acceptodds
Under review as a conference paper at ICLR 2027

Delta SNN: Learning Time-Variant Synaptic Weights in Spiking Neural Networks for Event Vision

Abstract

Spiking neural networks (SNNs) naturally align with event cameras through sparse, event-driven computation, yet effectively modeling the spatiotemporal dynamics of event streams remains challenging. Typically, SNNs fully share synaptic weights across timesteps to maintain parameter and computational efficiency. However, event distributions and neuronal states evolve over time, causing different timesteps to exhibit heterogeneous features and potentially conflicting optimization demands, while a fixed transformation must serve them all. To address this issue, we propose Delta SNN, a time-variant synaptic framework that relaxes complete temporal weight sharing. Delta SNN augments a shared synaptic backbone with compact temporal modulation, enabling adaptive transformations while retaining cross-time knowledge sharing. We further introduce binary synaptic routing to realize temporal modulation with sparse, addition-based inference. Theoretically, we show that time-variant weights substantially enhance the expressive capacity of SNNs. Experiments across three challenging event-vision tasks consistently improve strong baselines with only marginal overhead.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Network

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_xEmkOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes Delta SNN, a time-variant synaptic framework designed to improve the spatiotemporal modeling capability of SNN. It introduces compact temporal weight modulation and binary synaptic routing to enhance temporal adaptability while maintaining computational efficiency. Extensive experiments on event-based dataset demonstrate consistent improvements over strong baselines.

Critical strengths

1.The paper identifies an important limitation of temporal weight sharing in SNNs, supported by feature and gradient similarity analyses. 2.The proposed compact temporal modulation and binary synaptic routing enhance temporal modeling while preserving computational efficiency. 3.Extensive experiments and ablation studies, comparisons with alternative temporal adaptation methods, and efficiency analyses provide convincing empirical support.

Critical weaknesses
  1. The proposed method uses time-dependent projection matrices UtU_t, whereas the proof in Appendix A assumes a shared matrix UU. How does the theoretical bound on the effective dimension (1+q)(1+q) remain valid under the actual formulation?
  2. The proof of Theorem A.2 directly applies Lemma A.1, which concerns fixed-weight LIF neurons, to the time-varying setting. Could the authors justify why the same bound on the number of hyperplanes still holds?
  3. Equation (9) compares upper bounds rather than achievable expressivity. A larger upper bound does not necessarily imply greater expressive capacity. Could the authors provide a corresponding lower-bound analysis or revise this claim?
  4. The main text defines M(T)=(T2+T)/2M(T)=(T^2+T)/2, but the example for T=4T=4 uses M(4)=11M(4)=11 instead of 1010. The authors should reconcile these definitions and verify the subsequent calculations.
  5. The proposed method is interesting and empirically effective, and its practical contributions are sufficient to support acceptance. However, the theoretical analysis requires more rigorous justification. If the current theoretical claims cannot be rigorously established, I would encourage the authors to revise or remove the theoretical analysis rather than retain potentially unsupported conclusions.
Questions for authors

See weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer xEmk — Part 1 of 4

We thank the reviewer for recognizing the motivation, method design, and experimental results of this work, and for the specific suggestions on the theoretical analysis. Our analysis aims to explain the structural constraints imposed by temporal weight sharing and explore how compact temporal modulation alleviates these constraints. In response to your comments, we have revised Appendix A to further clarify the scope of each conclusion in the revision, add the counting derivation for time-dependent inputs and weights, and rigorously distinguish the theoretical insights offered by upper-bound analysis from constructive proofs of realizable expressivity. We respond point by point below.

 

W1. Time-dependent projections UtU_t versus a shared projection UU, and the scope of the effective-dimension bound

Thank you for pointing out this distinction between the theoretical assumptions and the actual parameterization. The original 1+q1+q dimension bound assumes a shared projection UU, whereas the actual method allows a different UtU_t at each timestep. The revised Theorem A.2 formulates the result uniformly using the actual rank of the weight sequence, with the shared-projection result as a special case.

For neuron kk, define

\left[w_{k,1},\ldots,w_{k,T}\right], \qquad r_k=\operatorname{rank}(B_k).$$ The input currents across all timesteps depend on $x$ only through $\lbrace w_{k,t}^{\top}x\rbrace_{t=1}^{T}$. Thus, after choosing a basis for the span of these weight vectors, the neuron's complete spike sequence can be expressed as a function of the corresponding $r_k$-dimensional projection, with $$r_k\le \min\lbrace d,T\rbrace.$$ For the actual parameterization $$w_{k,t}=w_k+U_t^\top\widehat m_{k,t},$$ the projection directions can differ across timesteps, giving the general bound $$r_k\le \min\lbrace d,T,1+Tq\rbrace.$$ When $U_t\equiv U$, the time-dependent $M_t$ can still produce time-varying weights, so **the shared-projection setting retains the core mechanism of temporal modulation**. In this case, all time-dependent weights lie in $\operatorname{span}\lbrace w_k,\operatorname{row}(U)\rbrace$, yielding $$r_k\le \min\lbrace d,T,1+q\rbrace.$$ Theorem A.2 now uses $r_k$ for the general case of time-dependent projections and explicitly restricts $1+q$ to the dimension bound under a shared projection. The revision will apply the same actual-rank formulation and shared-projection restriction to Theorem 4.1 in the main text. The independent counting derivation in W2 allows time-dependent weights and therefore combines directly with this actual-rank formulation.
AuthorsOpenReview ↗

Response to Reviewer xEmk — Part 2 of 4

W2. Extending the fixed-weight lemma to time-varying weights and counting hyperplanes

Thank you for this question. Our previous version omitted the treatment of x(t)=x+ytx(t)=x+y_t, and we provide the complete derivation here. Specifically, we restrict the hyperplane count and subsequent combinatorial upper bound to the non-leaky case β=1\beta=1.

Handling time-dependent input offsets. For the input model in the main text,

x(t)=x+yt,x(t)=x+y_t,

each yty_t is fixed and independent of the variable xx whose space is partitioned. Defining

bk,t=bk+wk,t⊤ytb_{k,t}=b_k+w_{k,t}^{\top}y_t

absorbs the time-dependent input offset into the affine constant term:

=w_{k,t}^{\top}x+b_{k,t}.$$ This treatment preserves both the time dependence of $w_{k,t}$ and the affine structure required for partitioning the input space. This connecting step is now included in Step 1 of the proof of Theorem A.2. **An independent counting proof for time-varying weights.** Under the soft-reset dynamics with $\beta=1$, expanding the recurrence gives $$s_k(t) =H\left( a_{k,t}^{\top}x+c_{k,t} -\vartheta\left[1+\sum_{\tau=1}^{t-1}s_k(\tau)\right] \right),$$ where $$a_{k,t}=\sum_{\tau=1}^{t}w_{k,\tau}, \qquad c_{k,t}=u_k(0)+\sum_{\tau=1}^{t}b_{k,\tau}.$$ The preceding spike history enters the threshold through the cumulative spike count $$m=\sum_{\tau=1}^{t-1}s_k(\tau) \in\lbrace 0,\ldots,t-1\rbrace.$$ Thus, timestep $t$ corresponds to at most the following $t$ candidate hyperplanes: $$a_{k,t}^{\top}x+c_{k,t} =\vartheta(1+m), \qquad m=0,\ldots,t-1.$$ The total number of candidate hyperplanes across all timesteps is therefore at most $$M(T)=\sum_{t=1}^{T}t=\frac{T(T+1)}2.$$ **This result follows directly from the dynamics with time-varying weights, without applying the fixed-weight lemma directly to the time-varying case.** The normal vectors can change across timesteps; coincident or degenerate hyperplanes and unreachable histories can only reduce the number of effective boundaries. Since every $a_{k,t}$ belongs to $\operatorname{span}\lbrace w_{k,1},\ldots,w_{k,T}\rbrace$, counting can be performed in the $r_k$-dimensional effective input space defined in W1. The upper bound on the number of regions in a hyperplane arrangement gives the following bound on the number of nonempty, full-dimensional activation regions of neuron $k$: $$N_k\le \sum_{j=0}^{r_k}\binom{M(T)}j.$$ Taking the common refinement of the partitions of $n$ neurons yields $$N_{\mathrm{layer}} \le \prod_{k=1}^{n} \left[ \sum_{j=0}^{r_k} \binom{M(T)}j \right], \qquad M(T)=\frac{T(T+1)}2.$$ Here, activation regions are the nonempty, full-dimensional realization sets of complete spike sequences. Each such set is convex, so the number of activation regions also satisfies $N_k\le2^T$. **The proof accommodates both fixed time-dependent input offsets and general time-dependent weights.** The $1+q$ dimension bound under a shared projection follows from the special case in W1. We have incorporated this derivation and its $\beta=1$ assumption into Theorem A.2 and its proof (Steps 1–4) in Appendix A. The revision will also incorporate this derivation and its $\beta=1$ assumption into the main theoretical section.
AuthorsOpenReview ↗

Response to Reviewer xEmk — Part 3 of 4

W3. The upper-bound comparison in Equation (9), the theoretical conjecture, and realizable expressivity gains

Thank you for pointing out the need to distinguish upper-bound analysis more clearly from comparisons of realizable expressivity. We introduce Equation (9) to explain the potential expressivity improvements from temporal modulation, starting from the input-partitioning constraints of fixed-weight SNNs. Prior work discusses differences in the representational mechanisms of discrete-time SNNs and ANNs and characterizes the parallel-hyperplane constraint under shared weights [1]. Under the input encoding, network architecture, and temporal budget analyzed in this work, time-dependent weights relax this geometric constraint, making richer input partitioning a reasonable conjecture with a structural basis.

Equation (9) provides theoretical insight at the level of upper bounds to motivate this conjecture; it does not prove that either upper bound is attainable. We have clarified the distinction between upper-bound analysis and realizable gains in Remark A.5 and Proposition A.6 in Appendix A. The revision will retain the upper-bound comparison and its geometric interpretation while making the wording reflect this role more accurately. The revision will align Equation (9) with the parameterization and dynamical assumptions in W1 and W2, explicitly restricting expressions involving 1+q1+q to the special case of a shared projection. The revision will replace expressions such as “establishes” and “confirms” with wording referring to potential gains, avoiding conflation of changes in upper bounds with a strict separation in realizable capability.

Building on this, we have added the following construction as Proposition A.6 in Appendix A to provide direct evidence of realizability for this conjecture.

Construction: strict separation with the same neuron count, input dimension, and number of timesteps. Under the above theoretical setting with β=1\beta=1, take

n=1,d=2,T=3,q=1,n=1,\qquad d=2,\qquad T=3,\qquad q=1,

and set

ϑ=1,u(0)=b=0.\vartheta=1,\qquad u(0)=b=0.

Choose the shared projection and binary routing as

(1,0)^\top, \qquad U= (-1,1), \qquad (M_1,M_2,M_3)=(0,1,1).$$ This gives $$w_1= (1,0)^\top, \qquad w_2=w_3= (0,1)^\top.$$ **This construction satisfies the compact-modulation and binary-routing constraints**; a shared projection is also an admissible choice within the time-dependent projection parameterization. For any target spike sequence $$r=(r_1,r_2,r_3)\in\lbrace 0,1\rbrace^3,$$ choose the static vector repeated as input over the three timesteps: $$x(r)= \left(r_1+\frac12,\frac18+\frac12r_2+\frac14r_3\right)^\top.$$ Let $z$ denote the second coordinate. The first-step output is $s_1=r_1$, and the post-reset potential is $u_1=1/2$. At the second step, $$s_2=H\left(z-\frac12\right)=r_2,$$ because $z\le 3/8<1/2$ when $r_2=0$, whereas $z\ge 5/8>1/2$ when $r_2=1$. At the third step, $$s_3 =H\left(2z-r_2-\frac12\right) =H\left(\frac{2r_3-1}{4}\right) =r_3.$$ The threshold decisions at all three timesteps have strictly positive margins. Thus, each target spike sequence is realized on a nonempty open input region, giving at least $2^3=8$ distinct activation regions. For the fixed-weight model under the same conditions $n=1$, $d=2$, $T=3$, and $\beta=1$, the normal vectors of all candidate boundaries are parallel, so at most $M(3)=6$ hyperplanes induce at most $7$ regions. Therefore, $$N_{\Delta}\ge 8>7\ge N_{\mathrm{Const}}.$$ **By exhibiting a realizable lower bound exceeding the baseline's upper bound, this construction rigorously proves a gain in partitioning capability under the specified setting.** The upper-bound comparison continues to provide geometric interpretation and theoretical insight, while the construction provides independent evidence of realizability. The paragraph following the proof of Proposition A.6 in Appendix A now states the scope of this comparison explicitly and distinguishes it from the experimental evaluation of parameter and computational efficiency.
Reviewer_i1CXOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Delta SNN, a spiking neural network framework for event-based vision that incorporates time-dependent low-rank weight modulation and binary synaptic routing into a shared synaptic backbone to enhance spatiotemporal modeling. The method is evaluated on object detection, human activity recognition, and video anomaly detection. Extensive experiments demonstrate that it can be effectively integrated into different SNN backbones, yielding consistent performance improvements across all evaluated datasets.

Critical strengths
  1. The paper provides a clear motivation for time-dependent synaptic adaptation, highlighting that fully shared synaptic weights may limit the model’s ability to accommodate evolving event distributions and neuronal states.

  2. The proposed module is compatible with different SNN architectures and preserves spike-driven computation, offering architectural flexibility while retaining the potential for energy-efficient inference.

  3. The experimental evaluation spans three vision tasks: object detection, human activity recognition, and video anomaly detection. Consistent performance gains across all evaluated datasets support the method’s effectiveness in diverse event-based vision settings.

Critical weaknesses
  1. Delta SNN introduces additional parameters and computation, making it difficult to distinguish the benefits of temporal modulation from those of increased model capacity, the auxiliary pathway, or changes in optimization. Comparisons with parameter- and compute-matched baselines, including a variant with a static auxiliary pathway, would help establish the contribution of the proposed temporal mechanism.

  2. The practical energy efficiency of Delta SNN requires further validation. Under the reported Gen1 setting [1], the method introduces approximately 8% more parameters and increases training and inference latency by approximately 5% and 8%, respectively. The authors should clarify whether the energy estimates include all additional operations and report single-sample inference latency or throughput under settings representative of practical deployment.

  3. Delta SNN combines a shared main pathway with a compact auxiliary pathway to enhance temporal processing, which is similar to existing methods [2]. The authors should explain how the proposed low-rank modulation and binary routing differ from these approaches and identify the specific advances beyond the shared architectural principle.

  4. The current analysis does not fully support the claimed expressivity advantage of Delta SNN. Equation (9) compares only upper bounds on the number of input-space regions, which is insufficient to establish that Delta SNN can realize more regions than conventional SNNs. The authors should also clarify whether the analysis remains valid under the binary routing constraint.

[1] A Large Scale Event-based Detection Dataset for Automotive.

[2] Algorithm–hardware co-design of neuromorphic networks with dual memory pathways, Nature Machine Intelligence.

Questions for authors
  1. How robust is Delta SNN to temporal shifts, variations in action speed, and changes in sequence length? How are the time-dependent modulation parameters selected or adjusted under these conditions? Experimental evidence would help establish whether the reported gains persist under temporal distribution shifts.

  2. Do the reported energy estimates account for auxiliary projections, accumulation operations for binary routing, memory accesses to intermediate data, and retrieval of time-dependent parameters?

  3. Could the authors compare Delta SNN with closely related methods under consistent experimental settings and comparable resource budgets? Such comparisons would help clarify its distinct contribution and quantify the trade-off between performance gains and additional computational overhead.

6 comments
AuthorsOpenReview ↗

Response to Reviewer i1CX — Part 1 of 3

We sincerely thank you for taking the time to review our paper carefully and for your thoughtful and constructive comments. We address your questions point by point below.

W1. Comparisons under matched resource budgets.

Thank you for this constructive suggestion. On Gen1 (T=4T=4, D=1D=1), we compare Delta SNN with static auxiliary pathways and wider/deeper SpikeYOLO baselines under matched resource budgets. These comparisons show that time-varying weights provide gains beyond static capacity expansion or adding a time-shared auxiliary pathway alone.

We introduce variants with a static auxiliary pathway of width qq or 4q4q at the same layers as Delta SNN, extending Table 21 (comparisons with static-pathway variants, on Gen1 with T=2T=2, D=4D=4). The auxiliary weights in our new variants are full-precision, shared across timesteps, and jointly trained with the backbone from scratch. The qq variant matches Delta SNN in FLOPs, while the 4q4q variant matches its parameter count. We additionally compare wider SpikeYOLO models matching parameters or FLOPs, and a deeper model matching both. FLOPs are computed for dense convolutions over a complete four-step clip, using Delta SNN as the reference. The comparison is summarized in the table below.

ModelMatched budgetParams (M)mAP@50 (%)mAP@50:95 (%)
SpikeYOLO—23.1266.038.4
+ Static pathway (qq)FLOPs23.5666.338.7
+ Static pathway (4q4q)Parameters24.8766.639.1
WiderParameters24.9166.538.9
WiderFLOPs23.5866.538.9
DeeperParameters+FLOPs24.9066.138.6
Delta SNNReference24.8768.941.3

The static auxiliary pathways and enlarged backbones yield only modest improvements over SpikeYOLO. Delta SNN surpasses the strongest static control by 2.3/2.2 points in mAP@50/mAP@50:95, improving mAP@50:95 by 2.9 points over SpikeYOLO versus 0.7 points for that control. Together, these comparisons support time-varying synaptic transformations as a more effective use of the additional resources than static capacity expansion or adding a time-shared auxiliary pathway alone.

 

W2. Energy estimates and inference latency.

Thank you for highlighting this point. We would like to clarify that our energy estimates account for backbone computation, auxiliary projections, and binary-routing accumulations (Table 28). We provide memory access and neuron update costs for representative modules (Table 31). We additionally extend the latter analysis to the full network (Q2 below).

Besides the per-sample latency measured on an H200 GPU in Table 4, we additionally benchmark on an RTX 4090 and include the parameter-matched widened SpikeYOLO from W1. These GPU evaluations use batch size 11 and inference mode on Gen1 with T=4T=4, D=1D=1:

GPUSpikeYOLO (ms/sample)Widened SpikeYOLO (ms/sample)Delta SNN (ms/sample)
H20028.4430.9430.66
RTX 409038.5740.9341.04

Delta SNN incurs a modest latency increase over the original SpikeYOLO (6.4–7.8%). Its latency is nearly identical to the parameter-matched widened baseline, yet both mAP@50 and mAP@50:95 are 2.4 points higher (W1).

AuthorsOpenReview ↗

Response to Reviewer i1CX — Part 2 of 3

W3. Comparison with DMP-SNN.

Thank you for highlighting this related work. We would like to clarify that Delta SNN addresses the bottleneck of temporal weight sharing, which remains underexplored in the SNN community. Our core contribution is to unlock the representational capacity of SNNs through time-varying synaptic weights, while preserving computational efficiency and neuromorphic compatibility. Below, we outline the key differences from DMP-SNN and present a controlled empirical comparison.

(1) Motivation and research problem. Dual Memory Pathway SNNs (DMP-SNN) in [1] aim to address the limited long-term memory of conventional SNNs, while our central research question is how to unleash the spatiotemporal modeling potential of SNNs through time-variant synaptic weights for event vision.

(2) Methodological differences. DMP-SNN projects incoming spikes into a memory input, maintains a weighted history in a compact memory state, and adds its readout to the membrane potential, with all weights shared across timesteps. Its nonbinary memory updates and readout involve dense multiply–accumulate operations. Delta SNN instead learns time-varying synaptic weights and uses binary routing to implement the modulation through spike-driven accumulation and selection, preserving addition-based inference. Neither time-varying synaptic-weight modulation nor binary synaptic routing is introduced in [1].

(3) Empirical comparison. We integrate DMP into SpikeYOLO on Gen1 (T=4T=4, D=1D=1) and include a widened DMP variant matching Delta SNN's parameter budget within 0.4%, with DMP's memory configuration unchanged. The results are shown below.

MethodParams(M)mAP@50(%)mAP@50:95(%)
SpikeYOLO23.1266.038.4
SpikeYOLO+DMP23.5366.138.9
SpikeYOLO+DMP, parameter-matched24.9766.939.0
SpikeYOLO+Delta SNN24.8768.941.3

Delta SNN exceeds parameter-matched DMP by 2.0/2.3 points in mAP@50/mAP@50:95 with slightly fewer parameters, supporting the practical value of our design. We have added the comparison to the revised manuscript in Appendix E.7.

 

W4. Expressivity analysis.

Thank you for your comments. We construct an explicit example under the binary routing constraint: a single Delta SNN neuron distinguishes eight input regions, whereas a constant-weight neuron under the same setting can distinguish at most seven. The eight regions produce eight different spike patterns, demonstrating that temporal modulation can distinguish more input patterns with the same neuron count, input dimension, and number of timesteps. The construction and proof are as follows.

(1) A compact construction with binary routing. Consider a single neuron with input dimension d=2d=2, T=3T=3 timesteps, and modulation width q=1q=1. We use non-leaky soft-reset dynamics (β=1\beta=1), threshold ϑ=1\vartheta=1, initial potential u(0)=0u(0)=0, and bias b=0b=0. A static input xx is repeated at all three timesteps, giving

\qquad u(t)=u(t-1)+w_t^\top x-s(t),$$ where $H$ is the binary threshold function. Set the routing scale to one and choose $$w_{\mathrm{base}}=(1,0)^\top, \qquad U=(-1,1), \qquad (M_1,M_2,M_3)=(0,1,1).$$ Then $w_t=w_{\mathrm{base}}+U^\top M_t$ gives $$w_1=(1,0)^\top, \qquad w_2=w_3=(0,1)^\top.$$ All routing values are binary, and the shared projection $U$ is an admissible choice within our time-dependent projection parameterization. **(2) Realizing eight distinct spike patterns.** For each target sequence $r=(r_1,r_2,r_3)\in\lbrace0,1\rbrace^3$, choose $$x(r)=\left(r_1+\frac12,\frac18+\frac12r_2+\frac14r_3\right)^\top.$$ Let $z$ denote the second coordinate. The first step yields $s(1)=r_1$ and $u(1)=1/2$. At the second step, $$s(2)=H\left(z-\frac12\right)=r_2,$$ because $z\le3/8$ when $r_2=0$, whereas $z\ge5/8$ when $r_2=1$. At the third step, $$s(3)=H\left(2z-r_2-\frac12\right)=H\left(\frac{2r_3-1}{4}\right)=r_3.$$ All threshold decisions have strictly positive margins. Thus, each of the eight spike patterns occurs on a nonempty open input region, giving $N_{\Delta}\ge8$. **(3) Comparison with constant weights.** For any fixed-weight neuron under the same input and dynamical settings, the decision at timestep $t$ takes the form $$s(t)=H\left(tw^\top x-1-\sum_{\tau=1}^{t-1}s(\tau)\right).$$ The preceding spike count is an integer from $0$ to $t-1$, giving at most $t$ candidate decision boundaries, all with normals parallel to $w$. Across three timesteps, at most $1+2+3=6$ parallel boundaries partition the input space into at most seven regions. Therefore, $$N_{\Delta}\ge8>7\ge N_{\mathrm{Const}}.$$ **This realizable lower bound exceeding the baseline's upper bound establishes a strict expressivity advantage under binary routing in the stated setting.** Following your suggestion, we have added this construction and its proof as Proposition A.6 in Appendix A. Thank you again for your valuable suggestions on our theoretical analysis.
AuthorsOpenReview ↗

Response to Reviewer i1CX — Part 3 of 3

Q1. Robustness under temporal distribution shifts.

Thank you for this question. Our existing experiments already cover different temporal configurations: Appendices D.1–D.2 report consistent gains for T=2,4,8T=2,4,8, input horizons of 30–500 ms, and temporal resolutions of 7.5–250 ms/bin, with models trained under the corresponding settings. Our benchmarks also include variable-length UCF-Crime-DVS sequences and HAR-DVS actions spanning diverse speeds, viewpoints, and lighting conditions.

To further test model robustness to temporal data distribution shifts between training and inference, we perturb all HAR-DVS test recordings and evaluate the frozen checkpoints of E-SpikeFormer, its parameter-matched static-modulation control, and Delta SNN (with four modulation states), all trained on the original HAR-DVS training data with T=8,D=4T=8, D=4. We apply three types of perturbations: (1) cropping each clip to the first or last 90% of its duration to assess robustness to incomplete observations; (2) increasing the playback speed to 2×2\times to assess robustness to variations in action speed; and (3) rebinning each clip over its original duration to T=16T=16 to assess robustness to longer sequence lengths. At timestep tt, we select modulation parameters (Mk,Uk)(M_k,U_k) with k=⌊4t/T⌋k=\lfloor4t/T\rfloor (zero-based indexing). All values below are Top-1 accuracy (%).

Test conditionE-SpikeFormerStatic modulationDelta SNN
Native input (T=8)48.948.851.0
Earlier crop28.521.542.2
Later crop28.626.647.8
Faster action speed (2×)47.046.349.4
Longer sequence (T=16)48.148.150.2

Delta SNN outperforms both controls in every condition, gaining 13.7–19.2 points over E-SpikeFormer under temporal cropping. These results support the role of time-variant weights in improving robustness to test-time distribution shifts.

 

Q2. Energy accounting.

Thank you for raising this important question about energy accounting. Our energy estimates include auxiliary projections and binary-routing accumulations, which together consume 0.023 mJ, only 0.20% of Delta SNN's total synaptic energy (11.53 mJ; Table 28). Following SpikedAttention's [2] energy model, we extend the module-level analysis of memory accesses and neuron updates in Appendix F.5 (Table 31) to the full network (all values below in mJ). Intermediate data denotes temporary results passed between operations in the backbone and modulation branch.

Energy componentSpikeYOLOWidened SpikeYOLODelta SNN
Backbone computation14.815.011.5
Modulation computation——0.023
Backbone parameter reads82.483.363.9
Modulation parameter reads——0.128
Intermediate data reads/writes3.053.155.35
Spike reads2.582.602.00
Spike writes0.0130.0140.013
Neuron updates1.381.431.38
Total104.22105.5084.29

With these costs included, Delta SNN consumes 19.1% less estimated total energy than SpikeYOLO and 20.1% less than the parameter-matched widened baseline. We attribute these savings to time-variant synaptic modulation improving information transmission under sparse firing, which reduces both computation and parameter reads.

 

Q3. Comparisons with related methods.

Please see our responses to W1 (static auxiliary pathways and wider/deeper SpikeYOLO) and W3 (DMP-SNN). These comparisons use consistent Gen1 evaluation settings (T=4T=4, D=1D=1) with SpikeYOLO-based models and include controls matched in parameter count, computation, or both. Delta SNN consistently outperforms these alternatives, supporting the benefit of time-varying synaptic weights under comparable resource budgets.

 

Thank you again for your time and detailed review.

 

References:

[1] Pengfei Sun, Zhe Su, Jascha Achterberg, Giacomo Indiveri, Dan F. M. Goodman, and Danyal Akarca. Algorithm–hardware co-design of neuromorphic networks with dual memory pathways. Nature Machine Intelligence, 8:901–912, 2026.

[2] Sangwoo Hwang, Seunghyun Lee, Dahoon Park, Donghun Lee, and Jaeha Kung. SpikedAttention: Training-Free and Fully Spike-Driven Transformer-to-SNN Conversion with Winner-Oriented Spike Shift for Softmax Operation. Advances in Neural Information Processing Systems, 37, 2024.

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

Delta SNN introduces time-varying synaptic weights into spiking neural networks by combining shared base weights with compact, timestep- or window-specific low-rank corrections. Binary routing allows the modulation branch to use selection and accumulation during spike-driven inference. The paper evaluates the method on event-based detection, action recognition, and anomaly detection, with additional benchmarks and analyses in the appendix.

Critical strengths
  1. Consistent improvements across tasks and backbones. Delta SNN improves several existing SNN models. For example, Gen1 mAP@50:95 increases from 40.4% to 43.3%, and HAR-DVS accuracy increases from 49.2% to 51.4% under the corresponding reported configurations.

  2. Useful controlled experiments. The ablations examine modulation width, temporal windows, and routing binarization. Parameter-matched static modulation and LoRA-style comparisons help assess whether the gains come from temporal variation rather than additional parameters alone.

  3. Detailed efficiency reporting. The paper reports GPU latency and memory costs alongside analytical energy estimates. The additional Lava functional simulations provide evidence that the proposed modulation can be implemented through spike-triggered accumulation and binary routing.

Critical weaknesses
  1. The methodological novelty appears incremental. The formulation closely resembles LoRA-style low-rank residual parameterization, with temporal variation and binary routing as the main distinctions. Although the controlled experiments support their effectiveness, the conceptual advance beyond existing low-rank adaptation ideas could be articulated more clearly.

  2. The practical advantage over ANNs remains unclear. On Gen1, Delta SNN still trails strong ANN baselines in accuracy, while its main advantage relies on estimated energy savings. End-to-end measurements on deployment hardware would help establish its practical accuracy–efficiency trade-off.

  3. The benefit of additional computation over the original SNN needs stronger justification. The method increases parameters and training and inference costs. Although parameter-matched static-modulation controls are included, comparisons against stronger configurations of the original SNN under similar computational budgets would better establish whether temporal modulation is the most effective use of the additional resources.

Questions for authors

As described in "Critical Weaknesses".

5 comments
AuthorsOpenReview ↗

Response to Reviewer uHEX — Part 1 of 2

We extend our sincere gratitude for dedicating your valuable time to reviewing our paper. In the following, we provide a comprehensive response to your comments.

W1. Advance beyond LoRA.

Thank you for highlighting the relevant study. We would like to clarify that Delta SNN addresses the bottleneck of temporal weight sharing, which remains underexplored in the SNN community. Our core contribution is to unlock the representational capacity of SNNs through time-varying synaptic weights, while preserving computational efficiency and neuromorphic compatibility. Below, we outline the key conceptual differences from LoRA and substantiate Delta SNN’s advantages through theoretical analysis and empirical comparisons.

(1) Distinct research objectives. Delta SNN is motivated by the temporal modeling bottleneck imposed by fixed synaptic weights: a shared transformation must process evolving event features and reconcile potentially conflicting gradients across timesteps. Our central research question is how to unleash the spatiotemporal modeling potential of SNNs through time-variant synaptic weights. LoRA instead targets the parameter-efficient adaptation of pre-trained large models through low-rank updates.

(2) Model operation and neuromorphic compatibility. At inference, LoRA applies a static weight adaptation, whereas Delta SNN performs time-varying synaptic transformations. Implementing these temporal transformations with full-precision LoRA-style factors requires energy-intensive floating-point multiplications, whereas pre-merging the factors into the backbone requires storing TT full weight matrices, substantially increasing storage costs. To preserve neuromorphic compatibility, Delta SNN introduces binary routing to enable energy-efficient, spike-driven, addition-only inference.

(3) Expressive advantage. Our method aims to address a theoretical structural constraint of conventional SNNs: temporal weight sharing makes each neuron’s decision boundaries parallel across timesteps. Static LoRA-style modulation increases the parameter count but cannot remove this constraint. In contrast, time-varying weights break this geometric restriction and extend the boundaries of conventional SNN expressivity.

(4) Empirical performance gains. We provide empirical evidence for the benefits of time-varying weights. Delta SNN substantially improves performance even with a modulation width of q=1q=1 (Figure 3), whereas all four static LoRA variants underperform the baseline (Table 22). Under the same parameter budget, Delta SNN exceeds the LoRA variant by 2.5 and 2.7 percentage points in mAP@50 and mAP@50:95, respectively. These results support the value of learned temporal variation in enhancing SNN performance beyond static parameter expansion.

 

W2. Accuracy–efficiency trade-off against ANNs.

Thank you for your constructive suggestions. We would like to clarify that our original Gen1 experiments follow the SpikeYOLO pipeline without data augmentation to isolate the effect of time-variant weights. However, applying standard data augmentation enables Delta SNN to achieve accuracy comparable to that of a strong ANN baseline. We draw on RVT [1], a strong ANN baseline that reports a 9.5-point gain in Gen1 validation mAP from horizontal flipping and zoom-in/out (Table 7 in [1]). We apply these standard augmentations to both SpikeYOLO and Delta SNN with other settings unchanged. Results are shown below.

ModelArchitectureData augmentationEnergy(mJ)mAP@50:95(%)
RVT [1]ANN✓116.747.2
SpikeYOLOSNN✗12.940.4
Delta SNNSNN✗12.043.3
SpikeYOLOSNN✓11.844.0
Delta SNNSNN✓11.747.5

With standard data augmentation, Delta SNN (T=4T=4, D=2D=2) reaches 47.5% mAP@50:95, comparable to RVT (47.2%), with about 90% lower estimated energy (11.7 vs. 116.7 mJ). It also exceeds SpikeYOLO with identical augmentation by 3.5 percentage points, demonstrating consistent gains from temporal modulation beyond augmentation.

Our results on other event vision datasets further demonstrate Delta SNN’s competitiveness with ANN models. It performs comparably to the ANN models TimeSformer and Video-SwinTrans on HAR-DVS (Table 2) and outperforms all reported ANN baselines on UCF-Crime-DVS (Table 3).

For energy comparisons, we report analytical estimates based on synaptic operation counts and per-operation energy costs. This is common practice in the SNN/neuromorphic computing community, where access to mature hardware platforms remains limited. To substantiate the efficiency advantage, we also account for memory-access costs (Appendix F.5, Table 31) and demonstrate neuromorphic compatibility through Lava simulations (Appendix F.4, Table 30). In addition, our ongoing work progresses toward deployment and measurements on neuromorphic chips.

References: [1] M. Gehrig and D. Scaramuzza. Recurrent Vision Transformers for Object Detection with Event Cameras. CVPR 2023.

AuthorsOpenReview ↗

Response to Reviewer uHEX — Part 2 of 2

W3. Comparisons against stronger configurations of the original SNN.

Thank you for this constructive suggestion. We further construct three static SpikeYOLO controls on Gen1 (T=4T=4, D=1D=1): two widened variants and one deepened variant. The two widened variants increase the global channel-width multiplier while keeping depth and topology unchanged. Since FLOPs also depend on feature-map resolution, a single width multiplier offers limited flexibility to match both budgets, so we match parameters and FLOPs separately. The deepened variant adds a residual block and adjusts its hidden width to match both. Each targeted budget is matched within 0.2%. FLOPs are computed for dense convolutions over a complete four-step clip and normalized to Delta SNN. The comparison is summarized in the table below.

ModelMatched budgetParams (M)FLOPs/DeltamAP@50 (%)mAP@50:95 (%)
SpikeYOLO—23.120.9855×66.038.4
SpikeYOLO (wider)Parameters24.911.0601×66.538.9
SpikeYOLO (wider)FLOPs23.581.0001×66.538.9
SpikeYOLO (deeper)Parameters+FLOPs24.900.9985×66.138.6
Delta SNNReference24.871.0000×68.941.3

The enlarged static baselines offer only marginal gains over the original SpikeYOLO. Delta SNN exceeds their best scores by 2.4 points on both metrics, despite parameter-matched widening using 6.0% more FLOPs. These results support time-varying synaptic transformations as a more effective use of the additional resources than static capacity expansion. We have added this comparison to Appendix E.6 of the revised manuscript.

Thank you again for your valuable comments.

AuthorsOpenReview ↗

A Brief Follow-up and Invitation for Further Feedback

Dear Reviewer uHEX,

Thank you for your careful review and constructive comments. As the discussion period draws to a close, we would appreciate your feedback on whether our responses have addressed your concerns. We would be happy to provide any further clarification.

Thank you again for your time and consideration.

Best regards,

The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

Dear Area Chair and Reviewers,

We sincerely appreciate the time and care you have devoted to reviewing our work. We have provided detailed responses to all concerns and questions and revised the manuscript accordingly. We believe these responses and revisions address the issues raised and clarify the scope and contributions of our work. Your constructive suggestions have helped us improve the paper’s clarity and completeness. All revisions are marked in blue, and our point-by-point responses include references to the relevant sections, figures, and tables to make the changes easy to locate. Thank you again for your thoughtful feedback. We welcome further discussion of any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.