acceptodds
Under review as a conference paper at ICLR 2027

How Reliable Is Online Training Through Time? A Convergence and Diagnostic Framework for Spiking Neural Networks

Abstract

Spiking neural networks (SNNs) offer remarkable energy efficiency, yet their training typically relies on backpropagation through time (BPTT) with surrogate gradients, whose memory consumption grows linearly with the number of time steps. Online Training Through Time (OTTT) avoids full temporal backpropagation by maintaining presynaptic activity traces and performing same-time spatial backpropagation, enabling direct training with memory cost independent of the number of time steps. However, whether its biased updates can still provide reliable optimization guarantees for the original discrete-spike objective remains an open theoretical question. In this work, we establish a non-asymptotic convergence framework for fixed-within-sequence OTTTA in feedforward LIF SNNs. We introduce a randomized-smoothed spiking model as an analytical reference and use its full soft-BPTT gradient as the differentiable benchmark. The systematic OTTTA bias separates into two components: a hard–soft mismatch controlled by a threshold-tail functional measuring near-threshold membrane activity, and a temporal-decoupling bias quantified by the gradient discrepancy between soft OTTT and full soft-BPTT. Under explicit smoothness, stochastic-variance, trajectory-stability, and restricted-PL assumptions, we derive a finite-step bound directly on the hard-spike risk, relative to a restricted minimizer of the smooth reference risk, with explicit optimization, variance, hard–soft, temporal-decoupling, and risk-transfer terms. Across different static and event-based benchmarks, the two approximation discrepancies exhibit distinct task-dependent behavior, while mechanism-specific checkpoint diagnostics show pronounced seed-level associations with accuracy on several tasks. Together, these results connect a finite-step convergence guarantee with practical diagnostics for characterizing the approximation behavior of OTTT.

Primary area:
optimization
Keywords:
Spiking Neural Networks, Online Training Through Time, Surrogate Gradients, Convergence Analysis

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_A9VZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the convergence behavior of Online Training Through Time (OTTT) for spiking neural networks, focusing on its fixed-parameter accumulation variant (OTTTA). The authors introduce a smoothed spiking model as an analytical reference and decompose the bias of OTTTA into two components: the hard–soft mismatch between discrete and smoothed spike trajectories, and the temporal-decoupling error between soft OTTT and full soft-BPTT. Under several regularity and stability assumptions, they derive a non-asymptotic bound on the hard-spike risk that explicitly accounts for these two sources of approximation error. The paper also proposes diagnostics for measuring these discrepancies and evaluates their behavior on CIFAR-10, CIFAR-100, CIFAR10-DVS, and DVS Gesture. The experiments examine discrepancy magnitudes, their evolution during training, their correlations with classification accuracy, and the effect of the smoothing scale.

Critical strengths

The paper separates hard–soft mismatch from temporal-decoupling error and incorporates both into a finite-step risk bound. This provides a useful framework for understanding different sources of gradient approximation error in OTTTA.

And the authors examine how the two discrepancies evolve during training and vary across datasets. The multi-seed experiments and additional sensitivity analyses provide some interesting observations about the relationship between gradient approximation errors and model performance.

Critical weaknesses

The main theorem relies on several strong assumptions, particularly the restricted PL condition and hard–soft trajectory stability. However, the paper does not discuss whether these assumptions hold, even approximately, for the networks used in the experiments. The resulting bound also contains residual terms related to the threshold-tail functional and temporal-decoupling discrepancy, but it is unclear whether these terms are sufficiently small to make the bound informative in practice. In addition, the finite-horizon stability factor in Appendix A.3 may grow rapidly with the number of time steps. I would like to see a more concrete analysis of these quantities and a clearer explanation of what the result tells us beyond existing convergence analyses of biased SGD.

And I find the experimental evaluation somewhat limited. While the paper analyzes the two approximation errors and their relationships with model accuracy, it does not directly examine whether reducing these errors leads to better training outcomes. More systematic ablations would be helpful, particularly regarding the number of time steps, smoothing scale, and network architectures. In addition, the experiments are restricted to four relatively small vision benchmarks. Since the proposed framework is intended to characterize online SNN training more generally, I would like to see evaluations on other modalities, such as auditory datasets like SHD and SSC, as well as larger-scale benchmarks such as ImageNet. Without broader experiments, it is difficult to assess how well the theoretical findings and proposed diagnostics generalize across different tasks and network configurations.

Questions for authors
  1. Can the authors provide evidence that the convergence bound is informative for the networks used in the experiments? In particular, it would be helpful to examine the magnitude of the residual terms and discuss whether the restricted-PL and trajectory-stability assumptions are reasonable in these settings.

  2. Do the proposed diagnostics provide actionable information about training? For example, if the dominant approximation error is reduced, does this consistently improve optimization or test accuracy?

  3. How does the relative importance of hard–soft mismatch and temporal decoupling change with the simulation horizon? Since the analysis focuses on temporal approximation errors, experiments varying the number of time steps within the same dataset would be particularly informative.

4 comments
AuthorsOpenReview ↗

Author Response (1/2): Risk Bound and Temporal Correction

Thank you for the constructive questions about what the bound explains and how the diagnostics can guide training.

Critical Weakness 1 · Question 1 · What the bound explains

Theorem 1 connects two properties of the OTTT-A update to finite-step hard-spike risk. Equation (18) separates the bias due to changing the spike trajectory from the bias due to approximating temporal backpropagation. In Eq. (13), these give the residuals 3LOTTT2Λδ2Tδ,K/μU3L_{\mathrm{OTTT}}^2\Lambda_\delta^2\mathcal{T}_{\delta,K}/\mu_U and 3Dδ,K2/μU3D_{\delta,K}^2/\mu_U, respectively. This identifies which term is targeted by changing the forward spike model or the temporal gradient rule.

On CIFAR10-DVS, the soft temporal discrepancy Dδ2D_\delta^2 averaged 358.95 across five final checkpoints. The observed hard–soft trajectory amplification had seed medians of 4.48–4.58 and a maximum of 7.46. These measurements quantify the temporal discrepancy and trajectory amplification at the tested checkpoints. Evaluating the numerical risk bound over training requires the averages in Eq. (12), uniform constants over the iterate region, and the threshold tail at the restricted comparator for the risk-transfer term. The membrane recurrence discussed in Q3 gives a sufficient condition for horizon-uniform amplification at fixed depth and with uniformly bounded layer operators.

A concrete sufficient case for PL is an affine readout on frozen features with a positive MSE weight and a positive-definite feature second moment (Karimi et al., 2016, Appendix B). Using a global lower bound on the soft risk in the descent argument behind Eq. (19) gives an averaged stationarity bound with variance and bias residuals (Ajalloeian & Stich, 2020, Sec. 4). Restricted PL converts this stationarity bound into the risk-gap comparison in Theorem 1.

Critical Weakness 2 · Question 2 · Does reducing the temporal gap improve training?

A targeted temporal correction improves mean final accuracy on CIFAR-10, CIFAR-100, and CIFAR10-DVS. We trained VGG11 with OTTT-A and with the two-step correction (h=2h=2), using three paired seeds per dataset. Each pair shared the initialization and data seeds, hard-forward model, surrogate, input length, batch size, optimizer settings, and 300-epoch schedule. We report test accuracy from the final checkpoint at epoch 299.

DatasetOTTT-A accuracy (%)Two-step correction accuracy (%)Mean paired gain (pp)
CIFAR-1093.2893.48+0.20
CIFAR-10070.2270.61+0.39
CIFAR10-DVS75.2075.77+0.57

On CIFAR10-DVS, where the temporal discrepancy dominates at the default setting, the correction increased mean final accuracy by 0.57 percentage points. This illustrates how the diagnostic can guide a targeted intervention with a measurable training benefit.

For each loss, the correction replaces the OTTT contributions from the most recent hh lags with their SG-BPTT counterparts on the same hard trajectory. The original OTTT trace supplies the older contributions. The update is

Gh=GOTTT+GBPTT≤h−GOTTT≤h.G_h=G_{\rm OTTT}+G_{\rm BPTT}^{\le h}-G_{\rm OTTT}^{\le h}.

The subtraction removes the original window contribution before its replacement is added. The same-time contributions agree, so h=0h=0 recovers OTTT-A and h=T−1h=T-1 gives full SG-BPTT. In VGG11 training mode, synthetic-batch checks verified both endpoints and confirmed that the h=2h=2 replay preserved stepwise logits. We will use final-checkpoint accuracy in Fig. 3(a) to match the diagnostic checkpoint.

AuthorsOpenReview ↗

Author Response (2/2): Simulation Horizon and Reproducibility

Critical Weaknesses 1–2 · Question 3 · Within-dataset horizon and stability

The dominant discrepancy changes within CIFAR10-DVS with fixed event coverage and fixed weights. We rebinned each of the same 64 complete test recordings into 5, 10, or 20 frames in acquisition order using equal-event-count bins, with the remainder in the last bin. All settings use the same five final checkpoints trained at T=10T=10, the same membrane time constant (τ=2\tau=2), and the loss averaged over the observed steps. The rebuilt 10-frame inputs match the original inputs exactly, and the T=10T=10 diagnostics reproduce the earlier results for all five checkpoints. Within each spike mode, OTTT and BPTT produced identical stepwise logits at all three settings.

Frames, full event windowOTTT hard–soft gapHard temporal gapSoft temporal gapHard temporal / hard–softSoft temporal / hard–softSeeds with ratio above one, hard / soft
5704.22644.22495.330.9150.7031/5 / 0/5
10323.71402.91374.021.2451.1555/5 / 5/5
20497.06517.09515.961.0401.0383/5 / 3/5

We average per-example squared gradient differences within each seed and divide the across-seed means. The counts use each seed’s own ratio. From 5 to 10 frames, both temporal gaps decreased in absolute terms. The hard–soft gap fell by a larger fraction, so both ratios rose in all five seeds. The hard ratio crossed one in four seeds and the soft ratio in all five. At 20 frames, both aggregate ratios returned close to one. Meanwhile, from 10 to 20 frames, mean OTTT–BPTT cosine fell from 0.9194 to 0.9002 on the hard trajectory and from 0.9292 to 0.9080 on the soft trajectory. The ratios track the relative sizes of the two discrepancies, and cosine tracks the alignment of the update directions.

The rebinning intervention jointly changes events per frame and the number of membrane updates per recording. The fixed-input intervention tests memory directly. Increasing the membrane time constant from 2 to 4 at ten frames raised the hard and soft temporal gaps to 1306.13 and 1425.97.

In Appendix A.3, the revised sufficient-condition argument will carry the previous membrane difference through Eq. (1). For the default sigmoid scale, unit threshold, and leak of 0.5, the temporal coefficient after separating the pointwise spike mismatch is bounded by 0.5 in each layer. Geometric summation gives a sufficient forward stability factor uniform in the horizon with shared initialization, fixed depth, and uniformly bounded effective layer operators. The size of that factor and the remaining Eq. (13) constants depend on the configuration.

Critical Weakness 2 · Reproducibility and scope

The benchmark-level discrepancy ordering holds across ten consecutive seeds (2022–2031) per vision dataset, using the original settings and 128 diagnostic examples per run at epoch 299. The soft temporal gap exceeds the OTTT hard–soft gap in 0/10 CIFAR-10 runs, 0/10 CIFAR-100 runs, 9/10 CIFAR10-DVS runs, and 10/10 DVS Gesture runs. The corresponding ratios of mean gaps are 0.276, 0.263, 1.220, and 4.270.

In the exploratory accuracy analysis, we pool the original five runs with five outcome-selected runs per dataset. The pooled Pearson coefficients have the same signs and similar magnitudes as those of the original five runs.

We also evaluated a feedforward LIF MLP on the auditory SHD benchmark at T=100T=100 using the same ten seed IDs. On 128 test examples per final checkpoint, the mean OTTT hard–soft and soft temporal gaps were 42.58 and 72.84, respectively. The ratio of these means was 1.711, and the soft temporal gap was larger in all ten runs. The SHD experiment extends the diagnostic comparison to an auditory dataset and a different architecture.

References

  • Ahmad Ajalloeian and Sebastian U. Stich. 2020. On the Convergence of SGD with Biased Gradients. arXiv:2008.00051. Full text
  • Hamed Karimi, Julie Nutini, and Mark Schmidt. 2016. Linear Convergence of Gradient and Proximal-Gradient Methods Under the Polyak-Łojasiewicz Condition. ECML PKDD, 795–811. Full text

Thank you again for asking whether the diagnostics can guide training and how the discrepancy ordering changes with the simulation horizon. As the discussion period draws to a close, we would appreciate your feedback on the paired two-step correction experiment in Q2 and the fixed-event-coverage comparison in Q3. Is there a specific point in either comparison that would benefit from further clarification?

Dear Reviewer,

This is a kind reminder that the discussion period will end in less than 24 hours. We would appreciate it if you could take a moment to review authors' responses and provide any further comments before the discussion period closes.

Thank you for your time and effort!

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

This paper studies fixed-within-sequence OTTT updates in feedforward LIF networks. Taking full soft-BPTT as a reference, it decomposes the approximation error into a hard–soft mismatch and a temporal-decoupling error. Under explicit smoothness, variance, stability, and restricted-PL assumptions, it derives a finite-step hard-spike risk bound relative to a restricted minimizer of the smooth risk. Experiments across four benchmarks characterize these discrepancies and their associations with accuracy.

Critical strengths
  1. The analysis connects two distinct approximation mechanisms to a hard-risk bound through biased descent and risk transfer (Sections4–6, Theorem 1). This provides a useful framework for understanding OTTT_A, with its assumptions and restricted comparator clearly stated.
  2. The experiments show that discrepancy magnitude, training dynamics, and directional agreement capture different aspects of the approximation under the tested configurations (Figure 2, Table 1, Appendix B.1). The reported exceptions and sensitivity analyses support informative descriptive conclusions, while appropriately limiting the diagnostic–accuracy interpretation (Section 7.3, Appendix B.2).
Critical weaknesses
  1. The guarantee’s practical applicability remains limited: its assumptions and constants are not instantiated for a concrete network, and the experiments do not evaluate the optimization bound (Appendices A.3 and B). This limits its usefulness for quantitatively assessing training reliability.
  2. The empirical evidence primarily covers one architecture family, with five seeds per task and a single-seed smoothing sweep (Appendix B). Despite the sensitivity analyses, the stability of these diagnostic patterns across architectures and training configurations remains uncertain.
Questions for authors
  1. Appendix B specifies that the diagnostics are computed in evaluation mode. Could the authors clarify how the implementation preserves the presynaptic-trace weight gradients in both OTTT passes while restoring complete membrane and reset dependencies for soft-BPTT, with identical soft-forward trajectories?
  2. In addition, could the authors provide stronger empirical evidence for the key assumptions, such as the restricted-PL condition, stability, and variance bounds? For example, reporting proxy measurements of these quantities during training would help assess whether the theoretical conditions are approximately satisfied in the tested settings.
5 comments
AuthorsOpenReview ↗

Author Response (1/2): Evaluation Setup, Stability, and Variance

Thank you for the thoughtful review and for asking how the diagnostics and theoretical assumptions relate to the implementation.

Question 1 · Trace gradients in evaluation mode

Both OTTT passes explicitly set force_ottt_wrap=True in evaluation mode. The wrappers use current activations for forward values and detached presynaptic traces for the wrapped weight gradients. The first convolution uses the current input directly. The OTTT passes detach the previous membrane and subtractive reset for temporal differentiation. The BPTT passes disable the wrappers and restore both temporal dependencies.

All passes share the initial state, step-normalized loss, and inputs, with dropout disabled. Across 128 examples at each of five final checkpoints per dataset on CIFAR-10 and CIFAR10-DVS, the two soft passes produced identical logits at every time step, as did the two hard passes. We also recovered the archived hard–soft and soft temporal gaps. The wrapper preserves the forward activations, and the identical logits verify identical stepwise loss values under the shared loss. Together with the explicit graph settings, these checks establish that the temporal gap compares the two backward computations on the same forward trajectory. We will document the switches and logit check in Appendix B.

Critical Weakness 1 · Question 2 · Stability, variance, and restricted PL

We measured trajectory amplification and sampling variance at the final checkpoints. For the stability ratio in Assumption 4, we used 32 fixed examples per checkpoint. The numerator is the sum of the norms of the hard–soft spike and margin differences. The denominator is the norm of the difference between the two spike maps evaluated at the hard margins. For the variance proxy, we used 128 hard-OTTT sample gradients. The trace of their sample covariance divided by 128 estimates the expected squared fluctuation of an i.i.d. batch-128 mean in evaluation mode.

DatasetStability ratio, range of seed mediansLargest observed ratioBatch-128 variance proxySquared norm of estimated mean direction
CIFAR-105.47–5.658.160.830.86
CIFAR10-DVS4.48–4.587.4613.6212.61

The CIFAR10-DVS results give a concrete reason to retain both the trajectory-stability factor and the sampling-variance term. The largest observed amplification was 7.46, and the batch-128 variance proxy of 13.62 was comparable to the squared mean update of 12.61. The recurrence below explains how a pointwise spike-map mismatch propagates through membrane state and across layers. The variance term accounts for fluctuations in the sampled update. The identical-logit comparison in Q1 pairs the two backward computations on each fixed example, so the measured temporal gap isolates the backward-rule difference even when sampling variance across examples is appreciable. The last two table columns average the evaluation-mode estimates across five seeds.

Sufficient conditions for stability and variance. With subtractive reset, unit threshold, and the default sigmoid scale and leak, the membrane recurrence retains the previous state difference with a coefficient of norm at most 0.5 in each layer. New spike-map mismatch at each step and propagation across layers can amplify the total trajectory difference even as the carried membrane difference decays. The revised Appendix A.3 will sum this geometric kernel to bound the gain uniformly in the horizon at fixed depth, given shared initialization and uniformly bounded effective layer operators.

For variance, a batch mean of BB conditionally i.i.d. sample updates has conditional variance at most M2/BM^2/B if each update has norm at most MM. At fixed horizon, bounded inputs, parameters, and layer and loss derivatives give a finite such bound by unrolling the update. The deterministic surrogate adds no local-estimator variance. A training-wide estimate uses the same bound with uniform control of sample updates under the training augmentation, dropout, and sampling protocol.

AuthorsOpenReview ↗

Author Response (2/2): Restricted PL and Robustness Across Seeds and Architectures

When restricted PL holds. A concrete sufficient case is an affine readout trained on frozen soft-spike features. Let Σ\Sigma be the second moment of the bias-augmented features over inputs and time steps. With a positive MSE weight β\beta and CC classes, the readout risk has curvature at least (2β/C)λmin⁡(Σ)(2\beta/C)\lambda_{\min}(\Sigma). Positive definiteness of Σ\Sigma gives a PL constant for the readout problem (Karimi et al., 2016, Appendix B).

Liu et al. (2022, Sec. 4) show that a uniform positive lower bound on the tangent kernel gives a PL* inequality for empirical squared loss. This suggests a direction for end-to-end soft VGG11. Future work could analyze the tangent kernel of the time-unrolled model on a deterministic region containing the OTTT-A iterates and seek a corresponding restricted PL inequality for the population risk under our combined cross-entropy and MSE loss.

Guarantee without PL. In the telescoping argument for Eq. (19), using the global lower bound of the soft risk gives a bound on the averaged squared soft-gradient norm under the smoothness and variance assumptions, with residual terms for update bias and stochastic variance (Ajalloeian & Stich, 2020, Sec. 4). Assumption 6 is used to convert this stationarity bound to the restricted risk comparison in Theorem 1. The theoretical update is vanilla fixed-step OTTT-A. The experimental checkpoints were trained with momentum SGD or Adam.

Critical Weakness 2 · Discrepancy ordering across seeds and architectures

We extended the diagnostic to a feedforward LIF MLP on the auditory SHD benchmark at T=100T=100. For each dataset, we use ten consecutive seeds (2022–2031) and 128 diagnostic examples per final checkpoint at epoch 299.

DatasetModelMean soft temporal gap / mean OTTT hard–soft gapRuns with larger soft temporal gap
SHDFeedforward LIF MLP1.71110/10
CIFAR-10VGG110.2760/10
CIFAR-100VGG110.2630/10
CIFAR10-DVSVGG111.2209/10
DVS GestureVGG114.27010/10

The soft temporal gap dominates on SHD and both event-based vision benchmarks. The OTTT hard–soft gap dominates both static benchmarks. The smoothing sweep covers CIFAR-10 and CIFAR10-DVS with one seed per dataset.

For the exploratory accuracy correlations, we combine the original five runs with five outcome-selected runs per dataset. As in the original analysis, we correlate final test accuracy with PtailP_{\mathrm{tail}} on static datasets and log⁡D2\log D^2 on event datasets. In all four datasets, the combined ten-run Pearson coefficients have the same signs and similar magnitudes as those from the original five runs.

DatasetPearson rr, original fivePearson rr, exploratory fivePearson rr, combined ten
CIFAR-10−0.816−0.839−0.816
CIFAR-100+0.450+0.605+0.450
CIFAR10-DVS−0.967−0.880−0.916
DVS Gesture−0.939−0.969−0.938

References

  • Ahmad Ajalloeian and Sebastian U. Stich. 2020. On the Convergence of SGD with Biased Gradients. arXiv:2008.00051. Full text
  • Hamed Karimi, Julie Nutini, and Mark Schmidt. 2016. Linear Convergence of Gradient and Proximal-Gradient Methods Under the Polyak-Łojasiewicz Condition. ECML PKDD, 795–811. Full text
  • Chaoyue Liu, Libin Zhu, and Mikhail Belkin. 2022. Loss landscapes and optimization in over-parameterized non-linear systems and neural networks. Applied and Computational Harmonic Analysis, 59, 85–116. DOI

Thank you again for your questions about the diagnostic setup. In our response to Q1, we described how we preserve the OTTT trace gradients in evaluation mode and verified that the two soft passes produce identical logits at every time step. Does this address your concern about comparing the two backward rules on the same soft trajectory? We would be happy to clarify any remaining implementation detail.

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

The authors test out how OTTTs two primary approximations a) surrogate gradient on the spiking non-differentiability, and b) the temporal gradient-propagation approximation vs RTRL impact training. They study this gradient impact by looking at a BPTT, OTTT with a differentiable forward spike function, and the original OTTT and examining gradient divergence (error/magnitude) from a single step starting from a common checkpoint. The authors then derive expressions for the two errors by examining the threshold approximation in some detail and assuming the remaining error is due to the temporal approximation. The authors also test how much the forward-function's steepness impacts performance. Finally, the authors offer empirical evidence of learning performance on VGG over CIFAR(10/100)/DVS scale datasets, demonstrating: that the spiking-non-spiking gradient discrepancy is larger on their static-image configurations and the temporal discrepancy is larger on their event-based configurations and accuracy changes with surrogate steepness.

Critical strengths
  1. I think this is a good question to ask and a useful line of reasoning to follow, i.e., does the impact of differring from full temporal backpropagation, impact learning and how does it interact with the spike-approximations.

  2. The experimental comparisons are reasonably careful, the authors calculate alternative gradients using the same weights and inputs. I think a strength. The authors very correctly look at both gradient magnitude and direction.

  3. The theory explicitly connects the differentiable reference back to the SNN's loss.

Critical weaknesses
  1. In my opinion, there's a slightly different way to think about this than the author's formulation in Fig. 1. Let's consider a 2x2 grid: one axis is whether the forward pass uses hard spikes or a differentiable spike function and the other is whether temporal gradients are approximated using OTTT or propagated using full BPTT/RTRL. This gives four cases:
    • Hard spikes + OTTT: studied by the authors
    • Differentiable/soft spikes + OTTT: studied by the authors
    • Differentiable/soft spikes + full BPTT: studied by the authors
    • Hard spikes + full BPTT, using surrogate derivatives: missing
  • That missing piece, IMHO, is somewhat important to be able to comprehensively study the interaction between the temporal approximation-induced gradient-error and the spike-induced gradient-error. As currently written, I'm not sure that the authors conclusions hold for an actual SNN? I'd possibly go a bit further and say that the authors ought to use RTRL vs. BPTT -- but I think that'd be a tad nit picky.
  1. In my opinion, changing hard spikes to differentiable outputs changes what information flows through the network. This impacts actual information flow and not just gradient calculation. I know why this is done, I don't have a good solution here, but I think this is an important point to note?

  2. I don't quite see any detailed study of the temporal aspect of gradient flow here at all? The authors use VGG11 with LIF neurons and a fixed time constant of two steps. That means there is extremely limited temporal dependency/state in their system? This means that we don't actually see how the temporal aspect of their task is impacted? Additionally, DVS gestures is not a very temporal task, simply averaging/spike-counting out of the first layer can give really high accuracy, using something like SHD is somewhat critical here. CIFAR-10/100 is just static images. Ideally I would have wanted to see an experiment that deliberately increases or decreases the need for temporal information and checks whether the proposed diagnostics capture that change.

  3. The authors provide correlations between their gradient-mismatch vs. accuracy -- I think some kind of actual causal intervnetion from the same checkpoint might be the more convincing step? Specifically one issue IMHO is that curvature information, which is kind of useful here is not considerd, and having a larger discrepancy needs to be modulated with what the curvature might be in that new neighborhood. Simply looking at the gradient doesn't directly tell you if changing that is needed/sufficient?

  4. I'm a little fuzzy on the exact thing the authors are showing here, the general idea that approximate gradients can support learning under suitable conditions quite well known. The OTTT paper provides alignment results, and OSTL and OTPE discuss temporal effects re: online learning. BPTT is of course not online so online vs. offline elements also probably interact? The authors do point to the original OTTT result but I'm not sure what the contribution is here? What does this tell us about OTTT? The temporal parts are somewhat lacking in my opinion? It would be really helpful to know where the author's assumptions hold and what information we're getting from the bound? In fact the experiments use a different optimizer (as the authors mention).

  5. Eq. 40 bounds the difference in the current membrane margins using only the difference in the preceding spikes. However, a LIF neuron also has state? This means that history/memory beyond a few steps might be extremely impactful to the final output?

Questions for authors

Could you complete the 2×2 comparison by including hard-forward BPTT with surrogate derivatives?

What happens when you vary the temporal dependency studied in the task or of the model itself, i.e., do your conclusions change with the membrane time constant or a controlled temporal dependency?

9 comments
AuthorsOpenReview ↗

Author Response (1/2): Gradient Comparisons and Temporal Controls

Thank you for the careful review and for suggesting the hard-forward BPTT comparison.

Critical Weaknesses 1–2 · Question 1 · Hard-forward BPTT and the effect of soft spikes

The discrepancy ordering holds when we measure temporal approximation directly on the binary SNN. We completed the 2×2 comparison with hard-forward SG-BPTT, which uses the same sigmoid surrogate derivative as hard OTTT and the full backward graph through membrane and reset. At each of five final checkpoints per dataset, the four passes share the parameters, 128 examples, and sequence loss. The two backward rules give identical stepwise logits within each forward mode, and the original three passes reproduce the corresponding entries in Table 1.

DatasetHard–soft gap with OTTTHard–soft gap with BPTTOTTT–BPTT gap on hard spikesOTTT–BPTT gap on soft spikes
CIFAR-1075.8088.9645.1420.77
CIFAR10-DVS299.17305.52388.98358.95

Both temporal gaps are smaller than both hard–soft gaps on CIFAR-10 and larger on CIFAR10-DVS. The hard-forward comparison establishes the ordering for the binary model, and the soft-forward comparison measures the temporal term used in the analysis. The mean squared difference between their temporal-gap vectors is 25.33 on CIFAR-10 and 78.64 on CIFAR10-DVS. The two comparisons therefore agree on the ordering while quantifying how the magnitude depends on the forward dynamics. Each table entry is the mean per-example squared gradient difference, averaged across seeds.

Softening the spike changes the downstream activations, membrane state, and reset. Figure 1 will show both forward trajectories. Hard-forward SG-BPTT gives the surrogate-gradient reference on the binary trajectory, and soft BPTT differentiates the smooth loss used in the analysis. At fixed parameters, exact RTRL and BPTT apply the same chain rule when they use the same graph, loss, and local derivatives.

Critical Weakness 3 · Question 2 · Model memory and simulation step count

We measured temporal gradient discrepancies on 64 CIFAR10-DVS recordings at five checkpoints trained with T=10T=10, keeping the weights fixed.

  • Membrane time constant. The diagnostics respond to a controlled change in model memory. With the same ten input frames and weights, increasing τ\tau from 2 to 4 raises the per-step membrane retention factor from 0.5 to 0.75. The hard temporal gap rose from 402.91 to 1306.13 and the soft gap from 374.02 to 1425.97. The corresponding OTTT–BPTT cosines fell from 0.919 to 0.844 and from 0.929 to 0.857.
  • Simulation step count. At τ=2\tau=2, we rebinned each complete recording into 5, 10, or 20 frames using equal-event-count bins with the remainder in the last bin. The rebuilt 10-frame inputs matched the originals exactly and reproduced the earlier diagnostics for all five checkpoints. From T=5T=5 to T=10T=10, the temporal-to-hard–soft ratios rose from 0.915 to 1.245 on the hard trajectory and from 0.703 to 1.155 on the soft trajectory. Both ratios increased at every checkpoint, driven mainly by a larger decrease in the hard–soft gap. At T=20T=20 the aggregate ratios were 1.040 and 1.038. Rebinning holds event coverage fixed while changing events per frame and the number of membrane updates.

On the auditory SHD benchmark at T=100T=100, a feedforward LIF MLP also showed a larger soft temporal gap than OTTT hard–soft gap in all ten runs (seeds 2022–2031), with a ratio of mean gaps of 1.711.

Critical Weakness 4 · Same-checkpoint direction intervention

At each checkpoint, we estimated the OTTT and SG-BPTT directions on the same 64 examples, scaled them to equal length, and measured hard-forward loss changes on 64 disjoint examples.

DatasetStep norm as % of parameter normMean loss change, hard OTTTMean loss change, hard SG-BPTT
CIFAR-100.01%−0.00299+0.00138
CIFAR-100.1%−0.01197−0.00859
CIFAR10-DVS0.01%−0.01421−0.01637
CIFAR10-DVS0.1%−0.01706−0.01842

The equal-length intervention addresses the concern about local geometry by measuring the realized loss after each update, including changes in spike decisions. At both step lengths, the mean loss change favors OTTT on CIFAR-10 and SG-BPTT on CIFAR10-DVS.

AuthorsOpenReview ↗

Author Response (2/2): Theoretical Contribution and LIF Stability

Critical Weakness 5 · Contribution relative to OTTT, OSTL, and OTPE

Our contribution is to derive the two bias terms for OTTT-A and carry them through to a finite-step bound on hard-spike risk. This connects a threshold statistic and a same-trajectory gradient gap to distinct terms in the risk comparison. Xiao et al. (2022) relate the OTTT direction to a spike-representation objective. OSTL (Bohnstingl et al., 2023) factors gradients into layer-local eligibility traces and learning signals and omits a cross-layer residual in its deep approximation. Base OTPE (Summe et al., 2024) accumulates layer-local spike sensitivities to estimate downstream temporal effects.

The decomposition gives a conditional extension to deferred OSTL and base OTPE, instantiated on the same feedforward LIF model with the same per-step loss and surrogate. At fixed weights, let GMhardG_M^{\rm hard} and GMsoftG_M^{\rm soft} be the directions of rule MM on the two forward trajectories. Taking expectations over samples gives

=\mathbb{E}(G_M^{\rm hard}-G_M^{\rm soft}) +\mathbb{E}(G_M^{\rm soft}-G_{\rm BPTT}^{\rm soft}).$$ With a trajectory Lipschitz constant $L_M$, the squared norm of the first term is bounded by $L_M^2\Lambda_\delta^2\mathcal{T}_\delta(W)$, using Assumption 4 and the threshold-tail argument. Jensen’s inequality bounds the squared norm of the second by the expected squared soft-update/BPTT gap of $M$. Given a variance bound for that rule and the remaining assumptions of Theorem 1, the descent and risk-transfer argument applies. The constants and the iterates in Eq. (13) are rule-specific. This extension uses fixed-step SGD with weights fixed during each sequence. The momentum SGD and Adam checkpoints provide empirical gradient diagnostics at fixed weights. **Critical Weakness 6 · Membrane history in Eq. (40)** Subtracting the hard and soft LIF recurrences in Eq. (1) gives the membrane difference in layer $l+1$, including its dependence on the previous membrane state. $$\Delta u_{t+1}^{l+1} =\lambda\big(\Delta u_t^{l+1}-V_{\rm th}\Delta s_t^{l+1}\big) +W^l\Delta s_{t+1}^l.$$ On the hard margins, let $d_t$ be the pointwise difference between the hard and smoothed spike maps. The spike difference equals $d_t$ plus the smoothed map’s secant slope times the membrane difference, coordinatewise. For sigmoid smoothing at $\delta=1/4$ and unit threshold, that slope lies in $[0,1]$. At the default leak $\lambda=0.5$, the coefficient on the previous membrane difference has norm at most $0.5$. The geometric kernel then sums to at most 2. Summing in time and propagating the bound through the layers gives a horizon-uniform stability factor at fixed depth, with shared initialization and bounded effective layer operators. A common factor over the training region requires uniform operator bounds there. Theorem 1 uses the network-level stability condition in Assumption 4. Appendix A.3 will use this recurrence to state a sufficient condition for it. **Details of Ethics Concerns · AI use** The AI Use Statement on p. 10 documents the assistance used and the authors’ verification. **References** - Thomas Bohnstingl, Stanisław Woźniak, Angeliki Pantazi, and Evangelos Eleftheriou. 2023. *Online Spatio-Temporal Learning in Deep Neural Networks*. IEEE Transactions on Neural Networks and Learning Systems, 34(11), 8894–8908. [DOI](https://doi.org/10.1109/TNNLS.2022.3153985) - Thomas M. Summe, Clemens JS Schaefer, and Siddharth Joshi. 2024. *Estimating Post-Synaptic Effects for Online Training of Feed-Forward SNNs*. IEEE International Conference on Neuromorphic Systems (ICONS), 264–271. [DOI](https://doi.org/10.1109/ICONS62911.2024.00046) - Mingqing Xiao, Qingyan Meng, Zongpeng Zhang, Di He, and Zhouchen Lin. 2022. *Online Training Through Time for Spiking Neural Networks*. NeurIPS. [Proceedings](https://proceedings.neurips.cc/paper_files/paper/2022/hash/82846e19e6d42ebfd4ace4361def29ae-Abstract-Conference.html)

Thank you again for suggesting the hard-forward BPTT comparison. As the discussion period draws to a close, we would appreciate your thoughts on whether the completed 2×2 comparison in our response to Q1 addresses your concern about the discrepancy ordering on the binary SNN. If there is a specific aspect of the comparison you would like us to clarify, we would be happy to do so.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of manuscript revisions

Thank you for the constructive feedback. We have uploaded a revised manuscript that incorporates changes made in response to the reviews. Revised passages are marked in blue.

  • Theory and presentation. We revised Figure 1 to distinguish the hard and smooth forward trajectories, expanded the LIF trajectory-stability derivation, and clarified the role of the restricted PL condition (Appendices A.3–A.4).
  • Gradient diagnostics. We extended the discrepancy comparison to ten seeds per dataset and included a feedforward LIF MLP on the auditory SHD benchmark. We also added matched hard- and soft-forward comparisons on CIFAR-10 and CIFAR10-DVS, with stepwise logit checks for OTTT and BPTT within each forward mode (§7.2 and Appendix B.1).
  • Temporal controls and training. We added experiments that vary the membrane time constant and input binning at fixed weights, a comparison of one-step loss changes at matched update norms, and a short-window correction. In the three paired CIFAR10-DVS runs, the correction increased mean final test accuracy from 75.20% to 75.77% (§7.4 and Appendices B.2–B.3).
  • Final-checkpoint analyses. We updated the exploratory diagnostic–accuracy correlations (§7.3) and revised the CIFAR-10 smoothing-scale sweep to report test accuracy and diagnostics at the same final checkpoint (§7.5 and Figure 3).

Our point-by-point replies explain how these revisions address the individual review comments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.