acceptodds
Under review as a conference paper at ICLR 2027

Applied but Not Received: Designing Effective Surrogate Gradients for Spiking Neural Networks

Abstract

Surrogate gradients are the standard tool for training spiking neural networks (SNNs), yet their true effect on the learning signal remains poorly understood. We show that the surrogate gradient (SG) applied at the spike nonlinearity is not the effective surrogate gradient (ESG) the network receives once it is propagated through the complete adjoint of the leaky integrate-and-fire (LIF) dynamics under backpropagation-through-time. Characterizing the ESG analytically, our analysis shows how common applied surrogate gradient definitions engineered to be smooth and symmetric yield effective surrogate gradients that are neither: discontinuous at threshold and asymmetric around it. These findings are consequential: the discontinuity inflates gradient variance, and the asymmetry biases learning toward non-spiking states, revealing an unintended mismatch between the forward and backward models. We turn this analysis of ESG dynamics into a design principle: rather than reshaping the applied surrogate alone, we design for the effective surrogate directly—shaping the applied surrogate and the LIF adjoint together—so that the effective gradient the network receives is continuous at threshold and restores intended credit assignment between silent and spiking regimes. The resulting design enables more stable and performant training of deep SNNs, demonstrating design of the effective surrogate gradient as a principled route to improved optimization.

Primary area:
optimization
Keywords:
Spiking neural networks, surrogate gradients, backpropagation-through-time, leaky integrate-and-fire, adjoint sensitivity, credit assignment, gradient variance, surrogate design, neuromorphic learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_hFpiOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper revisits an important issue in surrogate gradient (SG) training for spiking neural networks (SNNs): the surrogate gradient applied at the spike nonlinearity differs from the Effective Surrogate Gradient (ESG) actually received by the network after temporal backpropagation through LIF neuron dynamics. Through theoretical analysis, the authors reveal that conventional ESGs may exhibit discontinuities and asymmetry around the firing threshold. Based on these findings, they propose cdx−(t⋆)cdx-(t^\star), which redesigns temporal gradient propagation to improve credit assignment. Experiments on multiple neuromorphic classification tasks demonstrate the effectiveness of the proposed method.

Critical strengths
  1. A novel and theoretically valuable perspective. This paper shifts the focus of surrogate gradient research from local SGs to the ESGs resulting from temporal backpropagation. The proposed gradient propagation mechanism is motivated by this analysis, establishing a clear connection between the theoretical findings and the method design.
  2. Experimental results support the effectiveness of the method. The authors systematically analyze the ESGs produced by different surrogate gradients and backward propagation rules, and validate the proposed method across multiple neuromorphic datasets and network configurations, with particularly notable improvements in deeper networks.
Critical weaknesses

The proposed cdx−(t⋆)cdx-(t^\star) involves multiple design components, including gradient selection around threshold crossing and adjustments to gradient propagation during silent steps. Additional ablation studies on these components could help clarify their individual contributions.

Questions for authors

The analysis in this paper primarily focuses on LIF neurons with hard detached reset. Could the authors briefly discuss whether the proposed ESG analysis and method could be extended to other reset mechanisms (e.g., soft reset) or other neuron models?

3 comments
AuthorsOpenReview ↗

Response reviewer hfpi

Thank you for your careful assessment and your constructive suggestions. We are glad that you see the effective-surrogate perspective as novel and theoretically valuable, and we address your two points below.

On ablations. We agree that ablations make the contribution clearer. We intend to ablate the components that define cdx-t⋆ across the full experimental grid to isolate the source of its gains. We aim to report each variant next to the full method.

On resets. Our main study considers spiking LIF neurons with a hard reset. For resets the discontinuity has two sources: detaching the reset in the backward pass and the discontinuous reset itself in the forward pass. A hard reset sets the membrane to a fixed value after a spike, while a soft reset subtracts a fixed amount from it. In the backward pass a detached hard reset cuts the credit at each spike, while a detached soft reset carries it across spikes. A new Appendix D therefore covers hard and soft resets, each detached and attached. The effective-surrogate analysis there reveals that the discontinuity persists across these settings and that cdx-t⋆ keeps the credit more balanced than the conventional adjoint, most clearly under the hard reset. The effective surrogate thus shows how each choice of reset shapes the credit a neuron receives. We intend to summarise these reset results in the main matter of the revision.

On other neuron models. Our work studies the effective surrogate gradient that arises in spiking LIF neurons. Within this family our formulation nests binary neurons at β = 0 and approaches integrate-and-fire neurons as β tends to 1. In these neurons the applied surrogate covers only the step from the spike StS_t to the membrane UtU_t. The effective surrogate instead captures how the full adjoint transforms this gradient. Under the theoretical framework of §2.3 this analysis extends to any unit with a recurrent hidden state hth_t. A surrogate is needed only for the non-differentiable spike. For smooth units such as the GRU the analysis yields the effective gradient through the full adjoint. In both cases the effective (surrogate) gradient becomes

Geff(h)=E[λts ∣ ht=h].G_{\mathrm{eff}}(h)=\mathbb{E}\big[\lambda^{\mathrm s}_t \,\big|\, h_t=h\big].

Here λts\lambda^{\mathrm s}_t is the gradient that the unit passes upstream through its input. Mapping it requires a good exploration of the distribution of hth_t. The analysis also carries over to multi-dimensional states such as the two states of an oscillatory spiking neuron or the cell and hidden state of an LSTM. However, for such multi-dimensional adjoints the exploration becomes harder through the curse of dimensionality. The effective (surrogate) gradient thus extends to other neuron models as a tool for analysis. By contrast, our proposed cdx-t⋆ uses this tool to design an adjoint specifically for LIF neurons. We intend to describe this wider use of the analysis in the revision.

We hope this answers your questions. During the remaining rebuttal period we hope to share further insight into what drives the gains of cdx-t⋆.

Thanks for the response. In the additional experiments in the appendix, the authors discuss the results under soft reset. We can see that the proposed method does not significantly improve the gradient issues in this setting, possibly because the gradient discontinuity is already small. In addition, the authors committed to providing complete ablations, and I hope these can be added to the paper soon. This does not change my view that this is a valuable work, and I maintain my original rating.

AuthorsOpenReview ↗

Follow up reviewer hfpi (extended & ablation results)

Thank you for your reply and for maintaining your rating (our new public comment gives an overview of the revised and extended results). We agree with your reading of the soft reset. In simulation the conventional effective surrogate shows little distortion there when neurons fire often (Appendix D). The empirical results across all four reset variants show the same: cdx-t⋆ is ahead of or level with the conventional adjoint, and pooled over datasets its advantage is smallest where the reset leaves little to correct (Figure 19 and Table 22 in Appendix F.2). As promised, the main matter now summarises these reset results (Sections 3.2 and 3.3).

You noted the connection between our theoretical findings and the method design. The analysis traces the defect of the conventional adjoint to the gradient that time steps without a spike pass back to earlier steps. States just below threshold collect this gradient. Our simulations show this as a jump at threshold with more gradient below it (Sections 2.4 and 3.1). This points to three components of an adjoint. cdx-t⋆ adjusts all three:

  1. What a time step without a spike passes back in time.
  2. How much of its own gradient it sends to earlier layers.
  3. At which time step the surrogate of a spike is read.

The first two are the adjustments at time steps without a spike you named, and the third is the gradient selection around the threshold crossing.

The complete ablations of Appendix E cover 13 adjoint designs that change these components individually and in combination, across the experimental grid of Section 3.2 with ten seeds each. Figure 16 in Appendix E illustrates the structure of these designs. Of the adjustments at time steps without a spike, passing back only the gradient of the spike carries the gain in out-of-sample accuracy. Alone it already raises accuracy on every dataset. Sending part of a step's own gradient to earlier layers remains necessary for most runs to learn. The gradient selection around the threshold crossing changes accuracy little. cdx-t⋆ is consistently strong among the ablated designs, with none ahead of it on more than one dataset (Figure 17 and Table 15 in Appendix E). Because its ablation maps all three components alone and in combination, cdx-t⋆ serves as a reference point for the joint design of applied surrogate and adjoint. As you hoped, the paper now includes these ablations and builds its discussion on them (Sections 3.2 and 3.3).

In short, the connection you noted between our theoretical findings and the method design becomes measurable. Overall, the effective-surrogate analysis is the core contribution of our work. It diagnoses the conventional adjoint, separates the components an adjoint can change and indicates which change pays off. cdx-t⋆ adjusts all three, and its ablation measures what each component contributes.

We hope these additions to the manuscript answer your remaining points. Thank you again for your constructive engagement.

Reviewer_zes7OpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces the concept of the effective surrogate gradient, which takes into account that the surrogate gradient is propagated through the LIF adjoint such that its effective action differs from that of the original surrogate function. It shows how to change the surrogate functions such that the effective surrogate gradient is continuous and symmetric. Finally it shows that the learning performance for thus optimized networks is higher.

Critical strengths

The new concept of the ESG is very interesting, its usefulness and functionality are convincingly demonstrated.

Critical weaknesses

None.

Questions for authors

Why is in Fig. 3 for beta=0.5 a remaining discontinuity at threshold? Does this harm performance? Is it avoidable? Does the estimation of the ESG for a specific network and task cause some significant computational overhead for the task?

2 comments
AuthorsOpenReview ↗

Response Reviewer zes7

Thank you for your positive assessment and for these precise questions. We are glad that you found the concept of the effective surrogate interesting and its usefulness convincingly demonstrated. We regret that the presentation required substantial effort. In response to your presentation score, we have restructured the methodology and describe the changes in our public comment.

On the remaining jump. We can trace the jump you noticed at β = 0.5 to how cdx-t⋆ credits states just below threshold. Under the revised rule in our public comment it falls from 0.16 to a lower but nonzero 0.11 (Table 2). Its cause is the same under both rules and lies in how cdx-t⋆ credits the delay to the next spike. In the discrete-time LIF we evaluate, this delay cannot shrink smoothly to zero across threshold. A state just above threshold spikes now, while a state just below it spikes τ ≥ 1 steps later. Under cdx-t⋆ both receive a floor of 1−β1-\beta at once and a share β\beta through their spike. The leak then discounts that share for each additional step of delay, like a delayed reward:

1-\beta+\beta\cdot 1=1$$ $$\text{just below }\theta:\quad 1-\beta+\beta\cdot\beta^{\tau-1}\widetilde G\le 1$$ Here the delay discount $\beta^{\tau-1}$ and the surrogate value $\widetilde G$ selected at that spike are both at most one. The shortfall below threshold is the part of the spike share lost to the delay. With a spike at the next step nothing is lost, and Appendix A shows that the jump of cdx-t⋆ vanishes exactly in that case. At β = 0 nothing is carried and no jump arises. At β = 0.5 about a third of the states just below threshold spike two or more steps later (App. A.2.3). Despite this residual, cdx-t⋆ shows how adjoint redesign through the lens of the effective surrogate improves on the conventional adjoint in most configurations (Table 3). The redesign targets the whole credit profile rather than the jump alone. cdx-t⋆ splits its credit nearly evenly across threshold, whereas the conventional adjoint places two thirds of it below threshold (Table 2: 48/52 against 67/33 at β = 0.5). Our simulations compute the effective surrogate from a given applied surrogate and adjoint. Removing the jump entirely would require the reverse: choosing the applied surrogate, the adjoint or both so that a continuous effective surrogate is received. $$\text{Our simulations:}\quad G_{\mathrm{eff}}^{\mathrm{received}} \leftarrow G^{\mathrm{applied}},\ \text{adjoint} \quad \text{(analysis)}$$ $$\text{Exact removal:}\quad G_{\mathrm{eff}}^{\mathrm{desired}} \rightarrow G^{\mathrm{applied}},\ \text{and/or adjoint} \quad \text{(inverse)}$$ Both directions depend on θ, β and the membrane distribution. Our simulations match its mean and variance to $\mathcal{N}(\mu_u,\sigma_u^2)$. During training, estimating the quantities that adequately describe the membrane distribution (e.g. $\mu_u,\sigma_u^2$ above) for every neuron and step is computationally expensive. Additionally, we believe that a dependency on them in adjoint design likely incurs a high variance cost due to noisy or inaccurate estimation. Therefore, the proposed cdx-t⋆ aims to rebalance the credit profile cheaply and deterministically by relying solely on those elements already directly accessible to the conventional adjoint. The explanation for the jump has been included in our revised version and we intend to include the remaining aspects in further revisions. **On computational overhead.** In this work, the effective surrogate gradient is considered an offline analysis tool and is *never* computed during training. Guided by the effective surrogate, we propose cdx-t⋆ as an alternative to the conventional adjoint. Each backward step of the conventional adjoint already holds the two membrane values that cdx-t⋆ compares (Fig. 1). Thus cdx-t⋆ adds no memory cost. Like the conventional adjoint it needs a single surrogate-gradient evaluation per spike, here at the step that one added selection picks. We intend to summarise both costs in a prospective table. We hope these answers are helpful. During the remaining rebuttal period we hope to offer further insight into the cost of cdx-t⋆.
AuthorsOpenReview ↗

Follow up reviewer zes7 (extended & ablation results)

As a follow-up, we share new empirical results that extend the demonstration you found convincing (our new public comment summarises the revision). In the main experiments cdx-t⋆ improves on the conventional adjoint in most cells (Section 3.2). The new empirical results extend the analysis to further settings and ablate where this improvement comes from.

The effective surrogate describes how the gradient is transformed as it passes through the full LIF adjoint. It first locates the cause of the conventional jump. Under the conventional adjoint, time steps without a spike also pass their own gradient back to earlier time steps. States just below threshold collect this extra gradient. In simulation this opens the jump at threshold and places more gradient below it than above it (Sections 2.4 and 3.1). The analysis thus points to three components of an adjoint. cdx-t⋆ is designed to adjust all three:

  1. What a time step without a spike passes back in time.
  2. How much of its own gradient it sends to earlier layers.
  3. At which time step the surrogate of a spike is read.

Appendix E ablates 13 adjoint designs that change these components individually and in combination, each on the full experimental grid of Section 3.2 with ten seeds. Figure 16 in Appendix E illustrates each component and how the designs combine them. The gain in out-of-sample accuracy follows the first component, the one that removes the cause of the jump. Changing only this component already raises accuracy on every dataset. The other two matter less: one is needed for most runs to learn, and the other changes accuracy little. cdx-t⋆ is consistently strong among the ablated designs, with none ahead of it on more than one dataset (Figure 17 and Table 15 in Appendix E). Because its ablation maps all three adjustments alone and in combination, cdx-t⋆ serves as a reference point for the joint design of applied surrogate and adjoint. Appendix F extends the experiments to five further settings, including the other reset variants. In each cdx-t⋆ is ahead of or level with the conventional adjoint (Figure 18 with Table 17 in Appendix F.1 and Figure 19 with Table 22 in Appendix F.2). This bears on your question whether the jump harms performance: accuracy improves when the cause of the conventional jump is removed. Pooled over datasets, the advantage is also largest where the simulations of Appendix D show the most distortion. Section 3.3 now states this link.

Taken together, the usefulness you noted goes beyond seeing the gradient a neuron receives to designing the adjoint that shapes it. This makes the effective-surrogate analysis the core contribution of our work. It locates the defect of the conventional adjoint, splits the adjoint into components that can be changed separately and shows which change pays off. cdx-t⋆ adjusts all three, and its ablation shows where the gain comes from.

The revised PDF reflects all of this, and we hope it adds to our earlier answers. We thank you again for your precise questions.

Reviewer_8koEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper analyzes the effective surrogate gradient (ESG) obtained through temporal backpropagation in LIF neurons. Under hard reset with reset detachment, it identifies threshold discontinuity and asymmetry in the ESG and proposes the cdx-(t^\star) backward rule to mitigate them. Experiments on ECG, SHD, and DVS Gesture vary network depth and width and report accuracy improvements in most configurations.

Critical strengths

The paper examines the interaction between surrogate gradients and temporal backpropagation. Analyzing the gradient actually propagated through the network, beyond the shape of the local surrogate function, provides a useful perspective.

Critical weaknesses
  1. The harmful effects of the identified problem and the mechanism of improvement are insufficiently established. Synthetic experiments demonstrate ESG asymmetry and discontinuity, but do not establish that these properties cause inappropriate credit assignment or degrade actual training. In particular, the relationship between the independent, positive gradient signals used in the analysis and gradients during training remains unclear. Measurements of ESG, gradient variance, and firing rates during training, together with component ablations or comparisons controlling gradient magnitude, are needed to identify the source of improvement. The conditions for the continuity guarantee are also unclear. Equation (10) uses weighted selection, whereas Eqs. (17)–(18) in Appendix A rely on an unweighted maximum and additional assumptions. Table 2 also reports a residual jump of 0.16 for the proposed method at (\beta=0.5). The authors should specify when continuity holds for the actual rule and distinguish reducing the discontinuity from eliminating it.

  2. Generalization across reset settings, model scales and architectures, and tasks is insufficiently evaluated. The central derivation depends on the (\beta(1-S_t)) term induced by hard reset with reset detachment. Soft reset or differentiation through reset changes the temporal gradient. Whether the same problem and remedy persist under these settings remains unverified. Comparisons with representative alternatives, including soft reset used in prior direct-training methods [1, 2], are needed. The models in §3.2 are limited to LIF stacks with depths (L\in{2,3,4}) and hidden widths (H\in{32,64,128}). Varying depth and width within this range does not establish effectiveness in deeper or structurally different SNNs. Evaluation on deeper spiking CNNs and Transformers, including layer-wise ESG and gradient analysis, is needed to determine how the identified problem and its mitigation change with scale. ECG, SHD, and DVS Gesture cover different input modalities, but all are classification tasks. Effectiveness under other output and loss formulations remains unverified. Generalization across architectures and tasks requires further validation.

Questions for authors
  1. Does the continuity result in Eq. (18) hold for Eq. (10) with (\eta=1/2)? Please clarify the additional assumptions and explain the residual jump in Table 2.

  2. Are there existing analyses linking reduced ESG asymmetry or discontinuity to gradient variance or performance improvements during training? Are there results isolating the effects of individual components?

  3. What happens when reset detachment is disabled or soft reset is used? Are relevant experiments or analyses available?

2 comments
AuthorsOpenReview ↗

Response reviewer 8koE

Thank you for your critical review and for the opportunity to strengthen our work. We regret that the presentation required substantial effort; in response, we restructured the methodology (see our public comment).

Critical weakness 1.

  • 1a (Establishment of harmful effects): Prior work focuses on the applied surrogate gradient, generally assuming a symmetric shape and balanced credit. To our knowledge, we are the first to analyse the credit received after the full LIF adjoint, the effective surrogate gradient. As the reviewer mentions, our synthetic experiments show that for widely used methods it is asymmetric and discontinuous around the threshold. The reviewer justly asks: Why are these properties harmful?

    First, at a practical level, surrogate gradients are designed to be symmetric and continuous, yet yield effective surrogate gradients that are not. Although such a design could in theory learn better, the search for the best surrogate gradients and crediting schemes can only succeed if researchers know what they are actually designing.

    Second, at a more formal level, the surrogate gradient should be a useful backward proxy for the non-differentiable Heaviside function, whose common smooth approximations have symmetric and continuous derivatives around the threshold. The asymmetry unbalances credit assignment, leading to larger parameter adjustments below the threshold, and the discontinuity adds unintended gradient variance around it. Moreover, the conventional adjoint's credit (Eq. (8)) can peak far in time from the spike that influences the loss: a timing mismatch between the backward and forward pass.

    Third, at an empirical level, cdx-t⋆, which improves continuity and symmetry, performs better in most settings on three main neuromorphic benchmarks (Table 3), suggesting that these properties hamper SNN training. In the remainder of the rebuttal, we intend to ablate the sources of this improvement and share the results.

    We will make this reasoning explicit in the manuscript.

  • 1b (Requested extended analysis, Question 2): The reviewer questions whether the synthetic results generalize to real-world data, and requests ablations, which we agree would be insightful (see 1a). The effective surrogate gradient is a property of the neuron and its adjoint, not of a dataset. Under the assumptions of §2.3 and for a given applied surrogate, it depends only on the leak, the threshold and the membrane potential distribution, so a dataset shapes it only through these statistics. Synthetic data lets us set them directly and measure the effective surrogate gradient over the full membrane potential domain. Real-world data with the same statistics would not change its shape, but may cover some membrane regions sparsely, if at all.

  • 1c (Positive-gradient assumption): Our analysis does not require a positive gradient. It only assumes that the average gradient is stable within an inter-spike interval and independent of the membrane potential (§2.3). The credit profile's shape is then invariant: a negative average flips it; a different magnitude rescales it.

  • 1d (Continuity of cdx-t⋆, Question 1): We thank the reviewer for noting that, in the submission, Eq. (18) was derived for the unweighted maximum and does not hold for Eq. (10) at η=1/2\eta=1/2. cdx-t⋆ now takes exactly this maximum, so the continuity result holds for the rule as used (under the assumptions of 1c). The jump at β=0.5\beta=0.5 in Table 2 falls from 0.16 to 0.11, against 0.36 for the conventional adjoint. A state just above threshold spikes now and receives full credit (1−β)+β=1(1-\beta)+\beta=1, whereas one just below spikes τ≥1\tau\geq1 steps later and receives (1−β)+β⋅βτ−1G~(1-\beta)+\beta\cdot\beta^{\tau-1}\widetilde G, its spike share discounted by the leak per step of delay beyond the first. The jump thus vanishes exactly when all such states spike at the next step (App. A.2.3), so cdx-t⋆ reduces the discontinuity without removing it, as the revision and Limitations now state.

Critical weakness 2.

  • 2a (Resets, Question 3): With resets, the discontinuity has two sources: detaching the reset in the backward pass and the discontinuous reset itself in the forward pass. A new Appendix D covers hard and soft resets, each detached and attached, showing that the discontinuity persists across these settings and that cdx-t⋆ keeps the credit more balanced than the conventional adjoint, most clearly under the hard reset.
  • 2b (Empirical scope): We deliberately train LIF networks on the three neuromorphic benchmarks that are currently the community standard, for direct comparability. With the applied surrogate gradient fixed, accuracy differences trace back to the adjoint alone. Our Limitations (§3.3) set out this scope; the revision names the settings the reviewer raises as natural next steps.

We hope this response and the revised PDF address your concerns. Thank you again.

AuthorsOpenReview ↗

Follow up reviewer 8koe (extended & ablation results)

As a follow-up to our response, we report the empirical results announced in 1a (see our new public comment for an overview of the revised and extended results). In the main experiments cdx-t⋆ improves on the conventional adjoint in most cells (Section 3.2). The new empirical results extend analysis to alternative hyperparameters and reset variants, as well as ablate where improvements of cdx-t⋆ come from (Question 2).

The effective surrogate describes how the gradient is transformed as it passes through the full LIF adjoint. Under the conventional adjoint, time steps without a spike also pass their own gradient back to earlier steps. States just below threshold collect this extra gradient. In simulation the effective surrogate therefore jumps at threshold and puts more gradient below it (Sections 2.4 and 3.1). The analysis thus points to three separate components of an adjoint. cdx-t⋆ demonstrates design for the effective surrogate and adjusts all three:

  1. What a time step without a spike passes back in time.
  2. How much of its own gradient it sends to earlier layers.
  3. At which time step the surrogate of a spike is read.

Appendix E ablates 13 adjoint designs that change these components individually and in combination, across the full experimental grid of Section 3.2 with ten seeds each. Figure 16 in Appendix E illustrates these designs and what each component changes. The ablation shows that the gain in out-of-sample accuracy follows the first component, in which the analysis locates the defect. Changing only this component of the conventional adjoint already raises accuracy on every dataset. The second component is also needed. A time step without a spike must still send part of its own gradient to earlier layers, or most runs never learn. The third component sets the time step at which the surrogate of a spike is read. It changes accuracy little. cdx-t⋆ is consistently strong among the ablated designs, with none ahead of it on more than one dataset (Figure 17 and Table 15 in Appendix E). Because its ablation maps all three components alone and in combination, cdx-t⋆ serves as a reference point for the joint design of applied surrogate and adjoint. As announced in 1a, Section 3.2 now reports these findings and Section 3.3 reflects on them.

Appendix D extends the effective-surrogate analysis in simulation to all reset variants: hard and soft resets, each detached and attached. Together with the main experiments, Appendix F now answers Question 3 empirically for all four. Under every reset variant cdx-t⋆ is ahead of or level with the conventional adjoint on each dataset (Figure 19 and Table 22 in Appendix F.2). Pooled over datasets, its advantage is largest where Appendix D finds the effective surrogate of the conventional adjoint most distorted. Section 3.2 now reports these settings.

Overall, we believe that the analyses presented reinforce our contributions. Its core is the effective surrogate as an analysis tool that works in three steps. It diagnoses the conventional adjoint, identifies the separate components an adjoint can change and shows where a change pays off. cdx-t⋆ changes all three, and its ablation attributes the gain to the component the analysis singles out.

With these additions to the manuscript, we hope to have addressed the reviewer's concerns and questions. We thank the reviewer again for the critical assessment that allowed us to strengthen our work.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised PDF summary

With this public post we intend to describe the changes in the revised PDF. Two of the reviews found that our presentation required substantial effort, and we take this to heart. We have therefore restructured the methodology and dedicated the additional page to it, including new visual aids. Mindful of the load that reviewers carry during the discussion period, we have otherwise kept the main matter close to the reviewed version. Beyond the updated results, the remaining changes clarify the scope of our claims and extend the Limitations. New results go to the appendix.

Methodology. Every adjoint in the methodology (§2) is now presented in one consistent structure. A new figure shows the credit profile of each adjoint. Pictograms illustrate the features of the effective surrogate that the paper discusses: its discontinuity, its displaced modes and its imbalance of credit across threshold. They replace the appendix figure on how to read the shape scores (Figure 5 of the submission). With the new figure, Figures 2–4 of the submission are now Figures 3–5.

cdx-t⋆. One review pointed out that in the submitted version Eq. (18) was derived for the unweighted maximum and does not hold for Eq. (10) at η = 1/2. Following this point, cdx-t⋆ now takes exactly this maximum. The continuity result then holds for the rule as used, and the rule no longer has a hyperparameter to tune. We have updated all simulation and empirical results to this rule. To make the empirical results more robust, they now average over ten seeds instead of five and over the last 10% of training instead of the final epoch. The main conclusion holds: cdx-t⋆ yields higher out-of-sample accuracy and faster, steadier training in most configurations. In simulation the discontinuity at β = 0.5 in Table 2 falls from 0.16 to 0.11, against 0.36 for the conventional adjoint.

Appendix. Appendix A follows the restructured methodology. For each adjoint it gives the definition, the derivation of its effective surrogate and its bounds in one common format. A new Appendix D analyses the effective surrogate in simulation across four reset variants: hard and soft resets, each detached and attached. Its figures use a more expansive format than those in the main matter and sweep the membrane mean across firing rates. For continuity with the version under review, the effective-surrogate figures in the main matter keep their original format.

We hope these changes make the methodology easier to follow. Our answers to individual points are in the respective review threads.

AuthorsOpenReview ↗

Revised PDF summary - inclusion of new analyses (extensions & ablations)

This comment follows up on our first comment on the revised PDF. The reviewers asked for ablations and broader experiments, and we ran them to test the three contributions of our submission:

  • Analytical: the complete LIF adjoint turns smooth and symmetric applied surrogates into discontinuous and asymmetric effective surrogates.
  • Design: we design for the effective surrogate directly, with cdx-t⋆ as its realisation.
  • Empirical: cdx-t⋆ trains more accurately than the conventional adjoint in most configurations.

The new analyses reinforce each of them:

  • Simulation: Appendix D carries the effective-surrogate analysis to all four reset variants and shows how strongly each distorts the credit profile of the conventional adjoint.
  • Ablation: across 13 adjoint designs, the empirical gain follows the component of cdx-t⋆ that removes the cause of the conventional discontinuity and asymmetry in the credit profile (Appendix E).
  • Extensions: under alternative hyperparameters and reset variants, cdx-t⋆ is ahead of or level with the conventional adjoint on every dataset (Appendix F).
  • Supporting evidence: pooled over datasets, the empirical gains follow the component that the effective-surrogate analysis ties to the discontinuity and asymmetry in the credit profile and are largest under the reset variants where it finds them strongest.
  • Outlook: cdx-t⋆ demonstrates three ways to intervene in the adjoint that the effective-surrogate analysis exposes, and we believe its ablation makes it a reference point for next steps in the joint design of applied surrogate and adjoint.

The evidence behind these points required extending the experiments of the submission. We added 22 new configurations, each evaluated across the complete experimental grid of Section 3.2 with ten seeds. Together they comprise 5,940 additional networks, more than seven times the 810 networks of the submitted analysis. The revised PDF incorporates these results as follows:

Main matter, results and discussion. The results and the discussion of the main matter have been revised to incorporate the insights from these extended analyses. Section 3.2 now summarises the ablations and the extensions. Together with these extended and ablated analyses, the discussion in Section 3.3 emphasises how the effective surrogate analysis contributes to more informed surrogate gradient design. Both point to the new appendices for the full results.

Appendix E, Empirical analysis: ablations. cdx-t⋆ differs from the conventional adjoint in three separable components: the silent-step scale, the spike-step gate and the time rule. Appendix E ablates these components individually and in combination, in 13 configurations from the conventional adjoint to cdx-t⋆. Figure 16 illustrates the structure of these configurations. For Appendix E, Figure 17 and Table 15 summarise the ablations. cdx-t⋆ is consistently strong among the ablated designs, with none ahead of it on more than one dataset. Because the ablations map what each component contributes alone and in combination, cdx-t⋆ serves as the reference point named above.

Appendix F, Empirical analysis: extensions. The empirical analysis of the main matter takes the hard detached reset as its core case. Appendix F extends this core setting to alternative hyperparameters in F.1 (a steeper applied surrogate and a fixed leak) and to alternative reset variants in F.2 (the hard reset attached, and the soft reset detached or attached). The steeper surrogate moves here from Appendix C. In each setting it compares the three adjoints. For Appendix F.1, Figure 18 and Table 17 summarise the alternative hyperparameters: the advantage of cdx-t⋆ is retained. For Appendix F.2, Figure 19 and Table 22 summarise the alternative reset variants: the pooled advantage of cdx-t⋆ is largest where the simulation analysis across reset variants in Appendix D finds the effective surrogate of the conventional adjoint most distorted.

Our answers to individual points draw on these results in the respective review threads.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.