acceptodds
Under review as a conference paper at ICLR 2027

NLT-prop: Nonlocal Temporal Credit Assignment for Online Learning in Spiking Neural Networks

Abstract

Spiking neural networks (SNNs), inspired by biological neural networks, communicate through discrete spikes and offer the potential for energy-efficient information processing. They are commonly trained using backpropagation through time (BPTT) with surrogate gradients. However, BPTT requires propagating gradients backward in time, making it biologically implausible and incompatible with strictly online learning. Real-time recurrent learning (RTRL) enables online gradient computation, but maintaining full eligibility traces in recurrent networks requires memory that grows cubically with network width and incurs substantial computational costs. Many existing online methods reduce these costs by simplifying temporal dependencies between neurons, limiting their ability to capture nonlocal temporal influences. We propose NLT-prop, an online learning method that efficiently approximates nonlocal temporal credit assignment through recurrent connections. Starting from the forward recursion for eligibility traces, NLT-prop decomposes the temporal Jacobian into diagonal and off-diagonal components, separating local neuronal dynamics from interactions between neurons. It organizes temporal paths by their number of nonlocal transitions and uses structured low-rank approximations to propagate the retained contributions forward in time. For a fixed retained order, the memory required for eligibility traces scales linearly with network width. This formulation explicitly incorporates nonlocal temporal influences without storing the full activity history or introducing additional eligibility-trace decay parameters. Experiments on SHD, N-MNIST, CIFAR10-DVS, DVS Gesture, and DMTS demonstrate that NLT-prop achieves performance comparable to BPTT while supporting fully online learning.

Primary area:
optimization
Keywords:
spiking neural networks, online learning, temporal credit assignment, eligibility traces, low-rank approximation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nEX8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NLT-prop, an online learning method for spiking neural networks. The method divides the temporal Jacobian into local and nonlocal components and uses low-rank approximations to preserve nonlocal temporal information with linear eligibility-trace memory. Experiments on five datasets show competitive performance compared with BPTT and existing online learning methods.

Critical strengths

The paper proposes a structured low-rank approximation to preserve nonlocal temporal information with linear eligibility-trace memory. This provides a useful way to improve temporal credit assignment in online SNN learning while reducing memory requirements. The experimental results also show competitive performance.

Critical weaknesses

The proposed approximations and computational efficiency are not sufficiently validated. The method uses rank-one approximations and shares factors across different nonlocal orders, but the effects of these designs are not fully studied. In addition, the paper lacks direct comparisons of training time and computational cost with existing methods. These issues make it difficult to evaluate the effectiveness and practical benefits of the proposed method.

Questions for authors
  1. Why is rank-one approximation used? Can the authors compare different approximation ranks to show their effects on accuracy and memory usage?
  2. What is the effect of sharing the presynaptic factor across different nonlocal orders? Can the authors compare shared and independent factors?
  3. Can the authors provide training time and computational cost comparisons with existing online learning methods to demonstrate the practical efficiency of NLT-prop?
1 comment

Thank you for your constructive feedback. Your questions have helped us improve the rigor of both the work and its presentation. We have uploaded a revised PDF with the changes highlighted in blue. We address your questions below.

Comment/Question:

Why is rank-one approximation used? Can the authors compare different approximation ranks to show their effects on accuracy and memory usage?

Answer: We chose rank one to balance approximation error and computational efficiency. We have added an experiment comparing the accuracy and computational efficiency of NLT-prop with nonlocal trace ranks of 1, 2, and 3, together with a derivation of the updates for higher ranks (Appendix B.8, Table 9).

Higher ranks reduce accuracy on SHD and improve it by less than one percentage point on DVS Gesture. These limited accuracy benefits come with a substantial increase in training cost. On DVS Gesture, ranks 2 and 3 increase training time to 32.85 and 43.13 times that of rank one, respectively, and increase peak GPU memory usage by 16.7% and 41.0%. On SHD, training time increases to 4.55 and 4.37 times that of rank one. These results favor rank one as a practical choice.

Rank-one factorizations have also been used for local traces in existing online learning methods. We extend this approach to nonlocal contributions and use a different procedure to obtain the presynaptic and postsynaptic traces.

Comment/Question:

What is the effect of sharing the presynaptic factor across different nonlocal orders? Can the authors compare shared and independent factors?

Answer: Using the current presynaptic factor ptp^t for both local and nonlocal contributions, in place of pt−1p^{t-1} for the nonlocal contributions, is an approximation. Its purpose is to reduce the additional computational cost of retaining nonlocal terms while preserving accuracy.

The intuition comes from the common presynaptic origin of different propagation paths. Consider a source contribution BB introduced at time t−3t-3. In the absence of further source terms, all local and nonlocal paths originating from BB retain the same presynaptic activity and can therefore share a presynaptic factor (Eq. 32). If subsequent inputs have the same presynaptic direction, that shared direction can still represent the presynaptic dependence exactly. When subsequent inputs have different directions, factor sharing introduces an approximation error whose bound is given in Appendix A.6.

The added explanation in Appendix A.6 shows how this approximation reduces computation and storage. We also added an experiment comparing the shared factor with retaining a separate factor for the nonlocal contributions (Appendix B.9, Table 10). Sharing reduces training time, SynOps, and the spike rate, and improves accuracy by 2.64 percentage points. Two possible explanations for the accuracy gain are interactions with other approximation errors and a regularization effect from constraining each sample's gradient contribution to a common presynaptic direction.

Comment/Question:

Can the authors provide training time and computational cost comparisons with existing online learning methods to demonstrate the practical efficiency of NLT-prop?

Answer: Yes. We have added computational efficiency measurements for the different methods in Appendix B.7, Table 7. NLT-prop has training times close to those of the two main competing methods, OTTT and pp-prop, on SHD and DMTS, and is faster than both on DVS Gesture. It also offers advantages in peak GPU memory usage, spike rate, and SynOps.

The reductions in spike rate and SynOps are particularly useful for systems that exploit sparse activity. Their potential energy benefits are only partly reflected in our measurements on the A800. Neuromorphic hardware and implementations that exploit sparsity could make fuller use of these reductions to save energy and memory. Together with the accuracy improvements and the stronger learning and generalization of long-term dependencies shown in Section 4.7 and Figure 2, these benefits make the modest additional training time worthwhile.

Reviewer_XKX3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary
  • This manuscript presents an online learning rule for training deep spiking neural networks. Specifically, this work proposes to design a linear-memory eligibility trace that can capture non-local temporal dynamics. The idea is indeed interesting as it faces one of the main challenges of training SNNs which is the massive memory requirements of using BPTT and does this while preserving richer temporal dynamics. The non-local components seem to improve accuracy, but results do not follow a clear trend
Critical strengths
  • The idea of using low-rank approximations to estimate the non-local temporal components in the eligibility trace with linear memory looks very interesting. This is a very relevant work for the neuromorphic community and can spark some ideas for memory efficient learning rules with richer temporal dynamics for training SNN.
Critical weaknesses
  • The presentation is difficult to follow. While the high-level ideas are clear, how those are implemented is difficult to understand. While the ideas are interesting, is difficult to assess if the core mathematical derivations are correct.
  • The results do not look consistent; there do not seem to be a clear trend showing that non-local temporal dynamics are beneficial.
Questions for authors
  • Could you provide a more details on how the off-diagonal decomposition works in NLT-prop?
3 comments

Thank you for your valuable feedback. Below, we explain the mathematical basis of the off-diagonal decomposition in NLT-prop in more detail, following Section 3.4 and Appendix A.6.

Comment/Question:

Could you provide a more details on how the off-diagonal decomposition works in NLT-prop?

Answer: We explain the approximation of the off-diagonal Jacobian here, followed by the construction and compression of the nonlocal trace factors. The steps follow Appendix A.6.

Approximating the nonlocal Jacobian.

For the recurrent model, the off-diagonal Jacobian is Ct=offdiag⁡(W)Diag⁡(ψt−1)C^t=\operatorname{offdiag}(W)\operatorname{Diag}(\psi^{t-1}), where WW is the recurrent weight matrix and ψt−1\psi^{t-1} contains the surrogate derivatives. Its entries describe how the previous state of one neuron affects the current state of another.

We estimate a dominant singular component of this operator. The reason for this choice is that the leading singular component gives the best rank-one approximation in Frobenius norm. It provides a compact approximation of the interaction matrix using two vectors. A fixed number of alternating power iterations estimates this component.

The implementation uses the exact products with CtC^t and its transpose from Eqs. (67)–(69). For example,

Ctv=W(ψt−1⊙v)−diag⁡(W)⊙ψt−1⊙v.C^t v=W(\psi^{t-1}\odot v)-\operatorname{diag}(W)\odot\psi^{t-1}\odot v.

The vector is first multiplied elementwise by the surrogate derivatives, then by the recurrent weight matrix, and the diagonal contribution is subtracted. The transpose product is evaluated using the corresponding expression in Eq. (67).

We use the alternating power iteration and balanced-factor construction described in Appendix A.5. Following the initialization specified in Appendix A.6, we start from the normalized all-ones vector at the beginning of a sequence and reuse the preceding right direction at subsequent time steps. Each iteration alternates products with CtC^t and its transpose, normalizing after each multiplication. The final directions are scaled by the square root of the estimated singular value to obtain the balanced factors ata^t and btb^t.

Preserving the zero diagonal.

The outer product can introduce diagonal entries. We therefore apply the diagonal correction in Eqs. (70)–(72):

C~t=at(bt)⊤−Diag⁡(at⊙bt).\widetilde C^t=a^t(b^t)^\top-\operatorname{Diag}(a^t\odot b^t).

This correction preserves the structural separation between local and nonlocal propagation: self-transitions belong to the diagonal operator, while the nonlocal operator connects distinct neurons. It also removes approximation error in entries whose exact values are zero.

The corrected operator is represented by two vectors and an implicit diagonal subtraction. Its action is

C~tv=at((bt)⊤v)−at⊙bt⊙v.\widetilde C^t v=a^t((b^t)^\top v)-a^t\odot b^t\odot v.

The implementation evaluates one inner product and elementwise operations. Thus, applying the stored approximation costs O(N), where N is the number of neurons. The rank-one component supplies the compact representation, and diagonal masking enforces the known structure of the nonlocal Jacobian.

Answer (continued):

Constructing the incoming nonlocal contribution.

We choose a rank-one approximation H^νt=rνt(qνt)⊤\widehat H_\nu^t=r_\nu^t(q_\nu^t)^\top for each retained nonlocal order so that its neuronal contribution can be stored using two vectors. The initial condition follows from the model's fixed initial state. Since u0=0u^0=0 is independent of the weights, the initial sensitivity is E0=0E^0=0. Accordingly, the local and nonlocal eligibility factors are initialized to zero, including rν0=qν0=0r_\nu^0=q_\nu^0=0, as specified in Appendix A.6. The all-ones initialization described in Part 1 applies to the power-iteration direction used to estimate CtC^t.

At the first time step, the direct term B1B^1 establishes the local trace. This local trace then supplies the source for the first nonlocal update at the next step. At each update, Eq. (80) constructs a candidate from the saved factors at time t−1t-1 and the incoming source. Equation (81) projects that candidate back to rank one, producing the factors stored for the next step. Appendix A.6 treats the first and higher orders separately because their source factors have different structures.

For the first nonlocal order, the preceding local neuronal factor is Diag⁡(ℓt−1)\operatorname{Diag}(\ell^{t-1}). Applying the approximated nonlocal operator gives the source

=a^t w^\top-\operatorname{Diag}(a^t\odot w), \qquad w=b^t\odot\ell^{t-1}.$$ Multiplication by the local diagonal scales each column of the nonlocal operator. This transfers the local contributions to other neurons while retaining their source-neuron indices. The resulting source contains an outer product and a diagonal correction. To represent it with one pair of vectors, we apply power iteration to this source using the products in Eq. (73). The implementation evaluates those products directly from $a^t$ and $w$, using inner products and elementwise multiplication. The iteration starts from the normalized vector $w$, as specified in Appendix A.6, and returns the source factors $x_2,y_2$. For higher orders, the preceding neuronal factor is already an outer product. Therefore, $$\widetilde C^t\widehat H_{\nu-1}^{t-1} =(\widetilde C^t r_{\nu-1}^{t-1})(q_{\nu-1}^{t-1})^\top.$$ This equality explains the implementation directly: applying the operator changes the left factor and preserves the right factor. The source is obtained by setting $x_2=\widetilde C^t r_{\nu-1}^{t-1}$ and $y_2=q_{\nu-1}^{t-1}$. This is the construction in Eqs. (77)–(78). **Combining and compressing the two contributions.** The locally propagated contribution supplies the other pair, $x_1=d^t\odot r_\nu^{t-1}$ and $y_1=q_\nu^{t-1}$, where $d^t$ is the local Jacobian diagonal. The candidate update is consequently a sum of two outer products: $$H_{\nu,\mathrm{cand}}^t=x_1y_1^\top+x_2y_2^\top, \qquad \widehat H_\nu^t=\mathcal P_1[H_{\nu,\mathrm{cand}}^t].$$ Recompression keeps the stored representation at two vectors per order as contributions accumulate over time. The projection $\mathcal P_1$ selects the best rank-one approximation in Frobenius norm. The implementation exploits the fact that the candidate's columns lie in the span of $x_1$ and $x_2$. It orthogonalizes these two vectors, constructs the 2-by-2 representation of the candidate times its transpose, and solves the resulting eigenproblem analytically. The leading direction and its corresponding right factor give the compressed pair. Appendix A.5 derives this calculation, which is used in Eqs. (79)–(81) of Appendix A.6. All orders read factors from time $t-1$, so each update follows the stated temporal recurrence. The new factors are clipped after compression and replace the previous factors after the order updates. Finally, the current presynaptic factor $p^t$ is shared across the local and nonlocal contributions. Appendix A.6 identifies this as an additional approximation. It allows the postsynaptic contributions to be combined before forming the weight gradient, reducing the gradient outer-product computation and presynaptic-buffer storage. Its contribution to the approximation residual is included explicitly in Eqs. (82)–(89).

We provide further explanation and experimental evidence for two approximation choices in NLT-prop: the rank-one representation of nonlocal traces and the shared presynaptic factor.

Rank-one approximation.

We choose rank one to balance approximation quality and computational efficiency. Each nonlocal update produces a candidate that is a sum of two outer products. The compression described in Appendix A.5 gives its best rank-one approximation in Frobenius norm through a 2-by-2 eigenproblem, keeping each retained nonlocal order represented by two vectors.

We have added an experiment comparing nonlocal trace ranks of 1, 2, and 3, together with a derivation of the updates for higher ranks (Appendix B.8, Table 9). Higher ranks reduce mean accuracy on SHD and improve it by less than one percentage point on DVS Gesture. On SHD, ranks 2 and 3 require 4.55 and 4.37 times the training time of rank one, respectively. On DVS Gesture, they require 32.85 and 43.13 times the training time and increase peak GPU memory usage by 16.7% and 41.0%. These results support rank one as a practical choice for retaining nonlocal contributions at an affordable training cost.

Shared presynaptic factor.

We use the current presynaptic factor ptp^t for both local and nonlocal contributions, replacing pt−1p^{t-1} in the nonlocal terms. This allows the local and nonlocal postsynaptic contributions to be combined before forming the weight gradient. As derived in Appendix A.6, this reduces two gradient outer products to one and halves the presynaptic-buffer storage. The same appendix explicitly separates and bounds the approximation residual introduced by factor sharing.

We have also added a comparison between sharing ptp^t and retaining a separate pt−1p^{t-1} for the nonlocal contributions (Appendix B.9, Table 10). On SHD, sharing increases mean accuracy from 74.07% to 76.71%, a gain of 2.64 percentage points. Mean training time decreases from 580.26 to 573.51 ms per batch, SynOps decrease by 21.0%, and the mean spike rate falls from 15.86% to 11.37%.

The accuracy gain may reflect interactions with the other approximation errors. Sharing may also have a regularization effect by constraining each sample's gradient contribution to a common presynaptic direction. The measured results support this approximation as a way to reduce computation and synaptic activity while maintaining effective learning.

Comment/Question:

The results do not look consistent; there do not seem to be a clear trend showing that non-local temporal dynamics are beneficial.

Answer: The nonmonotonic accuracy trend as QQ increases in Table 2 is consistent with the analysis in Appendix A.9. Retaining additional orders recovers more nonlocal paths while introducing further approximation error. The benefit of each added order depends on how its contribution interacts with the remaining gradient error, including the effects of the other approximations. This balance provides an explanation for the nonmonotonic trend. Moreover, we can guarantee that introducing the first nonlocal order (Q=1Q=1) is beneficial.

To further isolate the contribution of nonlocal terms, we have added comparisons between Q=0Q=0 and Q=1Q=1 for computational efficiency (Appendix B.7, Table 8) and DMTS accuracy (Section 4.7, Table 4). The efficiency comparison shows that retaining the first nonlocal order increases training time by 24.8%, while reducing SynOps by 81.6% and the spike rate by 96.9%. These reductions in synaptic activity could translate into further energy and memory savings on neuromorphic hardware and in implementations that exploit sparse tensors.

The delay-dependent evaluation in Section 4.7, Figure 2, also shows strong performance on DMTS at long delays. On newly generated data with delays beyond the training range, NLT-prop achieves the highest mean accuracy among the compared methods across all evaluated delay groups. The DMTS ablation further shows that adding the first nonlocal order increases mean accuracy by 3.11 percentage points. This provides direct experimental support for the contribution of nonlocal temporal credit to delayed-memory learning.

Together, these experiments support the benefits of retaining nonlocal terms through improvements in learning performance and reductions in firing activity and SynOps.

Reviewer_1zxdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes NLT-prop, an online learning algorithm for spiking neural networks (SNNs) that aims to preserve nonlocal temporal credit assignment under a linear eligibility-trace memory budget.

The key idea is to decompose the temporal state Jacobian into diagonal and off-diagonal components and organize temporal sensitivities according to the number of cross-neuron transitions. The method combines separable local eligibility traces with deterministic low-rank approximations of nonlocal contributions, enabling online gradient updates without storing the full temporal computation graph.

The authors derive the corresponding update rules and approximation error bounds, and extend the approach to recurrent SNNs and feedback-connected convolutional networks. Experiments on five datasets compare NLT-prop against BPTT and four online learning baselines. The results demonstrate competitive predictive performance, with the highest mean accuracy among the evaluated methods on three datasets, alongside linear eligibility-trace storage at a fixed approximation order.

Critical strengths

1. A meaningful structural perspective on online temporal credit assignment.

Rather than compressing the global sensitivity tensor without distinguishing its temporal structure, NLT-prop organizes gradient contributions by the number of nonlocal transitions. This is more than a superficial reformulation of existing low-rank methods: it introduces an interpretable approximation dimension that explicitly distinguishes local dynamics from cross-neuron temporal interactions. The resulting recursive algorithm offers a potentially useful direction for designing memory-constrained online learning rules.

2. A technically developed method with meaningful theoretical and empirical support.

The paper provides detailed derivations, an approximation-error decomposition, and experimental validation across five datasets and two backbone families. In particular, the linear eligibility-storage claim is supported by both complexity analysis and memory measurements. The gradient-alignment experiments in Appendix B.6 also provide encouraging evidence that the compact representation can retain information relevant to BPTT gradients. These results support the practical feasibility of the proposed approach.

Critical weaknesses

1. The evidence does not sufficiently isolate the benefit of nonlocal temporal credit assignment.

The central claim is that retaining nonlocal temporal contributions improves online learning under a constrained memory budget. However, the direct evidence remains limited. In Table 2, introducing the first nonlocal order improves mean accuracy by only 0.86 percentage points on SHD and 1.06 points on DVS Gesture, with non-negligible variation across seeds. More importantly, DMTS is the task where NLT-prop shows some of its largest improvements over compact online baselines, yet no Q=0 comparison is reported for this task.

Consequently, the current results establish that the complete algorithm is competitive, but do not convincingly establish how much of its advantage is attributable specifically to nonlocal temporal propagation rather than other aspects of its eligibility-trace approximation. This is consequential because nonlocal propagation is the paper's central proposed contribution.

2. The practical value of the additional approximation machinery remains insufficiently characterized.

The linear eligibility-storage complexity is convincingly supported. However, NLT-prop introduces Jacobian-vector products, power iterations, and repeated low-rank compression, all of which incur additional computation. The paper does not provide a sufficiently direct comparison of training time or throughput against other linear-memory online methods.

Moreover, several reported accuracy advantages over the strongest competing online method are small relative to the reported run-to-run variability. Thus, while the memory and accuracy results are promising, the evidence does not yet establish whether the additional nonlocal approximation complexity provides a reliable and practically worthwhile advantage.

Questions for authors

Q1. Contribution of nonlocal propagation on the delayed-memory task.

Could the authors provide a Q=0 versus Q=1 comparison on DMTS under identical training settings and random seeds? Since DMTS exhibits relatively large improvements over OTTT and pp-prop, this comparison would help determine whether those improvements are specifically attributable to the retained nonlocal temporal contributions. A comparison across different delay lengths would also be informative if such measurements are available.

Q2. Gradient approximation and the role of Q.

Appendix B.6 provides useful gradient-alignment measurements against BPTT. Could the authors clarify how the alignment changes between Q=0 and Q=1 under identical weights and inputs? This would provide more direct evidence that the nonlocal terms recover otherwise missing temporal gradient information. In addition, can the authors clarify whether the nonmonotonic accuracy trends across Q correspond to changes in gradient approximation quality?

Q3. Computational overhead and practical efficiency.

Could the authors report or clarify the per-step training time or throughput of NLT-prop relative to OTTT and pp-prop under the same hardware and network settings? Since the proposed method introduces iterative factorization and additional Jacobian-vector products, this comparison is important for assessing the practical trade-off between improved temporal credit assignment and computational cost.

2 comments

We thank the reviewer for the thoughtful and constructive feedback. We have followed your suggestions to improve the paper and added experiments and figures that address your questions. The changes are highlighted in blue in the revised PDF. We respond to each point below.

Comment/Question:

The evidence does not sufficiently isolate the benefit of nonlocal temporal credit assignment.

Q1. Contribution of nonlocal propagation on the delayed-memory task.

Answer: Table 2 shows modest accuracy gains from adding nonlocal terms. These gains and the subsequent nonmonotonic changes are consistent with the analysis in Appendix A.9. Retaining additional orders recovers more nonlocal paths while introducing approximation error. The resulting improvement depends on the interplay between the recovered contributions and the approximation errors, including the relative directions of the different error components.

Following your suggestion, we added experiments in Section 4.7, Figure 2, and Table 4. On DMTS, NLT-prop with Q=1Q=1 improves accuracy by 3.11 percentage points over Q=0Q=0, a larger gain than those observed in the SHD and DVS Gesture ablations. We also grouped test trials by delay length and found that NLT-prop's accuracy increases steadily with the delay. The accuracy of pp-prop also increases but remains below that of NLT-prop. OTTT, which achieves the fastest training in our additional efficiency tests, has substantially lower accuracy than the other methods on the original DMTS test set.

We then generated a new DMTS test set with delays beyond the training range. NLT-prop achieves the highest mean accuracy among the compared methods across all unseen delay groups, with stronger performance at the longest delays. These results demonstrate its ability to generalize to substantially longer memory intervals. Since DMTS requires retaining sample information across delays containing distractors, the findings support the value of nonlocal terms for learning persistent memory representations that tolerate interference and generalize across time scales.

We also added a comparison of computational efficiency between Q=0Q=0 and Q=1Q=1 in Appendix B.7, Table 8. Including nonlocal terms increases training time by 24.8%, with little change in inference speed or peak GPU memory usage, while reducing SynOps by 81.6% and the spike rate by 96.9%. The current hardware and implementation capture only part of the potential benefit of this reduced activity. Neuromorphic hardware and implementations using sparse tensors and specialized algorithms could make greater use of this sparsity to reduce energy consumption and memory usage.

Comment/Question:

The practical value of the additional approximation machinery remains insufficiently characterized.

Q3. Computational overhead and practical efficiency.

Answer: Thank you for raising this point. We have added computational efficiency measurements in Appendix B.7, Table 7. NLT-prop has training times close to those of OTTT and pp-prop on SHD and DMTS, and is faster than both on DVS Gesture. It also offers advantages in measured peak GPU memory usage during training, firing rate, and SynOps. The firing-rate measurements in Appendix B.5, Figure 5, further show the relatively sparse activity of NLT-prop in deep networks.

Another benefit is its ability to learn long-term dependencies, as discussed in our response to Q1. Taken together, the gains in accuracy and long-term dependency learning, along with reductions in firing rate and SynOps, make the additional training cost worthwhile.

Comment/Question:

Q2. Gradient approximation and the role of Q.

Answer: We have added Figure 7 in Appendix B.6 to show how gradient alignment between NLT-prop and BPTT varies with QQ. The cosine similarity also changes nonmonotonically. Its trends overlap partly with those of accuracy, with differences at individual orders.

Several factors can explain these differences. The cosine similarity measures the gradient direction in the final hidden layer for particular samples and time steps. Accuracy also depends on gradient magnitudes, parameter updates in earlier layers, and the optimization trajectory throughout training. In addition, models trained with different orders have different learned weights, so these measurements correspond to different locations in the loss landscape. Together, these factors can affect the relationship between local gradient alignment and final predictive performance.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Summary of changes

We thank all the reviewers for their helpful feedback. We have uploaded a revised version of the paper that contains the following changes (highlighted in blue in the PDF):

  • Adding an experiment on DMTS that compares NLT-prop with Q=0Q=0 and Q=1Q=1, reports the accuracy of NLT-prop and the baselines across different delay groups, and evaluates their generalization to a new dataset with longer delays than those encountered during training. See Section 4.7, Figure 2, and Table 4. These results demonstrate the benefit of retaining nonlocal temporal credit for learning long-term dependencies and generalizing to longer delays.

  • Adding an experiment comparing the computational efficiency of NLT-prop and the baselines on SHD, DVS Gesture, and DMTS. See Appendix B.7 and Table 7. This experiment measures training speed, GPU energy consumption, peak memory usage, and SynOps. The results show that compression and approximation keep the computational cost of retaining nonlocal terms manageable. NLT-prop achieves training speeds comparable to OTTT and pp-prop, while offering advantages in peak memory usage and SynOps.

  • Adding an experiment comparing rank-one and higher-rank approximations of the nonlocal trace factors, together with a derivation of the updates for higher ranks. See Appendix B.8 and Table 8. Increasing the rank yields limited accuracy benefits at substantially greater training cost, supporting rank one as a practical choice.

  • Adding an experiment comparing the use of the shared presynaptic factor ptp^t with retaining a separate factor pt−1p^{t-1} for the nonlocal contributions. See Appendix B.9 and Table 9. On SHD, sharing ptp^t improves mean accuracy, slightly reduces training time, and lowers both SynOps and the mean firing rate. These results support the use of the shared factor for efficient gradient computation.

  • Adding Figure 7 in Appendix B.6, showing how the choice of QQ affects the cosine similarity between the gradients computed by NLT-prop and BPTT.

  • Correcting the BPTT test result on DVS Gesture in Table 1, where an incorrect value was reported in the original submission. We have also updated the corresponding descriptions in Section 4.4 and Section 5. The corrected results show that NLT-prop achieves the highest mean accuracy on four datasets and the second-highest on the remaining dataset among the compared methods.

  • Revising Appendix A.6 to explain the rationale for replacing the separate presynaptic factor pt−1p^{t-1} with the shared factor ptp^t in the nonlocal contributions, and to quantify the resulting savings in computation and storage. We have also removed a redundant equation, previously numbered Eq. (72).

  • Improving the wording and correcting cross-references to figures and tables.

We have uploaded an updated PDF with the following additional changes since the first revision:

  • Adding an experiment comparing the computational efficiency of NLT-prop with Q=0Q=0 and Q=1Q=1. See Table 8 in Appendix B.7. Tables 8 and 9 in the first revision are now Tables 9 and 10, respectively. This experiment shows that retaining nonlocal terms adds modest computational overhead while substantially reducing the spike rate and SynOps.

Dear Reviewers,

As the discussion period is approaching its deadline, I would like to ask whether my responses have adequately addressed your concerns. If there are any remaining questions or points that require further clarification, I would be very happy to provide additional explanations.

Thank you very much for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.