acceptodds
Under review as a conference paper at ICLR 2027

Dynamics-Projected Eligibility Traces for Online Training of Spiking Neural Networks

Abstract

Spiking neural networks (SNNs) are a promising route to low-power event-driven computing, but their training typically relies on backpropagation through time (BPTT), whose activation memory scales linearly with the number of time steps. Online learning rules avoid this growth by keeping only the current-step computational graph and optionally maintaining fixed-size eligibility traces (ETs) online to approximate weight sensitivities. Existing compact presynaptic ETs use fixed decay or derive it from presynaptic dynamics, although weight sensitivities evolve according to the postsynaptic temporal Jacobian, which varies across layers and training stages. In this work, we propose Dynamics-Projected Eligibility Traces (DPET), which derives trace decay by projecting this Jacobian onto one least-squares coefficient per sample and layer. This scalar projection preserves a single presynaptic ET while adapting its decay to the postsynaptic dynamics. A state-space formulation extends DPET to neurons with richer internal states. In addition, to complement the missing feedback from future time steps in online learning, we introduce an auxiliary objective that uses later-step predictions from completed minibatches as supervision. Experiments across image, event-based, point-cloud, and language classification show gains over existing online methods with substantially lower training memory compared to BPTT. DPET provides a practical solution for memory-constrained SNN training.

Primary area:
applications to neuroscience & cognitive science
Keywords:
spiking neural networks, online learning, memory-efficient training

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

The authors propose DPET, a method for training recurrent SNNs with eligibility traces as opposed to BPTT. The method reduces the memory and computation required for storing eligibility traces through one additional approximation: replacing the dynamics Jacobian with the average of its trace (i.e., a scalar). They train their system on image benchmarks and measure its memory usage and compute time.

Critical strengths
  • Lowering the resources required to train an SNN relative to BPTT, or even e-prop, is a sensible goal.
  • The proposed modification to e-prop both lowers the memory and compute.
  • They have an extensive set of empirical evaluations on image benchmarks.
Critical weaknesses
  • If the primary goal is memory reduction for training an SNN on a GPU, then activation checkpointing should be the BPTT baseline for memory usage. This has the added benefit of not needing to redesign batchnorm and achieves the exact same accuracy as full BPTT.
  • There are no tasks reported that actually require temporal dynamics, such as SHD or SSC, where the method will actually be exercised relative to BPTT. All evaluations are on static images with 4-10 time steps, involving nearly no need for temporal credit assignment.
  • e-prop is already an approximation to BPTT, so adding in another approximation would seem to require some additional error analysis, which seems like it belongs in the main text.
Questions for authors
  • If lagged temporal distillation doesn't help with missing future pathways, how does it help?
  • The bold rows in table 7 are misleading. The fair comparison is torch vs. torch and should be portrayed that way.
  • Figure 2's legend is a little confusing.
  • Do you introduce the membrane potential regularization? Or is that an existing method with some modifications?
8 comments
AuthorsOpenReview ↗

Response to Reviewer TQpe

Thank you for the suggestions on checkpointing and for asking us to clarify the approximation and auxiliary methods. Our responses are as follows.

W1. If the primary goal is memory reduction for training an SNN on a GPU, then activation checkpointing should be the BPTT baseline for memory usage.

We agree that checkpointed BPTT is an appropriate memory baseline. Gradient checkpointing (GC) recomputes activations during the backward pass instead of storing them, so it keeps the exact BPTT update and, as you note, needs no change to batch normalization. DPET, by contrast, requires BN after the spiking neuron or a projection without BN (Section 4.4).

For the BPTT baseline, we ran spatial GC in SpikingJelly's multi-step mode, where each layer processes all time steps before going on to the next layer. DPET is an online rule and processes one time step at a time through the whole network. We therefore report BPTT in both modes. All runs use SpikingJelly 2.0.0 on one RTX 5090, Spiking VGG11 for CIFAR-100 at T=6T=6, batch size 128 and BF16. The BPTT+GC condition uses the level-4 automatic layout search with binary spike compression [1]. All conditions use the torch backend. Each condition runs with 10 warm-up and 100 timed iterations. We report the maximum peak allocation and the median time per batch.

MethodStep modePeak memory (GiB)Time per batch (ms)
BPTTsingle-step4.828760.1
BPTTmulti-step4.916851.5
BPTT + GC + spike compressionmulti-step1.771467.3
DPETsingle-step1.044964.6
DPET + GC + spike compressionsingle-step0.869871.1

At T=6T=6, BPTT in multi-step execution mode is faster than BPTT in single-step mode, since the computation on stateless layers can be parallelized over the time axis. GC lowers the memory cost of BPTT while keeping the exact update at the cost of increasing the time per batch from 51.5 to 67.3 ms. DPET reduces peak memory by about 41.0% relative to GC with an update time close to single-step BPTT, at the cost of approximating the temporal paths that GC preserves. We will add this comparison to the paper.

Furthermore, spatial GC and DPET can be used together. Rather than checkpointing the multi-step BPTT graph, we checkpoint only a subset of spatial activations within each DPET time step and recompute them for that step’s local backward pass before advancing to the next step. This leaves DPET’s online update unchanged. Compared with pure DPET, DPET+GC (with spike compression) reduces peak memory by 16.8% and increases time per batch by 10.1%.

AuthorsOpenReview ↗

Response to Reviewer TQpe (cont.)

W2. There are no tasks reported that actually require temporal dynamics, such as SHD or SSC, where the method will actually be exercised relative to BPTT. All evaluations are on static images with 4-10 time steps, involving nearly no need for temporal credit assignment.

The 4–10 step range applies only to our main static-image benchmarks; the original manuscript already included Sequential CIFAR-100 at T=32T=32 (presenting one image column per step; Section 5.3, Table 6), where DPET (57.28±0.3657.28\pm0.36%) and DPET+LTD (57.50±0.4957.50\pm0.49%) clearly outperform SLTT (54.62±0.4054.62\pm0.40%) and approach BPTT+TET (57.86±0.2757.86\pm0.27%).

Even so, we would like to clarify the role of temporal benchmarks in our study. Online learning rules must approximate historical sensitivities, omit future feedback paths, and cannot use cross-time-step operations (e.g., temporal convolution), which inherently limits raw sequential modeling capacity compared with full BPTT. Hence, online learning is not designed to replace BPTT on tasks with rich temporal dependencies. Instead, temporal tasks serve as stress tests for temporal credit assignment, because optimization on these tasks is highly sensitive to temporal gradients, and clear gains over other online rules directly validate the fidelity of DPET's eligibility trace.

Following your suggestion, we also evaluated DPET on SHD under two complementary settings (also reported in our response to Reviewer VEDm, W1). All results are test accuracy, mean ± SD over three seeds, selected by best validation accuracy, without LTD or membrane regularization.

1. Feedforward LIF on SHD (T=250T=250): isolating within-layer temporal credit assignment. To directly test whether the eligibility trace preserves historical sensitivity across long horizons, we first use a two-layer feedforward LIF network (128 units per layer, fixed decay 0.95, T=250T=250, 120 epochs), where no recurrent weights exist and temporal credit must propagate entirely through membrane dynamics.

RuleBPTTSLTTOTTTNDOTDPET
SHD test acc. (%)81.38 ± 1.3061.41 ± 0.1176.30 ± 1.0870.77 ± 0.5579.78 ± 1.05

Without historical eligibility (ϕ=0\phi=0), SLTT drops to 61.4161.41%. DPET achieves the highest accuracy among all online rules (79.7879.78%, outperforming OTTT and NDOT) and closes most of the gap to BPTT (81.3881.38%), demonstrating that projecting postsynaptic dynamics onto the eligibility trace effectively preserves long-horizon temporal credit.

2. Recurrent LIF on SHD (T=100T=100): evaluating performance in a higher-capacity recurrent model. We next evaluate a two-layer recurrent LIF architecture (1024 units per layer, learned decays, T=100T=100, 100 epochs, per-step cross-entropy), extending DPET with eligibility traces for recurrent weights.

RuleSHD test acc. (%)
BPTT87.13 ± 0.79
SLTT84.41 ± 0.90
OTTT85.96 ± 1.07
NDOT85.03 ± 1.32
pp-prop (α=0.93\alpha=0.93)86.81 ± 0.71
Weighted DPET85.48 ± 1.21
DPET86.75 ± 0.85

With recurrent synaptic connections, all rules achieve stronger accuracy. DPET reaches 86.7586.75%, matching the performance of pp-prop . Together, the two SHD experiments show that DPET preserves long-range membrane-level credit.

AuthorsOpenReview ↗

Response to Reviewer TQpe (cont. 2)

W3. e-prop is already an approximation to BPTT, so adding in another approximation would seem to require some additional error analysis, which seems like it belongs in the main text.

Thank you for this suggestion. Your description is accurate: DPET adds one approximation to e-prop. Like e-prop, it omits feedback through future deeper states; in addition, it replaces the per-neuron eligibility traces with a single shared trace.

Consider one sample in a feedforward dense LIF layer with current input x[t]\mathbf x[t]. For each output neuron jj, e-prop keeps an exact eligibility Ej[t]\mathbf E_j[t] (the sensitivity of that neuron's membrane potential to its incoming weights, accumulated over past steps). It evolves with the neuron's own one-step Jacobian Φj[t]\Phi_j[t], as Ej[t]=Φj[t]Ej[t−1]+x[t]\mathbf E_j[t]=\Phi_j[t]\mathbf E_j[t-1]+\mathbf x[t]. DPET keeps one input-sized trace E^[t]\widehat{\mathbf E}[t] for the whole layer, with Φj[t]\Phi_j[t] replaced by the layer mean ϕ[t]\phi[t], as E^[t]=ϕ[t]E^[t−1]+x[t]\widehat{\mathbf E}[t]=\phi[t]\widehat{\mathbf E}[t-1]+\mathbf x[t]. Their gap dj[t]=Ej[t]−E^[t]\mathbf d_j[t]=\mathbf E_j[t]-\widehat{\mathbf E}[t] follows

dj[t]=Φj[t]dj[t−1]+(Φj[t]−ϕ[t])E^[t−1].\mathbf d_j[t]=\Phi_j[t]\mathbf d_j[t-1]+(\Phi_j[t]-\phi[t])\widehat{\mathbf E}[t-1].

The first term carries the earlier gap forward; the second is the new gap introduced at step tt. Since E^[t−1]\widehat{\mathbf E}[t-1] is the same vector for every jj, the sum of the squared norms of the newly introduced gaps is ∥E^[t−1]∥2∑j(Φj[t]−ϕ[t])2\|\widehat{\mathbf E}[t-1]\|^2\sum_j(\Phi_j[t]-\phi[t])^2, which the mean ϕ[t]\phi[t] minimizes. The newly introduced gap is zero when all neurons share the same Jacobian at that step.

With the same current spatial error signal ζnow,j[t]\zeta_{\mathrm{now},j}[t], the gradient rows computed from DPET's trace and from exact eligibility differ by ζnow,j[t]dj[t]⊤\zeta_{\mathrm{now},j}[t]\mathbf d_j[t]^\top. Minimizing the new gap at each step does not guarantee the smallest accumulated difference (Appendix A.6), so we also measure it. Table 5 uses e-prop's exact eligibility as its exact local reference. Against it, DPET's error is 0.154, compared with 0.287–0.353 for SLTT, OTTT and NDOT; only pp-prop, which stores an additional postsynaptic trace, is lower. Against full BPTT+SG, where the error also includes the effect of omitted future feedback, DPET has the lowest error of these compact rules (0.342).

In the revision, we will move this analysis into Section 4.1, next to the projection objective.

Q1. If lagged temporal distillation doesn't help with missing future pathways, how does it help?

Thank you for the question. Without future feedback, the update at step tt is driven by the current loss L[t]\mathcal L[t]; there is no explicit target from later-step predictions. LTD supplies this on the target side: the step-tt output is trained to match the network's later-step predictions for the same class, collected from previously completed minibatches. It therefore needs no future Jacobians of the current sample, while gradients flow only through the current-step predictions.

Adding LTD to DPET raises mean accuracy from 76.97±0.0976.97\pm0.09% to 77.34±0.4077.34\pm0.40% on CIFAR-100 and from 84.13±0.3584.13\pm0.35% to 85.33±0.4785.33\pm0.47% on CIFAR10-DVS (Table 4). We will revise Section 4.2 to state this relation directly.

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

This work proposes an eligibility trace-based training method for spiking neural networks. This method can achieve higher memory efficiency compared with widely used BPTT-based approaches. The authors conducted comprehensive evaluations and achieved state-of-the-art accuracy.

Critical strengths

This work focuses on a meaningful problem: reducing the memory footprint of SNN training. Though BPTT-based SNN training approaches achieve performance that is comparable to DNNs, the memory consumption increases linearly. This work achieves state-of-the-art accuracy with a reduced memory footprint.

Critical weaknesses

Since one motivation is that temporal truncation discards historical information, this work seeks to faithfully track local eligibility and preserve temporal credit. But this work mainly uses static image datasets or static-image-derived DVS datasets. One criticism regarding neuromorphic datasets is that for datasets derived from static images, the temporal dimension of the data carries little information [1]; this is also why Spatial Learning Through Time (Meng et al., 2023) works.

It would be more convincing if the authors evaluated on datasets where the spike timing truly matters, such as DVS Gesture mentioned in [1] and Spiking Heidelberg Digits in [2]. They exhibit time-varying complex patterns in the temporal domain, and therefore can more rigorously reflect the strength of this algorithm.

The methodology in Section 5.3 and Table 5 is problematic. The methodology of evaluating gradient fidelity via relative L2L_2 error is flawed. In SNNs, there is no true gradient in BPTT due to the use of surrogate gradients. The L2L_2 distance simply measures how closely one approximation mimics another, which does not guarantee better navigation of the actual discrete loss surface. As deep SNNs are highly non-linear and non-convex, distinct estimators can have entirely different training trajectories toward different low-loss basins. Furthermore, the exact choice of surrogate gradient function is flexible, and a different surrogate gradient can lead to a completely different trajectory. Measuring an algorithm's gradient estimation on another estimator's trajectory provides no rigorous evidence of how effectively that algorithm would actually assign temporal credit.

Evaluating on a dataset with rich and complex temporal dynamics and achieving high accuracy would sufficiently justify the temporal credit fidelity.

[1] Is neuromorphic mnist neuromorphic? analyzing the discriminative power of neuromorphic datasets in the time domain [2] The heidelberg spiking data sets for the systematic evaluation of spiking neural networks

Questions for authors

Recently, several works have applied SNNs to regression tasks such as time-series forecasting, where no class labels exist. Because LTD relies entirely on a class-indexed teacher bank, how can the proposed method be extended to tasks beyond classification? (Note: I am not asking the authors to conduct additional experiments for this open-ended question; this is purely for discussion and will not negatively impact my score.)

4 comments
AuthorsOpenReview ↗

Response to Reviewer VEDm

Thank you for the questions about temporal learning, gradient comparisons and regression tasks. Our responses are as follows.

W1. It would be more convincing if the authors evaluated on datasets where the spike timing truly matters, such as DVS Gesture and Spiking Heidelberg Digits.

We first clarify the role of temporal benchmarks in our study. Online learning rules must approximate historical sensitivities, omit future feedback paths, and cannot use cross-time-step operations (e.g., temporal convolution), which inherently limits raw sequential modeling capacity compared with full BPTT. Hence, online learning is not designed to replace BPTT on tasks with rich temporal dependencies. However, temporal tasks can serve as stress tests for temporal credit assignment, because optimization on these tasks is highly sensitive to temporal gradients. In the original manuscript, Sequential CIFAR-100 (T=32, presenting one image column per step; Section 5.3, Table 6) served this diagnostic role. DPET reaches 57.28±0.3657.28\pm0.36% (and 57.50±0.4957.50\pm0.49% with LTD), clearly outperforming other online rules and approaching BPTT+TET (57.86±0.2757.86\pm0.27%).

Following your suggestion, we evaluated DPET on DVS Gesture. We follow SLTT's experimental setup, using Spiking VGG11 with T=20. We use post-neuron batch normalization for compatibility with DPET. DPET achieves 97.87±0.3597.87\pm0.35% test accuracy, and DPET with both auxiliary objectives achieves 98.01±0.2098.01\pm0.20%. For reference, the published accuracies of SLTT [1] and OTTT$_A$ [2] are 97.9297.92% and 96.8896.88%, respectively.

We further evaluated DPET on SHD under two complementary settings. All results are test accuracy, mean ± SD over three seeds, selected by best validation accuracy, without LTD or membrane regularization.

1. Feedforward LIF on SHD (T=250T=250): isolating within-layer temporal credit assignment.

To directly test whether the eligibility trace preserves historical sensitivity across long horizons, we first use a two-layer feedforward LIF network (128 units per layer, fixed decay 0.95, T=250, 120 epochs), where no recurrent weights exist and temporal credit must propagate entirely through membrane dynamics.

RuleBPTTSLTTOTTTNDOTDPET
SHD test acc. (%)81.38 ± 1.3061.41 ± 0.1176.30 ± 1.0870.77 ± 0.5579.78 ± 1.05

Without historical eligibility (ϕ=0\phi=0), SLTT drops to 61.4161.41%. DPET achieves the highest accuracy among all online rules (79.7879.78%) and narrows the gap to BPTT (81.3881.38%).

2. Recurrent LIF on SHD (T=100T=100): evaluating performance in a higher-capacity recurrent model.

We next evaluate a two-layer recurrent LIF architecture (1024 units per layer, learned decays, T=100T=100, 100 epochs, per-step cross-entropy), extending DPET with eligibility traces for recurrent weights.

RuleSHD test acc. (%)
BPTT87.13 ± 0.79
SLTT84.41 ± 0.90
OTTT85.96 ± 1.07
NDOT85.03 ± 1.32
pp-prop (α=0.93\alpha=0.93)86.81 ± 0.71
Weighted DPET85.48 ± 1.21
DPET86.75 ± 0.85

In the recurrent setting, all tested rules achieve higher accuracy. DPET reaches 86.7586.75%, close to pp-prop's 86.8186.81%. Together, the two SHD experiments show that DPET preserves long-range membrane-level credit.

AuthorsOpenReview ↗

Response to Reviewer VEDm (cont.)

W2. The methodology in Section 5.3 and Table 5 is problematic. The methodology of evaluating gradient fidelity via relative error is flawed.

Thank you for raising this point. We agree that BPTT with surrogate gradients (BPTT+SG) is not the true gradient of the discrete spiking loss, and that matching it more closely does not guarantee better optimization. We will state this explicitly in Section 5.3.

Nevertheless, BPTT+SG is a principled reference rather than an arbitrary heuristic. Gygax & Zenke [3] show that, for a stochastic spiking neuron with escape noise, the surrogate gradient equals the derivative of the neuron's expected output when the surrogate derivative matches the derivative of the escape-noise function. Through stochastic automatic differentiation, they extend this correspondence to multi-layer SNNs, with deterministic networks as a special case. We will cite this result in the same statement in Section 5.3.

The purpose of Table 5 is more specific: DPET and the other online rules in Table 5 are all formulated as online approximations to BPTT+SG under the same surrogate function (Section 3.2). Table 5 therefore measures how closely each online rule approximates this shared target update.

We also agree that training trajectories matter. In Table 5, we fixed the weights and forward states across estimators to isolate the gradient approximation error, using SLTT-trained checkpoints. To check whether this choice affects the conclusion, we also evaluated DPET on checkpoints trained by DPET+LTD itself. The relative errors (0.1560.156 against the exact local gradient and 0.3400.340 against full BPTT+SG) remain very close to those on SLTT-trained checkpoints (0.1540.154 and 0.3420.342). While we did not test other surrogate functions, changing the surrogate would change the BPTT+SG reference for all online rules alike.

Finally, we agree that evaluating on datasets with rich temporal dynamics is a more direct test of temporal credit assignment, which is why we added the SHD experiments in our response to W1.

Q1. Because LTD relies entirely on a class-indexed teacher bank, how can the proposed method be extended to tasks beyond classification?

LTD is an auxiliary objective designed specifically for classification. For regression tasks, we simply omit LTD and keep the rest of DPET unchanged.

To verify DPET on regression, we evaluate it on the Solar and PEMS-BAY time-series forecasting tasks following [4]. Because the convolutional spike encoder in [4] maps each continuous input window into a T=4T=4 spike sequence across all time steps at once, we first pretrain a shared encoder offline and freeze its weights, and then train the downstream spiking predictor online from identical initializations using SLTT, OTTT, or DPET with mean per-step MSE loss. We report test R2R^2 (higher is better) / RRSE (lower is better) averaged on 3 random seeds.

Task (input / horizon)SLTTOTTTDPET
Solar (168 / 24)0.9406 / 0.25170.9400 / 0.25280.9397 / 0.2530
PEMS-BAY (12 / 96)0.8071 / 0.47620.8134 / 0.46480.8131 / 0.4651

Without LTD, DPET trains normally on both regression benchmarks, achieving performance comparable to SLTT and OTTT on Solar and outperforming SLTT on PEMS-BAY.

[1] Meng et al. Towards Memory- and Time-Efficient Backpropagation for Training Spiking Neural Networks. In ICCV, 2023.

[2] Xiao et al. Online Training Through Time for Spiking Neural Networks. In NeurIPS, 2022.

[3] Gygax and Zenke. Elucidating the Theoretical Underpinnings of Surrogate Gradient Learning in Spiking Neural Networks. arXiv:2404.14964, 2024.

[4] Lv et al. Efficient and Effective Time-Series Forecasting with Spiking Neural Networks. In ICML, 2024.

Thanks for replying and the efforts for additional evaluations. I don't have additional concerns.

Thank you for your response. We are glad to hear that our clarifications were helpful. If you feel that your concerns have been sufficiently addressed, we would greatly appreciate it if you could take this into account when finalizing your assessment. Should any further questions arise during the discussion phase, please do not hesitate to let us know.

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

This paper proposes Dynamics-Projected Eligibility Traces (DPET) for online training of spiking neural networks. The method uses a least-squares projection of the postsynaptic temporal Jacobian to obtain an adaptive decay coefficient for eligibility traces. It also introduces lagged temporal distillation (LTD) and membrane potential regularization to improve training. Experiments across different tasks show competitive performance with reduced memory usage compared with BPTT.

Critical strengths

The paper proposes a simple and efficient method to adapt eligibility-trace decay based on postsynaptic neuronal dynamics. The method maintains low memory usage while achieving competitive performance on several benchmarks. The paper also provides detailed gradient analysis and computational efficiency comparisons.

Critical weaknesses

The proposed Jacobian projection treats all neurons equally without considering their different contributions to the parameter gradients. Neurons with small or negligible gradient contributions may still affect the shared decay coefficient. Therefore, minimizing the Jacobian approximation error does not necessarily lead to an accurate eligibility-trace or gradient approximation. The effectiveness of this design requires further validation.

Questions for authors
  1. The proposed Jacobian projection treats all neurons equally, regardless of their contributions to the gradients. Can the authors explain this choice and compare it with alternative strategies that consider different neuronal contributions?
  2. In Table 5, pp-prop achieves a lower error against the exact local gradient than DPET. Can the authors further explain this result and analyze whether the uniform projection contributes to the difference?
3 comments
AuthorsOpenReview ↗

Response to Reviewer syRz

Thank you for the questions about the projection criterion and neuronal contributions. Our responses are as follows.

W1. Minimizing the Jacobian approximation error does not necessarily lead to an accurate eligibility-trace or gradient approximation

We agree that minimizing Jacobian error does not ensure an accurate eligibility-trace or gradient estimate. Using the paper's notation (Section 3.2),

=\sum_{t=1}^{T}\boldsymbol\zeta^l[t]\mathbf E^l[t].$$ The gradient therefore depends on the learning signals $\boldsymbol\zeta^l[t]$ and on eligibility $\mathbf E^l[t]$, which accumulates temporal Jacobian products. Minimizing a one-step Jacobian approximation error cannot guarantee minimum overall gradient error. DPET's motivation is to retain neuronal-dynamics information in a compact, input-shaped trace, making local eligibility practical for deep SNNs. The scalar fit adapts the trace to postsynaptic dynamics within this storage constraint. DPET does not guarantee an optimal eligibility-trace or gradient estimate, and this fact was already explicitly stated in the original submission (Section 4.1 and Appendix A.6, “Layer-Wise Reduction Error”). We will also move this error analysis into Section 4.1 and discuss this limitation in the Limitations section of the revised manuscript. > Q1. The proposed Jacobian projection treats all neurons equally, regardless of their contributions to the gradients. Can the authors explain this choice and compare it with alternative strategies that consider different neuronal contributions? We are not sure whether the reviewer's concern is about sharing one scalar across neurons when updating eligibility, or about choosing that scalar through an unweighted mean without accounting for the current spatial error signal. Here, we discuss these aspects respectively. **(1) Sharing one scalar across neurons** DPET exploits the linearity of $g_{\mathbf W}^{\prime\,l}$ in its input to accumulate historical inputs in one trace (Section 4.1). For a temporal factor to be absorbed into the same input for all output neurons, it must be a common scalar. We therefore approximate $\mathbf\Phi^l[t]$ by $\phi^l[t]\mathbf I$ for each sample and time step. Appendix A.3 proves this requirement for the shared-input construction. **(2) Using an unweighted mean to choose the scalar** The decay chosen at step $t$ enters the trace that all later steps reuse, and each later step weights that trace with its own spatial error signal, which is not yet available at step $t$. The current signal $\boldsymbol\zeta_{\mathrm{now}}^l[t]$ describes only the present step, so we do not use it to choose the decay. We fit the neuronal dynamics $\mathbf\Phi^l[t]$ directly, because they are what the trace carries forward. The unweighted least-squares fit gives $\phi^l[t]=\operatorname{tr}(\mathbf\Phi^l[t])/n_l$. Each neuron's current spatial error signal $\zeta_{\mathrm{now},j}^l[t]$ then weights its gradient contribution, as in Section 4.2. To test whether using the current spatial error signal in the projection itself improves this design, we also evaluated the weighted fit $$\phi_w^l[t] =\frac{\sum_j(\zeta_{\mathrm{now},j}^l[t])^2\Phi_j^l[t]} {\sum_j(\zeta_{\mathrm{now},j}^l[t])^2}.$$ Here, $\Phi_j^l[t]=[\mathbf\Phi^l[t]]_{jj}$, and $\phi_w^l[t]$ denotes the least-squares fit weighted by squared current spatial error signals. Following the protocol in Section 5.3 (Spiking VGG11 on CIFAR-100, $T=6$), we compare the two projections across three checkpoints trained with different random seeds and report layer-averaged relative $L_2$ gradient error. | Method | Error against exact local gradient | Error against full BPTT+SG gradient | | ------------- | ---------------------------------: | ----------------------------------: | | Uniform DPET | 0.1540 ± 0.0013 | 0.3419 ± 0.0023 | | Weighted DPET | 0.1341 ± 0.0006 | 0.4072 ± 0.0024 | Uniform projection gives lower error against full BPTT+SG, although the weighted variant has lower local error. We also tested whether the projection choice affects learning from SHD's auditory spike streams. Both variants use the same recurrent model ($T=100$), per-step cross-entropy and 100 epochs. | Projection | SHD test acc. (%) | | ------------- | ----------------: | | Uniform DPET | 86.75 ± 0.85 | | Weighted DPET | 85.48 ± 1.21 | Uniform DPET has higher SHD accuracy, which supports our choice of dynamics-only fit for the eligibility trace.
AuthorsOpenReview ↗

Response to Reviewer syRz (cont.)

Q2. In Table 5, pp-prop achieves a lower error against the exact local gradient than DPET. Can the authors further explain this result and analyze whether the uniform projection contributes to the difference?

pp-prop maintains a presynaptic trace vector and a postsynaptic trace vector, whose outer product forms a rank-one eligibility approximation. Its postsynaptic trace preserves neuron-specific sensitivities for each output unit, whereas DPET updates a single shared input trace with a uniform scalar decay. Preserving this per-neuron sensitivity allows pp-prop to more closely match the exact local gradient (Section 5.3; Appendix C.5), whereas collapsing across neurons via uniform projection increases local error.

To test whether uniform averaging contributes to this local-error gap, we compare DPET with the weighted variant described in our response to Q1. The two variants use the same eligibility trace representation and differ only in how the decay coefficient is chosen. Using the same CIFAR-100 (T=6T=6, SLTT-trained VGG11) and recurrent SHD (T=100T=100) settings as in Q1, we compare pp-prop against both DPET variants across three seeds:

MethodError against exact local gradientError against full BPTT+SG gradientSHD test acc. (%)
pp-prop0.1281 ± 0.00090.4247 ± 0.002886.81 ± 0.71
Uniform DPET0.1540 ± 0.00130.3419 ± 0.002386.75 ± 0.85
Weighted DPET0.1341 ± 0.00060.4072 ± 0.002485.48 ± 1.21

Weighted DPET reduces the exact-local gradient error from 0.1540 to 0.1341, narrowing the gap to pp-prop. This suggests that uniform averaging contributes to the local-error gap. However, the ranking reverses when comparing against full BPTT+SG gradients. Uniform DPET achieves the lowest error against full BPTT+SG and similar SHD accuracy to pp-prop, without requiring its additional postsynaptic trace.

Dear Reviewer syRz,

As the discussion period is nearing its close, we wanted to politely follow up in case you have had a chance to read our response and the revised manuscript.

In our reply above, we responded to your questions on the uniform Jacobian projection—clarifying its approximation scope in Section 4.1, adding a gradient-weighted ablation (Weighted DPET) where Uniform DPET gives lower full-BPTT gradient error and higher SHD accuracy, and analyzing the exact-local versus full-BPTT comparison with pp-prop.

Please let us know if these clarifications resolve your concerns or if any questions remain before the deadline. Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Reponse to All the Reviewers

We thank the reviewers for their constructive suggestions. We highlight below the relevant content in the original submission and the additional evidence provided in our responses.

1. Temporal evaluation and its purpose. The online setting considered here restricts cross-time-step operation and gradient propagation in ways that full BPTT does not (e.g., temporal convolution cannot be used in online settings), which inherently limits sequential modeling capacity. Matching or surpassing the absolute performance of full BPTT on temporal benchmarks is therefore not our primary evaluation criterion. We instead use temporal benchmarks as diagnostics of temporal credit assignment, comparing online rules under matched constraints, with BPTT serving as a reference. The original submission includes Sequential CIFAR-100 at T=32T=32, with one image column presented per step (Section 5.3, Table 6). Following the reviewers’ suggestions, we have extended this evaluation to DVS Gesture and to SHD with feedforward and recurrent LIF networks.

2. Projection objective and approximation analysis. Appendix A.6 of the original submission analyzes the layer-wise reduction error and distinguishes one-step Jacobian fitting from accumulated eligibility and gradient errors. We will bring this analysis into Section 4.1. The new weighted-projection comparison further examines the choice of uniform averaging.

3. Reference used in the gradient diagnostics. Section 3.2 defines the reference as BPTT with surrogate gradients. Our comparisons measure approximation error under a fixed surrogate function and matched weights and forward states, complementing the task-level evaluation. We will clarify this interpretation in Section 5.3.

4. Role and attribution of the auxiliary objectives. Section 4.2 presents LTD as auxiliary supervision and states that it does not reconstruct the missing future-feedback paths; Table 4 evaluates its contribution separately. Section 4.3 acknowledges prior membrane-potential regularization. We will clarify our specific contribution in the Introduction.

5. Efficiency comparison with checkpointed BPTT. On CIFAR-100 VGG11, DPET reduces peak memory by 41.0% relative to spatial gradient checkpointing (GC) at comparable update speed and can also be combined with GC. In follow-up experiments on larger backbones under fixed GPU memory budgets, DPET supports larger batch sizes without recomputation and improves end-to-end training throughput over checkpointed BPTT by 25.2% on ScanObjectNN (Spike Discrepancy Transformer; ~1 hour 40 minutes saved per 300-epoch run, 85.91% vs. 86.03% test acc.) and 22.0% on QNLI (SpikingLM; ~28 minutes saved per 10-epoch run, 83.72% vs. 83.62% val. acc.).

We also provide regression experiments without LTD using frozen encoders and online training of the downstream predictors. We will incorporate these additional results and the clarifications above into the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.