a.

2. "…the bounds reported in the main text are roughly 397 to 13000 times larger than the measured error."

We partially agree, since this factor has two parts and only one of them was slack of the bound, which the sharpened Theorem 3.4 brings down to 1.7 to 2.3 times wherever the bound is below 1.

  • Slack. The bound over the mismatch probability that it bounds.
  • Cancellation. The mismatch probability over the descriptor error. The descriptor is an average over neurons in which spikes gained and spikes lost cancel, which is not slack and which no bound on the mismatch probability can see.

What changed in the theorem. The exponent goes from Etemp1/3\mathcal E_{\rm temp}^{1/3}, which came from Markov's inequality on ∥Z−Zdc∥2\lVert\mathbf Z-\mathbf Z_{\rm dc}\rVert^2, to Etemp1/2\mathcal E_{\rm temp}^{1/2}, the exponent measured for small perturbations. The new proof looks only at the first time step tt at which the two spike trains of an input aa differ.

  1. Same resets. Up to that step the reset histories agree, so the two membrane potentials differ by exactly Ya,tY_{a,t}, the transmitted temporal part as the membrane has accumulated it.
  2. Near a threshold. The rate-projected potential Xa,tX_{a,t} therefore lies within ∣Ya,t∣\lvert Y_{a,t}\rvert of one of the ∣Θt∣\lvert\Theta_t\rvert values that its threshold can take after the resets of earlier spikes.
  3. Density. The pair is jointly Gaussian, and the density of Xa,tX_{a,t} given Ya,tY_{a,t} is at most (2πva,t)−1/2(2\pi v_{a,t})^{-1/2}, with va,tv_{a,t} the conditional variance. A union over inputs, times and threshold values gives the bound below.

∥T(c)−Tdc(c)∥∞≤P(Ψ(Z)≠Ψ(Zdc))≤∑a,t∣Θt∣⋅E∣Ya,t∣2πva,t≤C⋅Etemp1/2,\lVert\mathcal T(c)-\mathcal T_{\rm dc}(c)\rVert_\infty \le \mathbb P(\Psi(\mathbf Z)\ne\Psi(\mathbf Z_{\rm dc})) \le \sum_{a,t}\frac{\lvert\Theta_t\rvert\cdot\mathbb E\lvert Y_{a,t}\rvert}{\sqrt{2\pi v_{a,t}}} \le C\cdot\mathcal E_{\rm temp}^{1/2},

with C=∑a,t∣Θt∣∥ℓt∥(2πva,t)−1/2C=\sum_{a,t}\lvert\Theta_t\rvert\lVert\boldsymbol\ell_t\rVert(2\pi v_{a,t})^{-1/2} where ℓt\boldsymbol\ell_t is the tt-th row of the leaky-integration matrix, with entries λt−q\lambda^{t-q} for q≤tq\le t. This constant is half that of Lemma A.32, with conditional variances.

Singular fields. When some va,tv_{a,t} vanishes, Theorem A.11a bounds its term by one of two further forms, so singular fields stay covered as in the submission.

How tight it is. The setting is the one-layer test of Figure 3b with the parameters of the supplementary code (T=4T=4, three inputs, λ=0.5\lambda=0.5, ρ=0.3\rho=0.3, ϑ=0.35\vartheta=0.35, μ=0.4\mu=0.4, and the synapse P0+εP⊥\mathbf P_0+\varepsilon\mathbf P_\perp, the time average plus ε\varepsilon times the rest). The result is below (the sharpened bound is the sum in the display, and the mismatch probability is measured under the coupling of Theorem 3.4):

ε\varepsilonDescriptor error, measured (×10−3\times10^{-3})Mismatch probability, measuredBound in the submissionSharpened boundBound over descriptor error, submission → sharpenedBound over mismatch probability, submission → sharpened
0.3224.00.3729.511.10397 → 4626 → 3.0
0.1611.90.2395.990.549506 → 4625 → 2.3
0.083.440.1443.770.2741,096 → 8026 → 1.9
0.040.4550.07832.380.1375,230 → 30230 → 1.8
0.020.1770.04001.500.06868,479 → 38837 → 1.7
0.010.0930.02020.9440.034310,149 → 36947 → 1.7
0.0050.0450.01020.5950.017113,218 → 38159 → 1.7

The two parts of the factor, and the sharpened bound in other settings (ranges over the rows above, and the further configurations take leak from 0 to 1, reset up to 0.5, thresholds from −0.2 to 1, zero mean, T=6T=6 and an exponential base operator):

QuantityValue
slack, the bound over the mismatch probability, submission → sharpened25 to 59 → 1.7 to 3.0, and 1.7 to 2.3 where the sharpened bound is below 1
cancellation, the mismatch probability over the descriptor error16 to 226
the bound is below 1, and so informative as a bound on a probability, submission → sharpenedε≤0.01\varepsilon\le0.01 → ε≤0.16\varepsilon\le0.16
the bound of the submission over the sharpened bound9 to 35
slack of the sharpened bound in twelve further configurations1.3 to 4.7
slack of the sharpened bound on two singular fields1.3 to 2.8

The mismatch probability bounds the total-variation distance between the spike-train laws of the network and of its rate projection, which is a stronger statement than a bound on the descriptor.

We thank the reviewer for insisting on this, which led to a bound with the exponent that the measured gap shows for small perturbations. The revision states Theorem 3.4 in this form, with the proof in Appendix A.2.6 and the first table above as Appendix Table 6.

3. "The appendix further notes that the anti-concentration constant may grow rapidly with the time horizon. Therefore, while the current theory well explains why rate simplification can occur, it does not yet supply sufficiently practical error diagnostics. This limits the utility of these results as criteria for real networks."

The submission did not stress that this growth belongs to the bound and not to the mechanism, and we clarify it here. With the measured density in place of the bound on it, the same expression gives the mismatch probability to within 4% for ε≤0.04\varepsilon\le0.04, which makes it a practical diagnostic in this regime, at the cost of one pass over the rate-projected network.

  • What the constant is. A bound on the density of the rate-projected membrane potential at its threshold, taken as one Gaussian term for each reset offset, that is, for each value that the threshold can take after the resets of earlier spikes.
  • What the mechanism depends on. At each step only the offset of the realized spike history is in play, so the density itself does not carry the count of offsets.

The number of offsets per input grows with the horizon TT (Appendix A.5):

RegimeReset offsets per input
no reset, ρ=0\rho=0TT
potential not carried over, λ=0\lambda=02T−12T-1
integrator of conversion, λ=1\lambda=1T(T+1)/2T(T+1)/2
intermediate membrane leak with reset2T−12^T-1

The setting is that of point 2, where the count is 15 per input, with the density estimated from 4×1064\times10^6 samples. The result is below (the expression is the density of potentials at threshold times the mean size of the transmitted temporal part):

ε\varepsilonSharpened bound, with the union over the 15 offsetsSame expression with the measured density at thresholdMismatch probability, measured
0.040.1370.08120.0783
0.020.06860.04060.0400
0.010.03430.02030.0202
0.0050.01710.01010.0102

For trained networks, where the theorems for Gaussian weights need not hold, Section 4 uses a criterion that needs no bound, a permutation of spike times at fixed counts. The result is below (change of the logits under the permutation, as a fraction of their norm):

NetworkChange of the logits
trained on a timing code0.92 to 0.99
converted or rate-trainedat most 0.08

We thank the reviewer for raising this, which led us to separate the density that the mechanism depends on from the count of offsets that only the bound carries. The revision adds the second table as Appendix Table 11.

Reviewer_RUrvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

A commonly cited potential advantage of spike based learning/ coding is the ability to utilise the exact spike timing of a neuron and relative spike timings of different neurons. In neuroscience, through rate based models or as described in the paper through an SNN to ANN conversion potentially task relevant information encoded in the precise timings of spikes is lost when we only use the number or rates of spiking. Thus it is natural to ask: When do spike timings matter? And this is the core question of the paper.

To do so the authors consider a feedforward network where weights are drawn iid from a gaussian distribution and Leaky Integrate and Fire Dynamics. They first prove that (in the finite neuron number limit) a layer depends upon the previous layer through only its mean spike count (m) and the temporal second moment (Q), thus the paper then focuses around comparing the true network dynamics with one depending upon a flattened version of these two statistics where we average over the full time horizon for every input, thus representing rate coding. As any information on when with the time horizon for each input a spike occurs is lost upon the described projection/ flattening.

The first step they make is considering the infinite width limit of the original set up, proving each neuron becomes independent (and not just conditionally independent upon the previous layer) and that the sufficient statistics (m,Q) evolve deterministically given the (m,Q) of the previous layer. They do so while still treating the hard reset in the LIF exactly. Next in theorems 3.3 and 3.4 they prove when the flattening and thus rate coding does not change the output of their deterministic kernel and provide an upper bound on the difference in the outputs respectively. Thus using the infinite limit result they manage to show when the rate coding projection doesn't change the network dynamics and upper bound these changes. Finally, in Theorem 3.5 they then provide an example of when spike trains with identical mean rates for the full time horizon and each bin lead to different temporal second moments, Q, and thus inducing different dynamics in the next layer. Then they finish with several experiments testing their predictions (P1-6), including checking their upper bound from theorem 3.4 or the infinite limit result.

Overall this paper aimed to explore when the rate projection in their model does and doesnt affect the network dynamics, and thus when the timing of spikes matters.

Critical strengths
  • The question is very well posed and of interest to a broad audience

  • The ability of theorem 3.2 to treat the hard LIF reset exactly is clever

  • All the proofs we checked contained no errors, thus we have confidence in the soundness of the theoretical results

  • The framework allows you to determine at which stage the timing information is lost, the synaptic filter, LIF or readout weights

Critical weaknesses
  • The paper is very poorly written, it took a considerable amount of time and effort to understand the text. The text makes very little attempt to explain the results derived, reading as extremely dense technical prose constantly introducing new jargon and requiring repeated referencing back to the appendix. The figures, core contributions and conclusions are very inaccessible on their own, meaning it would take an interested reader considerable time to understand the contributions of the paper. Perhaps to the authors credit we are still unsure if we understand the results to their full credit, for example I am unsure exactly what to make of Theorem 3.4 I of course understand the high level intuition of upper bounding the difference in the induced (m,Q) between when the previous layer is and isn't rate projected, but this bound seems relatively arbitrary and I am unsure how useful it is especially given later it is said the true value is of order 10,000 times smaller in the experiments. This is just one example of where some further intuition and understanding on what the results tell us and why they are of interest would be greatly appreciated. There are several notational issues throughout the paper which do not help the readability, for example the choice to write the paths in theorem 3.5 as {12,23,34,14} rather than for example {(1,2),(2,3),(3,4),(1,4)} caused us great confusion, why is the bound on ϵtemp\epsilon_{temp} presented under theorem 3.4 and not the more natural position above, why not mentioned the use of Markov equality here or hoeffdings for eq.4, why are the seemingly main predictions of the paper (P1-6) not included in the main text, only being referred to as P1-6, with little effort to explain their meaning, where even worse they aren't referred to as (P1-6) in the appendix thus finding them is very hard, the figures are completely inaccessible despite spending considerable time on the paper I do not understand all of them I am sure they are explained somewhere deep in the appendix but a good paper should have clear understandable results such that an interested reader can at a glance understand the core ideas of the paper and decided whether or not to read further, if met with figure 2 which is testing 'the limit' and P1 and P2 it is hard to see what can be understood. Overall, on this comment the overuse and poor referencing of the appendix and the extremely dense, hard to comprehend writing style are our main concerns regarding the presentation

  • The choice of iid gaussian weights and the infinite layer limit is what allowed the clever result of theorem 3.2, but we would appreciate a justification of these choices as they clearly do not reflect real networks and where central to the ability to avoid any smoothing or surrogates, We can appreciate the difficulty of the result, but we think the choices highlighted should be addressed and justified in the main text.

  • Theorem 3.3 and 3.4 both feel relatively ad hoc and uninformative though as discussed we may not appreciated the full significance due to the poor communication.

  • We are unsure what the main takeaways of this paper are, with this not being helped by either the contributions or the conclusions. Mechanistically, I have now taken the time to understand the core theoretically results from the paper but I am still unsure what this tells me about real networks and thus when/ why spike timings matter

Questions for authors
  • Please improve the presentation and communication of the results as discussed, with a better communication and thus appreciation of the results on our side we would consider improving our score.
4 comments

We thank the reviewer for checking the proofs and for the detailed reading. Each point below led to a concrete change in the paper.

1. "…this bound seems relatively arbitrary and I am unsure how useful it is especially given later it is said the true value is of order 10,000 times smaller in the experiments."

We partially agree. We apologize that the submission did not develop this bound further, which made its use hard to see, and we do so below. The bound is not arbitrary, and the revision brings it to within 1.7 to 2.3 times of the probability it bounds, wherever it is below 1.

Why it has this form. Two things must happen together for timing to change a spike, and the bound multiplies them for each input aa and time step tt.

  • Temporal drive reaches the membrane. Ya,tY_{a,t} is the temporal part of the drive that the membrane has accumulated by step tt, of mean size E∣Ya,t∣\mathbb E\lvert Y_{a,t}\rvert.
  • A potential sits that close to its threshold. The rate-projected potential has density at most (2πva,t)−1/2(2\pi v_{a,t})^{-1/2}, with va,tv_{a,t} its variance given Ya,tY_{a,t}, at each of the ∣Θt∣\lvert\Theta_t\rvert values that the threshold can take after the resets of earlier spikes.

∥T(c)−Tdc(c)∥∞≤P(the two spike trains differ)≤∑a,t∣Θt∣⋅E∣Ya,t∣2πva,t.\lVert\mathcal T(c)-\mathcal T_{\rm dc}(c)\rVert_\infty \le \mathbb P(\text{the two spike trains differ}) \le \sum_{a,t}\frac{\lvert\Theta_t\rvert\cdot\mathbb E\lvert Y_{a,t}\rvert}{\sqrt{2\pi v_{a,t}}}.

The left-hand side is the descriptor gap, the difference between the (m,Q)(\mathbf m,\mathbf Q) that the layer emits and the one that its rate projection emits. The display holds when every va,tv_{a,t} is positive, and Theorem A.11a covers the other cases.

What it is used for. No continuously updated synapse satisfies the exact condition of Theorem 3.3 (Proposition 3.3a), so for such synapses the paper's answer rests on this bound.

  1. One layer. A synapse that passes temporal structure with gain at most ε\varepsilon changes the spike trains of the next layer with probability O(ε)O(\varepsilon), at most 3.43ε3.43\varepsilon in the test of Figure 3b.
  2. Depth. Theorem A.12 carries the bound through the layers, so when the transmitted temporal part vanishes at every layer the readout kernel approaches that of the rate projection.
  3. Temporal resolution. Applied to averages over windows of Δ\Delta steps, it tells over which windows spikes can be averaged, which gives the network a temporal resolution Δ∗\Delta^\ast (Appendix A.2.12, with numbers for exponential synapses in Appendix Tables 7 to 9).
  4. What to measure. Timing acts through the part that the synapse transmits, not through the timing present in the spikes, which Section 4 tests (point 4).

How tight it is. The factor of 10,000 is the product of two effects, and only the first is slack of the bound. The second is cancellation, since the descriptor gap averages over neurons, in which spikes gained and spikes lost cancel, and no bound on the share of spike trains that differ can see this. The bound therefore controls more than the gap, the change of every statistic of the output spikes and not only of (m,Q)(\mathbf m,\mathbf Q).

The setting is the one-layer test of Figure 3b with the parameters of the supplementary code, where the synapse is P0+εP⊥\mathbf P_0+\varepsilon\mathbf P_\perp, the time average plus ε\varepsilon times the rest. The result is below (ranges over seven values of ε\varepsilon from 0.005 to 0.32, with the gap and the share measured):

One-layer test of Figure 3bBound in the submissionSharpened bound
Slack, the bound over the share of spike trains that differ25 to 591.7 to 3.0, and 1.7 to 2.3 where the bound is below 1
Cancellation, that share over the descriptor gap16 to 22616 to 226
Product, the bound over the descriptor gap397 to 13,21846 to 388
The bound is below 1, and so informative, forε≤0.01\varepsilon\le0.01ε≤0.16\varepsilon\le0.16
The bound shrinks with the perturbation asε2/3\varepsilon^{2/3}ε\varepsilon, as the measured gap does for small perturbations
Obtained fromMarkov's inequality on the whole temporal partthe Gaussian density at the first time step where the two spike trains differ

We thank the reviewer for pressing on this point, which gave the paper a sharper bound and a statement of what the bound is for. The revision states it as Theorem 3.4, with the proof in Appendix A.2.6.

2. "Theorem 3.3 and 3.4 both feel relatively ad hoc and uninformative though as discussed we may not appreciated the full significance due to the poor communication."

We may not have made clear what the two theorems are for, and we clarify it here. They are not ad hoc. They are the two halves of one answer, since Theorem 3.3 covers the synapses that remove spike timing exactly and Theorem 3.4 bounds how much timing can matter for every other synapse.

What each result covers (a spike counter hands every time step a fixed gain times the number of spikes in the window):

Synapse or neuronResultWhat it says
spike counter γv0⊤\boldsymbol\gamma\mathbf v_0^\top, for example the time averageTheorem 3.3timing cannot matter, exactly
causal spike counter, with the count released at the last stepTheorem 3.3the only causal synapse of this kind
counter over windows of Δ\Delta steps, with each count delivered at or after the last step of its windowTheorem 3.3 at resolution Δ\Delta (Appendix A.2.12)timing inside a window cannot matter
any other synapse, for example an exponential filterTheorem 3.4, on windows of Δ\Delta steps for the exponential filter (Appendix A.2.12)spike trains differ with probability at most the temporal part of the drive that reaches the membrane times a bound on the density of potentials at threshold
integrate-and-fire neuron of conversionProposition A.17the firing rate is the clipped summed drive up to the terminal membrane potential divided by ϑT\vartheta T
inputs with equal counts and equal per-step rates but different pairwise timingTheorem 3.5one layer separates them, also in the second moment of its output rates

This remark gave Section 3 a statement of what each theorem is for, and we thank the reviewer for it. The revision adds this table as Appendix Table 5.

3. "The choice of iid gaussian weights and the infinite layer limit is what allowed the clever result of theorem 3.2, but we would appreciate a justification of these choices…"

We agree that the main text should justify these choices, and the justification is that they are the setting in which the hard threshold and the reset can be treated exactly, at a finite-width cost that we measure at 0.009 for width 8,192.

  • Gaussian weights. The drive of the next layer is exactly Gaussian given the spikes of the previous layer, at any width (Theorem 3.1), so nothing has to be smoothed.
  • Infinite width. It serves one purpose, to make the descriptor (m,Q)(\mathbf m,\mathbf Q) deterministic (Theorem 3.2), as in the Gaussian-process limit of conventional networks (Lee et al., 2018).
  • Trained networks. The paper does not claim that they satisfy the theorems, and Section 4 uses the theory's quantities on them as diagnostics.

The setting is Experiment I of Appendix A.4.2 (depth three, T=4T=4). The result is below (relative Frobenius distance between a finite network's second moment and the limit recursion, and the relative kernel error at layers one to three when the hard threshold is replaced by a logistic function):

Width321285122,0488,192
Distance to the limit recursion ↓\downarrow0.1730.0750.0480.0210.009
ThresholdLayer 1 ↓\downarrowLayer 2 ↓\downarrowLayer 3 ↓\downarrow
logistic, slope 10.210.280.34
logistic, slope 100.0550.0850.106
logistic, slope 1,0000.00070.00110.0012

A smooth threshold reproduces the hard-threshold kernel only when it is nearly a step, which is why the analysis keeps the threshold exact.

We thank the reviewer for asking for this. The main text used these two choices without saying what each one buys, and the revision gives the justification where the model is introduced, in Section 3.1, which makes clear what the theory assumes and why.

4. "…I am still unsure what this tells me about real networks and thus when/ why spike timings matter"

Spike timing matters in a network when three conditions hold together, and each of them corresponds to a quantity that forward passes measure.

  1. The spike trains carry timing structure beyond their spike counts.
  2. The synapses pass that structure on.
  3. The neurons and the readout that follow respond to it.

When one of the three fails the network computes with rates, which is why a converted network can be replaced by its rate network while a network trained on a timing code cannot. The conditions, the result behind each and what the paper measures:

ConditionQuantityResult in the theoryWhat the paper measures (Section 4)
timing structure is presentthe non-constant part of Q\mathbf QTheorem 3.5inputs with equal counts and equal per-step rates stay separated at all 224 neuron settings
the synapse passes it onthe change of the second moment after the synapse when spike times are shuffled at fixed countsTheorems 3.3 and 3.4, through the transmitted temporal energy Etemp\mathcal E_{\rm temp}this transmitted difference, normalized for scale, explains the output change with R2R^2 0.83, the raw difference with 0.77
neurons and readout respond to itthe readout kernel and its alignment with the taskCorollary A.30permuting spike times at fixed counts changes the logits of timing-trained networks by 0.92 to 0.99 of their norm and those of converted and rate-trained networks by at most 0.08, and across 27 trained networks the benefit of timing follows its alignment with the task (r=0.73r=0.73) and not the amount of timing structure (r=0.12r=0.12)
conversion, where the neuron countsintegrate-and-fire neuronProposition A.1714 converted checkpoints recover their rate networks, for example ImageNet ResNet-18 with top-1 accuracy 0.698 against 0.698

We are grateful for this question, which led us to state the paper's answer in one place. The revision gives the three conditions in the introduction and in the conclusion.

5. "…why are the seemingly main predictions of the paper (P1-6) not included in the main text…", "the figures are completely inaccessible…" and "There are several notational issues throughout the paper…"

We agree with each of these points, and the revision follows all of them. The six predictions and where each is tested:

PredictionTested in
P1finite networks approach the limit recursion with widthFigures 2 and 3a
P2the gap between a network and its rate projection vanishes with the transmitted temporal energyFigures 3b, 3c, 4a and 4b
P3inputs with equal spike counts can give different outputsFigure 4c
P4the timing difference after the synapse, the transmitted difference, predicts the output change better than the raw oneFigure 5b
P5the benefit of timing follows its alignment with the task, not its amountFigures 5a and 5c
P6converted and timing-trained networks separate when spike times are permuted at fixed countsFigure 6

These suggestions make the main text and the figures readable on their own, and we thank the reviewer for them. The revision opens Section 4 with the predictions.

We are here for an in-depth discussion and further feedback. We would be glad to answer and address any issue you raise, and to keep improving the paper together with you.

Dear Reviewer RUrv,

We greatly appreciate your constructive feedback. Your questions on the bound of Theorem 3.4 and your comments on the presentation helped us sharpen the theorem and make the main text readable on its own.

As the Author-Reviewer Discussion period closes in a few hours, on November 18 AoE, we would be grateful if you could review our responses and the revised paper, and update your assessment if your concerns have been adequately addressed.

We hope that the additional evidence and clarifications provide sufficient grounds for a favorable reconsideration.

Sincerely,

The Authors

Comments on the paper

4 comments
AuthorsOpenReview ↗

Summary of the revision

We thank all reviewers for their valuable comments. Following them, the revision states which operators Theorem 3.3 covers, replaces the bound of Theorem 3.4 by one that is linear in the transmitted temporal part, separates the density that this bound depends on from the count of thresholds in its constant, gives the range in which it controls exponential synapses, lists what each statement covers and what each assumption excludes, states the predictions and the three conditions under which timing matters in the main text, and defines every object with its type under one notation convention. Following the later comment of Reviewer Ru8B, the second revision also restores the margins around figures and tables, fixes the missing spaces, sets Table 1 at the table size of the template, moves Figure 9 to the appendix, removes one clause from the setup and rewords four additions that read like replies. The revised paper is uploaded. Every change is marked in blue, Figures 1–21, Tables 1–3 and all theorems and equations keep their numbers, the new statements are Proposition 3.3a and Theorem A.11a, the new results are in Tables 4–11, Figures 3 and 10 are redrawn, and one row of Table 3 is updated (lines 729–731). Rows marked (new) are the changes of the second revision. The additions and where they are:

AdditionRaised byMain textAppendix
(a)The operators that Theorem 3.3 covers (Proposition 3.3a), also at block resolutionmzAQ, Ru8BSection 3.4, Proposition 3.3a; lines 233–238A.2.2; lines 1158–1171
(b)Bound of Theorem 3.4 linear in the transmitted temporal part, in place of the Etemp1/3\mathcal E_{\rm temp}^{1/3} boundRUrv, mzAQSection 3.4, Theorem 3.4; lines 247–267. Section 4, Figure 3(b); lines 368–370, 390–394A.2.6, Theorem A.11a, Table 6; lines 1282–1360. A.2.7, Figure 10; lines 1365–1370. A.2.8; lines 1441–1444
(c)Count of thresholds in the constant of the bound, and the measured density at thresholdmzAQSection 3.4; lines 261–264A.5, Table 11; lines 2504–2516
(d)Exponential synapses at finite temporal resolutionmzAQSection 3.4; lines 274–276A.2.12, Tables 7–9; lines 1507–1560
(e)What each statement coversRUrv, Ru8BSection 3.4; lines 223–224, 283A.2, Table 5; lines 1088–1105
(f)What each assumption is needed for and what it excludesRu8BSection 3.3; line 197A.5, Table 10; lines 2430–2453
(g)Why Gaussian weights and the wide limitRUrvSection 3.1; lines 138–140
(h)The three conditions under which spike timing mattersRUrvSections 1 and 5; lines 58–61, 538–539
(i)Predictions P1–P6 stated in the main textRUrv, Ru8BSection 4; lines 337–343A.4, experiments labelled P1–P6; lines 2077, 2116, 2155, 2196, 2230, 2240, 2246, 2331
(j)Figure captions that say what each panel plotsRUrvFigures 1–8; lines 67–68, 362–365, 390–394, 405–407, 425–427, 452–456, 467–470, 507–508Figure 9; lines 2317–2318
(k)Types of all objects and one notation conventionRu8B, RUrvSection 3.1; lines 129–135, 146–152, 161. Sections 3.2 to 3.5; lines 184, 200, 243–244, 298–299, 323–330Appendix A, Table 4; lines 756–795
(l)The two parts of the paper, kernel theory and rate criterionRu8BSections 3 and 4; lines 125–126, 337–338
(m)What the theory offers to neuroscienceRu8BSection 2; lines 104–107
(n)AI use statement, one use addedRu8BAI use statement; line 543
(o) (new)Missing spaces restoredRu8BSection 3.1; lines 129–130A.4, eight headings; lines 2077–2331
(p) (new)Margins restored around every figure and table, 8 to 10 pt to the text and 5 pt to the captionRu8BFigures 1–8, Tables 1–2All figures and tables
(q) (new)Table 1 set at the table size of the templateRu8BSection 4, Table 1; lines 442–454
(r) (new)Figure 9, the numerical checks of the integrator identity, moved to the appendix, with its paragraph citing it thereRu8BSection 4; lines 490–495A.4.9, Figure 9; lines 2317–2318
(s) (new)Clause on Theorems 3.3 and 3.4 removed from the setup, so that the operators Theorem 3.3 covers are stated once, with the theorem, and four other additions that read like replies reworded as statements of the paperRu8BSection 3.1; lines 131, 138–140. Section 3.3; lines 200–201. Section 3.4; lines 233–238, 261–265, 274–276

We are here for an in-depth discussion and would be glad to answer any further question.

AuthorsOpenReview ↗

Reminder of our replies and the revised paper

Dear Reviewers,

Thank you again for the careful reviews. Each of them changed the paper, and we would like to make sure that our replies have reached you.

A reply is posted under each review, and the revised paper is uploaded with every change marked in blue. In brief, the replies cover the following.

  • Reviewer RUrv. What the bound of Theorem 3.4 is for and how tight it is, why the theory uses Gaussian weights and the wide limit, and what the results say about when spike timing matters in real networks.
  • Reviewer mzAQ. What the condition of Theorem 3.3 means for lower-triangular causal operators, where the factor between the bound of Theorem 3.4 and the measured error comes from, and for which exponential synapses the bound gives non-trivial control.
  • Reviewer Ru8B. Which class of operators Theorem 3.3 covers, what each assumption of the theory excludes, and the definitions and notation that let the main text be read on its own.

The discussion period is close to its end. We would be grateful to hear whether the replies settle your concerns, and we would be glad to answer any further question while the discussion is open.

Sincerely, The authors

Area_Chair_zL1HOpenReview ↗

Reminder: Author Rebuttal Available

Dear Reviewers,

This is a gentle reminder that the authors have submitted their rebuttal for the paper. Please review it at your earliest convenience.

Thank you for your continued time and dedication to the review process!

Best regards,

AC of Submission 33455

AuthorsOpenReview ↗

Final summary: what the paper achieves

We thank the AC and the reviewers for the discussion. The paper asks when a spiking network depends on spike timing and when spike counts suffice. It first describes exactly what a layer passes on (a, b), then proves when this reduces to rates (c, d) and when it cannot (e), which gives the answer (f). It carries the answer to trained and converted networks (g, h) and to neuroscience (i), within a stated scope (j) and in a main text that reads on its own (k). Rows marked (revised) were sharpened in the discussion. After reading our responses and the revised paper, Reviewer Ru8B raised the score and writes that the paper could be accepted on content and would be interesting to many readers. What the paper achieves:

What the paper achievesEvidenceIn the reviewsWhere
(a)With Gaussian weights a layer passes on only the mean and the temporal second moment of its spike trains, at any width and with the hard threshold and the reset exactProved, and confirmed on finite networksTheorems such as this one "seem quite important" (Ru8B)Theorem 3.1, Figure 2
(b)In wide networks this becomes a kernel over spike trains, the spiking counterpart of the NNGP kernel, of which rate coding keeps the time-averaged partProved, and the distance of finite networks to the limit falls to 0.009 at width 8,192Treating the hard reset exactly "is clever", and "All the proofs we checked contained no errors" (RUrv)Theorem 3.2, Figure 2
(c) (revised)If the synapse removes all timing, the network computes with rates exactly. The synapses that do so are the spike counters, and the only causal one releases the count at the last stepProved, with the class as an if and only ifAsked by mzAQ and Ru8BTheorem 3.3, Proposition 3.3a
(d) (revised)For every other synapse, timing changes the spike trains with a probability bounded linearly in the temporal part that the synapse transmitsWhere below 1, the bound lies within 1.7 to 2.3 times of that probability, and with the measured density at threshold it gives it to within 4% for ε≤0.04\varepsilon\le0.04. For exponential synapses with a time constant of at least 2 steps it is informative on windows of up to 4 stepsAsked by RUrv and mzAQTheorem 3.4, Tables 6–9 and 11
(e)Conversely, timing can be irreducible. One layer separates two inputs with equal spike counts and equal rates at every step, even when only rates are read outA gap in closed form, positive at 224 of 224 neuron settings"more definitive than merely showing distinct spike rasters" (mzAQ)Theorem 3.5, Figure 4
(f) (revised)The answer. Timing matters only when the spike trains carry it, the synapses transmit it, and the neurons and the readout respond to itThe transmitted timing difference explains the output change with R2R^2 0.83, against 0.77 for the raw one. In 27 trained networks the benefit of timing follows its alignment with the task, r=r= 0.73, and not its amount, r=0.12r=0.12Asked by RUrv, who notes that the framework lets one "determine at which stage the timing information is lost"Sections 1 and 5, Figure 5
(g)On trained networks, a test that tells those that use timing from rate networksPermuting spike times at fixed counts sends timing-trained networks to chance and changes their logits by 0.92 and 0.99 of their norm, against at most 0.08 for converted and rate-trained networksFigure 6, Table 1
(h)Why conversion keeps the accuracy of the source network. The integrate-and-fire neuron counts, with an error of order 1/T1/T14 checkpoints from vision, language, speech, control and graphs recover their rate networks, for example ImageNet ResNet-18 at 0.698 against 0.698, and fail with leak or without reset, as predictedProposition A.17, Figures 7–8
(i) (revised)A testable form of the rate-versus-timing question of neuroscience, for feedforward pathwaysThe index Δ∗\Delta^\ast defines the encoding time window, and shuffles of spike times that preserve counts are the testAsked by Ru8BSection 2
(j) (revised)A stated scope. Every assumption has its use and the set it excludes, and trained networks are treated with diagnostics, not with theoremsGaussian weights are what allow the exact treatment of threshold and reset, and anti-concentration holds in 75 of 75 sampled configurationsAsked by RUrv and Ru8BSections 3.1 and 3.3, Table 10
(k) (revised)A main text that reads on its ownEvery object has its type where it first appears, under one notation, the predictions precede their experiments, captions say what each panel plots, and figures and tables have their marginsAsked by RUrv and Ru8BSections 3 and 4

We sincerely thank the AC and all reviewers again for the time and care given to our paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.