acceptodds
Under review as a conference paper at ICLR 2027

Making a Differentiable ANN-SNN Continuum using Spiking Threshold Gating Cells

Abstract

Spiking neural networks (SNNs), inspired by the brain and motivated by the potential for energy efficiency, replace real-valued activations with spikes generated by stateful neurons, but the discontinuity of spike generation complicates gradient-based training. Building on Threshold Gating (TG), which expresses neural nonlinearities as affine branches selected by sigmoid gates with tunable sharpness, we introduce Spiking Threshold Gating (), a parameterized cell enabling a differentiable transition from stateless ANN activations to stateful spiking neurons. Unlike prior ANN-to-SNN conversion methods, allows ANN pretraining and subsequent hardening within the same cell, using exact gradients of the smooth cell during training rather than surrogate gradients. We then introduce a curriculum that progressively sharpens the cell’s gates during training, achieving state-of-the-art results on several event-based datasets. We additionally propose a novel input-routed parameterization that separates membrane-driven firing from input-dependent event payloads, further improving performance through flexible, multi-branch emissions.

Primary area:
optimization
Keywords:
SNNs, Neuromorphic computing, Training, Backpropagation, gradients, gradient-based optimization, Spiking neural networks

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kRAMOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a novel Spiking Threshold Gating (STG) technique that bridges the gap between ANN activations and spiking neurons through different parameter configurations within a unified framework.

Critical strengths
  1. The design of STG is novel, particularly in its ability to unify ANN activations and spiking neurons within a single framework.

  2. The authors provide extensive theoretical analysis of the proposed STG framework, which helps establish the theoretical foundation of the proposed ANN–SNN continuum.

Critical weaknesses
  1. Most of the main experimental results in Table 3 are obtained using graded or multi-bit quantized emissions rather than strictly binary spikes. Although graded emissions may improve gradient propagation and representational capacity, they differ substantially from conventional binary spiking neurons. Therefore, direct comparisons with SNNs using strictly 0–1 outputs may not be entirely fair, as the proposed models benefit from additional representational capacity.

  2. Since the authors claim that "ANNs with typical activation functions and canonical spiking neurons become part of a differentiable continuum," experimental evidence is needed to validate this claim. However, the current experiments primarily compare the proposed method with SNNs trained using surrogate gradients, which does not directly demonstrate the effectiveness of the proposed ANN–SNN continuum. Comparisons with conventional ANN-to-SNN conversion methods would provide stronger support for this central claim.

Questions for authors

Please see the Weaknesses part.

3 comments
AuthorsOpenReview ↗

Response to the respected reviewer (kRAM)

We thank the reviewer for the positive assessment of the unification of ANN activations and spiking neurons and of the theoretical analysis. We answer the two weaknesses.

Graded and multi-bit emission against binary spikes. We agree that binary and graded emission should be compared separately. First, two of the comparisons in Table 3 (Section 6) are already strictly binary. On MSWC-FSCIL, the only change to the NeuroBench network is the neuron: STG replaces RadLIF and the emission stays binary, as with RadLIF. The accuracy rises from 75.27% to 83.32% (+8.05 points), and it is 83.54% with an exact Heaviside gate (Table 7, Section 6.4). On DVS-Gesture, the DTS-SNN row is also binary: 95.83% against 94.44% for the same network with LIF.

Second, where Table 3 uses graded emission, binary or graded output is a setting of the same cell, so we trained both on the same networks (Table 10, Appendix C, 3 seeds).

binary, curriculum (ours)binary, fixed + STEgraded, curriculum (ours)
DVS-Gesture, MS-ResNet-2094.32 ± 1.00—95.71 ± 0.44
SSC81.78 ± 0.3078.81 ± 0.2284.40 ± 0.24
SHD94.06 ± 0.8295.05 ± 1.06 (2 of 3 seeds diverge)96.04 ± 0.60

Binary emission costs 1.4 to 2.6 points against graded emission, which quantifies the additional capacity that the reviewer points to. The binary models are still above binary baselines: 94.32% against 87.37% for LIF in the same MS-ResNet-20, and 81.78% against 80.69% for DCLS-Delays on SSC. Finally, when the trained gate of a binary model is replaced by an exact Heaviside, the accuracy is kept (Table 7, Section 6.4), and the firing pattern shows that every channel sends either zero or one fixed value (Figure 3, Appendix C). Section 4 gives the reason (lines 260–262).

Evidence for the ANN–SNN continuum, and comparison with ANN-to-SNN conversion. We would like to clarify one point first. The submitted version already contains a direct demonstration of the continuum, beyond comparisons with surrogate-gradient SNNs: the CIFAR10-DVS model starts from an ImageNet-pretrained ANN (ConvNeXt-Small), its activations are replaced by STG cells, and the sharpness curriculum carries it to spiking emission (94.10 ± 0.17%, and 93.47 ± 0.32% with 4-bit emission; Section 6.1). We did not state this clearly enough. We now say it explicitly in the contributions (lines 78–79), in Section 3 (line 241) and in Section 6.1 (lines 382–384), and we added a curve of accuracy against sharpness for this model (Figure 4, Appendix C).

However, we agree that a comparison with ANN-to-SNN conversion strengthens this, and we added one with MST [36] (Table 4, Section 6.1). MST converts an ImageNet-pretrained ANN and reaches 88.12% on CIFAR10-DVS with T=512. It is not fully spiking, as its released code keeps a real-valued softmax in the attention (marked in Table 4). Yet STG is better: a ConvNeXt-Tiny of the same size (28M) reaches 91.23 ± 1.60% with T=16. From the same ImageNet weights, a standard LIF neuron trained with surrogate gradients reaches 87.30 ± 2.38% against 92.43 ± 0.81% for STG. So the continuum, and not only the backbone, carries the pretrained ANN into the spiking regime. The same holds at ImageNet scale: a pretrained ConvNeXt-Tiny with STG cells reaches 71.07% on I2E-ImageNet with 8-bit emission and an exact Heaviside gate, against 60.50% for the published MS-ResNet-34 with binary LIF trained from scratch [18] (Table 3, Section 6; details in Table 11, Appendix C).

Reviewer_kRAMOpenReview ↗

After Rebuttal

Thank you for the response. My concerns have been fully addressed. After carefully reading the rebuttal and the other reviewers' comments, I have decided to maintain my original rating.

Thank you for reading our rebuttal and confirming that your questions were addressed

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

This paper proposes Spiking Threshold Gating (STG), a parameterized neural cell that aims to establish a differentiable continuum between conventional ANN activations and spiking neuron dynamics. Building upon Threshold Gating (TG), STG introduces membrane and adaptation states, sigmoid-gated affine state updates, and configurable output branches. The authors further propose a sharpness curriculum that progressively transitions the cell from smooth to sharp gating, enabling training with exact gradients of the smooth computation rather than surrogate gradients. An input-routed emission mechanism is also introduced to decouple membrane-controlled firing from input-dependent output amplitudes. The method is evaluated on several event-based benchmarks, including SHD, SSC, CIFAR10-DVS, DVS-Gesture, and MSWC-FSCIL. The authors report competitive classification performance and investigate quantized emissions and event rates.

Critical strengths

The proposed STG formulation integrates membrane dynamics, adaptive thresholds, and configurable emission branches within a common mathematical framework. The explicit parameter settings for recovering conventional activation functions and several spiking neuron models provide a useful conceptual connection between ANNs and SNNs.

The paper includes several experiments that suggest the proposed cell can improve performance over conventional neuron formulations. In particular, the same-width SHD comparison and the matched SSC comparison provide preliminary evidence of the method's potential. The theoretical derivations and detailed implementation descriptions also contribute to the clarity of the proposed framework.

Critical weaknesses

I find the novelty of the proposed method somewhat limited. STG mainly extends sigmoid-gated affine representations with recurrent states, while related work such as UltraLIF has already explored differentiable spiking neurons trained without surrogate gradients. The theoretical results largely concern parameter equivalences and differentiability, rather than establishing a clearly distinct training mechanism. My main concern is the lack of experimental evidence for the claimed ANN-to-SNN continuum. Although the authors show that ANN activations and spiking neurons can be represented by different configurations of the same cell, they do not demonstrate that a pretrained ANN can actually be transformed into a hard-spiking SNN while retaining its performance. This makes the practical significance of the proposed continuum unclear.

Several experimental comparisons make it difficult to attribute the reported improvements to STG itself. For example, the CIFAR10-DVS results use an ImageNet-pretrained ConvNeXt-Small, while the compared methods differ substantially in architecture and training settings. The benefit of the sharpness curriculum is also not consistent across experiments. Although it slightly improves performance on SSC, Appendix C shows that fixed sharpness achieves better accuracy than annealing on DVS-Gesture (86.62% vs. 83.84%). In addition, the contribution of input routing is not evaluated through a controlled ablation. I am also concerned that the SHD models are selected based on test accuracy (Appendix B), which may lead to overly optimistic results.

Finally, the evaluation is limited to relatively small event-based datasets, without experiments on large-scale benchmarks such as ImageNet. While the authors use ImageNet-pretrained weights for CIFAR10-DVS, this is not equivalent to evaluating STG on ImageNet itself. Given that the paper proposes STG as a general framework connecting ANNs and SNNs, I would like to see evidence that the approach remains effective on larger and more challenging tasks.

Questions for authors
  1. The ANN-to-SNN continuum is mainly illustrated through different cell configurations. Have the authors tried starting from a pretrained ANN and progressively transitioning to a hard-spiking STG model? It would be helpful to see how accuracy and spike sparsity change during this process, and whether additional adjustments to the state dynamics or output branches are needed.

  2. How much of the reported improvement comes from the STG cell itself, rather than the sharpness curriculum or input routing? A comparison with UltraLIF and standard surrogate-gradient LIF under the same architecture and training budget would help clarify this. I would also like to understand why fixed sharpness performs better than annealing on DVS-Gesture in Appendix C.

1 comment
AuthorsOpenReview ↗

Response to the respected reviewer (UaHJ)

We are grateful to the reviewer for the careful review, our response is below:

Novelty and UltraLIF. UltraLIF also trains without surrogate gradients. Its mathematics, its methodology and its cell are different from ours. UltraLIF's membrane is a log-sum-exp, so the input can only raise the membrane, and reset and adaptation are not recovered by parameter settings. Our point is not just differentiability and annealing alone, it is also a continuum, the STG cell: the same STG cell represents ANN activations on one hand and spiking neurons on the other (Table 2, Section 3), and the ramp moves a network from one to the other. Even coming to annealing, the base maths is different: UltraLIF uses ultradiscretization (max-plus algebra), which requires an expression without subtraction in positive variables, while STG sharpens sigmoid gates over affine branches with signed coefficients. To demonstrate, as requested, we compared in two ways. First, in UltraLIF's own released code, with only the neuron swapped, STG is more accurate on all three benchmarks, e.g. 72.73 vs. 57.95% on DVS-Gesture (Table 8, Section 7, T=1). Second, in the same architecture and training budget (3 seeds), LIF, UltraPLIF and STG give 87.37, 81.82 and 95.71% on DVS-Gesture (MS-ResNet-20) and 94.72, 82.68 and 96.04% on SHD.

From a pretrained ANN to a spiking SNN. This was the reviewer's main concern, and it is exactly what the CIFAR10-DVS experiment in the submitted version does; it is a major point of our work. To make it explicit, we added it to the contributions (lines 78–79), Section 3 (line 241) and Section 6.1 (lines 382–384). In the CIFAR10-DVS experiment that we reported in the submitted version, we do start from an ImageNet-pretrained ConvNeXt-Small, replace its activations by STG cells, and train along the ramp as the model turns from an ANN to becoming an SNN. No additional adjustment of the state dynamics or the output branches was needed. With an exact Heaviside in place of the trained gate, the quantized model gives 91.80% at 9.6% events (Table 7, Section 6.4). We have added a sharpness-accuracy curve in the appendix, which was not in the submitted version (Figure 4, Appendix C).

Large-scale benchmarks. The reviewer also questioned if our methodology can be scaled for bigger datasets. We do not see a reason why it would not, we have added a new result for ImageNet to demonstrate the scaling (Table 3; Table 11, Appendix C). On I2E-ImageNet (event-based ImageNet, T=8) [18], we take an ImageNet-pretrained ConvNeXt-Tiny (29.1M parameters), replace its activations by STG cells, and train along the ramp. We still used a pretrained backbone, because we see that as our point, that we can traverse from a pretrained ANN and go to an SNN. The model reaches 72.50% with the trained gate and, after one more pass with a sharper gate, 71.55% with an exact Heaviside gate. With 8-bit emission and the exact Heaviside gate it reaches 71.07% (70.03% at 16.0% events), and a pretrained ResNet-34 (22.0M) reaches 65.75%. The best published result is 60.50% (MS-ResNet-34 with binary LIF, trained from scratch); our models are pretrained and emit graded values, so this is evidence of scaling, though not a fully matched comparison.

Attributing the improvement. We agree that separate ablation experiments to isolate the contributions from the input-routed or curriculum with exact gradients is needed. Firstly, we have demonstrated why each contribution is valuable: curriculum is what unlocks going from pretrained ANNs to SNNs, exact gradients with annealing give a better network than surrogate gradients, and this holds even if a vanilla spiking neuron (with Heaviside) is the destination (on MSWC-FSCIL, STG in place of RadLIF gives 83.54% with an exact Heaviside against 75.27% published; Tables 3 and 7, Section 6), the input-routed gives a boost in accuracy (with tradeoff with energy consumption). The ablation study is shown here: For the training, the ramp and the gradient are crossed on the same cell (Table 5, Section 6.2). With exact gradients, annealing adds 0.5 to 2.2 points. At fixed sharpness, the exact gradient adds 1.1 to 3.8 points over the surrogate. For the cell, STG against LIF in the same network is given above (and 92.43 vs. 87.30% on CIFAR10-DVS, Table 4, Section 6.1). For the input routing, switching it on gives 96.64 vs. 94.45% on SHD and 84.40 vs. 83.11% on SSC.

Fixed sharpness vs. annealing (Appendix C). This is addressed in our response to Reviewer KMwX (first point): with a ramp spread over several epochs, annealing matches or exceeds fixed sharpness (Table 9).

SHD model selection. Regarding the concern about test-accuracy selection on SHD: there is no validation split, so this is common practice. We also report the average over epochs 100 onward, which needs no selection: 95.65 ± 0.46%, only 0.4 points below the selected 96.04% (Appendix B, lines 750–751).

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

This paper introduces Spiking Threshold Gating (STG), a parameterized neuron model intended to unify conventional ANN activations and SNN inference within a common differentiable formulation. Building on Threshold Gating (TG), STG progressively increases gate sharpness during training, allowing optimization to begin with smooth, differentiable dynamics and gradually approach hard threshold decisions. The authors argue that this enables training with exact gradients of the smooth forward computation, avoiding surrogate-gradients. They also introduce input-routed emissions, in which the membrane determines branch selection while the current input supplies the emitted payload, providing a direct gradient path that bypasses the membrane recurrence.

The paper reports competitive accuracy and experiments with quantized emissions, including 4-bit and 8-bit configurations. The results suggest that smooth-to-sharp training is a promising approach to training SNNs, although the extent to which the improvements generalize across architectures and tasks remains to be established.

Critical strengths
  • The paper is generally well organized. The progression from the STG formulation to the ANN–SNN continuum, the sharpness curriculum, and the experimental evaluation is logical. The distinction between gate selection and emitted payload is useful for understanding the proposed input-routing mechanism.

  • Unified parameterization for ANNs and SNNs: The paper provides a common formulation for stateless activations, leaky integration, reset mechanisms, adaptive thresholds, and different emission types. This is a useful modeling abstraction that facilitates systematic investigation of neuron dynamics.

  • A smooth-to-sharp training strategy: The proposed curriculum provides a way to train a differentiable cell while progressively approaching discrete branch selection.

  • Diverse emission configurations: The quantized-emission experiments are a useful first step toward examining the trade-off between accuracy, emission precision, and event activity.

Critical weaknesses
  • The curriculum results are configuration-dependent. The matched SSC experiment supports the proposed smooth-to-sharp approach, but the additional DVS-Gesture ResNet-20 experiment favors fixed sharpness over annealing. It is not clear where and how the progressive sharpening is beneficial.

  • The experimental comparisons do not consistently isolate the proposed method. The CIFAR10-DVS result uses an ImageNet-pretrained ConvNeXt-Small and achieves 93.47% with quantized emissions. This is substantially different from the architectures and training conditions of the listed baselines. Though promising, the result does not establish that STG itself explains the performance difference. Similar concerns apply to comparisons involving different training budgets, temporal resolutions, and emission configurations.

  • Although the paper includes hard-gate baselines and reports configurations that become binary in the sharp limit, it is unclear whether the principal curriculum-trained models are evaluated using actual hard-threshold inference. Since the main results are generally reported at finite sharpness, a direct evaluation of the trained models with hard threshold decisions would better establish the connection between smooth training and the claimed spiking regime.

Questions for authors
  • Contribution of the curriculum: In the SSC experiment, the curriculum improves accuracy over hard-gate training with surrogate gradients, while the DVS-Gesture ResNet-20 experiment favors fixed sharpness over annealing. How much of the observed benefit is attributable to progressive sharpening itself, as opposed to using smooth gates during training?

  • CIFAR10-DVS comparison: How do the results change when ConvNeXt-Small is trained without ImageNet pretraining, and can the authors compare STG against a strong baseline under the same pretraining, data split, temporal resolution, and quantization conditions? This would help determine how much of the improvement is attributable to STG rather than the backbone and pretraining.

  • Graded payloads: How does the same network with binary emissions compare with inference using graded outputs? Although graded payloads are supported by some neuromorphic platforms, it would still be beneficial to see the performance-quantization tradeoff.

1 comment
AuthorsOpenReview ↗

Response to the respected reviewer (KMwX)

We thank the reviewer for the feedback, our answers to the concerns and questions are as follows:

Configuration Dependence and Appendix C DVS-Gesture Experiment (ResNet-20). We stated a possible limitation to be a very abrupt sharpening of the curriculum, which causes the training to diverge. Our audit of the Appendix C results showed that as the underlying reason: the ramp of the sharpness completed too quickly. However, we additionally probed why other experiments did not show that sensitivity to the sharpness climb. We discovered from prior literature that ResNet-20 with the neuron after the residual addition (the form used in Appendix C) is especially prone, since it cannot realize an identity mapping (Fang et al., SEW ResNet, 2021; Hu et al., MS-ResNet, 2024). To prove this point, we carried out an additional experiment with a ramp spread over several epochs (a gentler climb of the sharpness τo\tau_o), on the same network and on another variant of ResNet-20 avoiding the issue with the previous topology (MS-ResNet-20) (Table 9, Appendix C, 3 seeds):

annealfixed
DVS-Gesture, ResNet-2086.9986.62
DVS-Gesture, MS-ResNet-2095.71 ± 0.4493.56 ± 1.65
SHD96.04 ± 0.6095.56 ± 0.47
SSC84.40 ± 0.2483.70 ± 0.25

The MS-ResNet-20 on the other hand shows a clear gain from annealing, and so do SHD and SSC. The ablation of how much gain comes from progressive sharpening and how much from exact gradients is further mentioned in Table 5 (Section 6.2).

Consequently, we modified line number 406–407 to read (now lines 421–422): On ResNet-20 with DVS-Gesture, a ramp that completes within a fraction of an epoch ends below fixed sharpness (83.84 vs. 86.62%), whereas a gentler ramp spread is better, although, to restate, this sensitivity to ramp is for ResNet-20. We are sorry that a lack of clarification of this part raised concerns, but we are at the same time grateful to the reviewers for allowing us to get to the bottom of this.

CIFAR10-DVS and Crediting to Backbone. We see the CIFAR10-DVS result differently. Starting from a pretrained ANN backbone is the purpose of this experiment, not a bug: it shows that the continuum carries a pretrained ANN into an SNN, which unlocks the large zoo of pretrained ANN models for spiking networks. This is what a continuum from ANN activations to spiking neurons is for (Table 2, Section 3). To make this explicit, we added a line in the contributions, in Section 3 (line 241) and in Section 6.1. That is why we do not see the point of training from scratch as we would anyways train STG as an ANN then traverse the continuum towards an SNN.

That said, it may be asked: we ought to compare with prior methods that also start from an ImageNet-pretrained backbone with a similar number of parameters. For that, we add the results from MST paper[36] which also attempts ANN-to-SNN conversion (using a backbone pretrained with ImageNet) and it reported an accuracy of 88.12% and uses 512 time steps and 28M parameters, in comparison, we also trained ConvNeXt-Tiny with 28M parameters but T=16 only, our accuracy is 91.23 ± 1.60% as compared to 88.12%. Note that MST is not fully spiking (its released code keeps a real-valued softmax in the attention at every time step, now marked in the table) and it needs 512 time steps, and yet STG scores better. We have added this as a small table (now Table 4, Section 6.1) in the paper. The same table additionally has the ablation experiment that the reviewer asked for: a standard LIF neuron in the same model (ConvNeXt-Small), with everything else as in our run, reaches 87.30 ± 2.38%, against 92.43 ± 0.81% for STG, quantifying the gain by the STG cell.

Hard Threshold. The reviewer also asked for an evaluation with a fully hard (exact Heaviside) emission gate. Firstly, we have argued that for many cases of actual hardware deployment, the exact Heaviside (infinite τo\tau_o) is an overkill: depending upon if the circuit is analog or digital, a certain value of the emission sharpness τo\tau_o would be fully hard (Section 4, Eq. 8, we slightly rephrased and added few lines to make it clearer). For example, with 4-bit emission the gate is indistinguishable from a hard one beyond a margin of 3.37/τo3.37/\tau_o, i.e. 0.34 at τo=10\tau_o=10 and 0.07 at τo=50\tau_o=50 (lines 263–265). However, upon the request of the reviewer, we also evaluate the trained models with an exact Heaviside in place of the emission gate, without retraining (Table 7, Section 6.4). Section 4 gives the reason why this works: for binary emission, the transmitted bit is the same as a Heaviside spike at any sharpness (lines 260–262, with the conclusion stated at lines 274–276).

With the exact Heaviside, the binary models keep their accuracy, and the 4-bit graded models lose 0.55 points on SHD and 1.2 on CIFAR10-DVS (Table 7, Section 6.4). For binary against graded emission, see our response to Reviewer (kRAM).

Comments on the paper

1 comment

We thank the reviewers for their reviews. Their feedback allowed us to add results that strengthen the conclusions of the paper. The central point of the paper is a continuum: the same STG cell represents ANN activations on one hand and spiking neurons on the other, and a curriculum that sharpens its gates moves a network from one to the other. This addresses the training problem of spiking networks, the discontinuity of the spike, with exact gradients instead of surrogate gradients, and it lets an SNN start from a pretrained ANN. These gains are verified experimentally, with further results added during the rebuttal. The cell also has an input-routed form, in which the membrane decides the firing and the emitted value is read from the input, and its effect is also quantified.

The concerns of the reviewers broadly fall into four areas, which we address with clarifications and new results. All changes in the revised paper are marked in red, and the details are in the individual responses. We will upload a final version before the closure of the rebuttal period.

1. Novelty with respect to UltraLIF (UaHJ). The two methods differ in their mathematics and in their scope. We also compare them directly, in UltraLIF's own code and in the same architecture.

2. From a pretrained ANN to an SNN, and scaling (UaHJ, KMwX, kRAM). We add a comparison with ANN-to-SNN conversion and with a LIF neuron from the same pretrained weights. We also add an ImageNet-scale result: 71.07% on I2E-ImageNet with 8-bit emission and an exact Heaviside gate.

3. What contributes to the results (KMwX, UaHJ, kRAM). We add ablations of the ramp and the exact gradient, of binary against graded emission, and of the input routing. We also repeat the Appendix C experiment with a ramp spread over several epochs, where annealing matches or exceeds fixed sharpness. That was raised by two reviewers but we have clarified it with additional experiments.

4. Hard-threshold inference (KMwX, UaHJ). We evaluate the trained models with an exact Heaviside in place of the emission gate, without retraining. The binary models stay within one point, and the quantized graded models lose about one point or less.

We believe that the revision addresses the concerns raised. Some of them came from points that the submitted version did not state clearly enough, such as the transfer from a pretrained ANN to an SNN, and these are now explicit and supported by new results. We believe that this work addresses a central difficulty of SNN training, the non-differentiable spike, and that it opens pretrained ANNs to spiking networks. We kindly ask the reviewers to take the revised paper and our responses into account in their final assessment, and to reconsider their scores if they find their concerns resolved. We remain available for any further questions or experiments during the rest of the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.