acceptodds
Under review as a conference paper at ICLR 2027

RIPE-MAMBASPIKE: esolution-ndependent Spiking–State-Space Interfaces for arameter-fficient Event-Based Vision

Abstract

Spiking–Mamba hybrids reach strong accuracy on event-based vision, but existing designs often require tens of millions of parameters. Much of that cost comes from how the spiking front-end is connected to the state-space backbone rather than from the hybrid architecture itself. In a representative model, a single resolution-dependent projection accounts for 33.55M of 36.25M parameters. To that end, we introduce RIPE-MambaSpike (esolution-ndependent, arameter-fficient), which replaces that projection with a hierarchical multi-resolution bridge of fixed channel width. Its deployed footprint is 0.870M parameters, constant at fixed time steps and widths across a 43× range of input areas. Reparameterized spiking stages, temporal decoupled modulation, and a dynamic convex-hull-bounded dual-stream membrane-potential attention preserve accuracy under this compact design. Result-wise, RIPE-MambaSpike is pareto-optimal on CIFAR10-DVS, N-Caltech101, and DailyDVS-200. Notably, on the 200-class DailyDVS–200, a scaled 8.04M configuration achieves 45.7% top-1 accuracy, the best reported spiking result on that benchmark, and outperforms prior spiking methods with 3.0-15.1× fewer parameters than dense ANNs. Overall, our findings demonstrate that competitive event-based recognition does not require resolution-dependent parameter growth.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Spiking Neural Networks, Mamba, State-Space Models, Event-Based Vision, Parameter-Efficient Learning, Structural Reparameterization, Efficient Deep Learning.

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

This paper questions why spiking state-space models for event-based vision are so large, and locates the cost at the interface joining the spiking front end to the state-space backbone rather than in the model's capacity. Profiling a released Mamba-Spike model, the authors find that nearly all of its parameters sit in a single projection layer that flattens the front-end feature map before the scan, and that this layer's size grows quadratically with sensor resolution. Their method, RIPE-MambaSpike, pools every spiking stage to a common grid and projects at a fixed width, which removes the dependence of the deployed parameter count on input resolution. The remaining components follow the same principle of paying cost during training and recovering it at inference. The spiking stages train as several parallel convolution branches and fuse into a single kernel once training ends. A per-channel first-order temporal filter, carrying one learnable scalar per channel, precedes each convolution. The attention module draws its queries and keys from the membrane potential and its values from the previous spike tensor, and uses a strictly positive feature map so that the output remains inside the convex hull of the value vectors. The authors support these claims with parameter counts obtained by direct instantiation across input resolutions, a controlled swap of the interface inside the released baseline with all other components held fixed, a component ablation, and five-seed results on five event benchmarks. They report a Pareto-optimal position among published methods with reported parameter counts on three of those benchmarks and the strongest spiking result on the largest benchmark they evaluate, while stating that the model does not lead on accuracy on any of them.

Critical strengths
  1. The paper makes its central diagnosis causal rather than correlational, and does so with an unusually clean control. The claim that the interface rather than the capacity determines the parameter cost is supported by an intervention that replaces only the projection joining the front end to the scan, while the spiking front end, the spike-to-activation conversion, the Mamba blocks, the normalization and the classifier are imported unmodified from the released baseline. The intervention removes roughly 7.6$\times$ of the model while gaining 4.40 percentage points on the same protocol, and the diagnosed layer is shown by direct instantiation to hold 92.6% of parameters at 1282128^{2} and 97.4% at 2242224^{2}. For the single model profiled, this establishes the diagnosis far more convincingly than a parameter table alone would.
  2. The reported numbers are checkable, and the paper is explicit about the limits of its efficiency claim. The deployed parameter counts can be reproduced without downloading any data, which makes the central quantity easy to verify. The energy discussion is confined to an estimate of accumulate-only cost for the layers where the binary-operand property actually holds. The rest of the network is left uncosted rather than folded into the headline, and the paper states outright that it makes no energy, operation-count or latency claim on ordinary runtimes. It also notes that resolution-independence applies to weight storage alone. Activation memory still grows with resolution, and at the larger batch sizes tested the model can end up slower and larger in total memory than the baseline.
Critical weaknesses
  1. The central claim is stated at the level of a class of models, while the evidence base is a single profiled architecture and an intervention performed on that same architecture. The abstract and the introduction assert that the parameter cost of Spiking-Mamba hybrids is a property of the interface rather than of capacity, and the conclusion generalizes this into a practical recommendation for the field ("competitive event-based recognition does not require resolution-dependent parameter growth"). The measurement behind it profiles one released model, and the causal evidence is a swap performed inside that same model. The paper does acknowledge the limitation once, in the discussion ("the diagnosis rests on one released model, so we scope it to flatten-then-project bridges"), but that qualification does not appear in the title, the abstract, or the contribution list, where the general framing remains. The consequence is that a reader cannot determine whether the paper has identified a property of this model class or a property of one particular implementation, and the recommendation offered to the field is supported only for the bridge family the authors have already selected. This is consequential because it is the paper's central claim, not a supporting detail.
  2. The comparative claims are made on a cross-protocol basis, and the controlled comparison is against a reproduction that the paper itself could not reconcile with the published baseline. The reported Pareto-optimality on three benchmarks and the 8.80-point lead on DailyDVS-200 are statements about the published literature, yet the entries they are measured against are quoted from other papers trained under different protocols and augmentation, which the paper marks but does not correct for. Where a controlled comparison does exist, it is against the authors' own reproduction of the released baseline: on CIFAR10-DVS their model reaches 81.40, below the published baseline figure of 92.50 and above their own reproductions of the released code at 65.30 and 48.90. The consequence is that the paper's competitive position depends on which reference a reader adopts, and under the published figure the model is not more accurate, only smaller. Combined with the fact that the model leads on accuracy on no benchmark, the contribution is a parameter-efficiency frontier position rather than an advance that a reader can rely on without choosing a protocol.
Questions for authors
  1. The appendix specifies a parameter-matched sigmoid-gated attention as the natural alternative to the proposed bounded attention and states that it was not trained. Would training that control be feasible, so that the module's claimed advantage can be measured rather than argued?
  2. The presented diagnosis as a property of spiking state-space hybrids, but it is measured on one model. Could the authors add the parameter share held by the front-end-to-scan interface for a second released hybrid that uses a different kind of bridge, for instance a window-based or token-pruned coupling? A second data point would show whether the interface dominates across the family or only in flatten-then-project designs, which is what I need in order to judge how far the finding generalizes.
  3. The paper states that the deployed model scans every token the bridge produces, and Table 13 lists the full grid as the deployed configuration; the released per-benchmark defaults, however, set the retained-token budget below the full grid for three of the five benchmarks. Could you confirm which setting produced the reported numbers, in particular for CIFAR10-DVS and N-Caltech101, where the budget sweep of Table 13 is not reported? Table 13 establishes the budget is accuracy-neutral on DVS-Gesture, so this is a question of reproducing the stated configuration rather than of the reported accuracy.
4 comments
AuthorsOpenReview ↗

Response to Reviewer oekq

We thank the reviewer for a careful review and for recognizing the interface swap as a causal control.

W1 / Q2: one profiled model, class-level claim. We agree that the scope belonged in the abstract and thesis, not only in the Discussion. As requested, we profiled a second Spiking–Mamba hybrid with public code and a different bridge. SpikMamba (Chen et al., 2024) tokenizes per 4×4 patch at a fixed width with learned positional embeddings, without a global flatten-and-project. Instantiated at 34²–640² (script in the supplement):

InputMamba-Spike interfaceSpikMamba interface
34²2.10 M (43.8%)0.036 M (2.0%)
128²33.55 M (92.6%)0.281 M (13.9%)
224²102.76 M (97.4%)0.822 M (32.0%)
640²838.9 M (99.7%)6.57 M (79.0%)

The second model confirms and refines the diagnosis. In both hybrids, the interface is the only resolution-dependent component (SpikMamba's sequence stack is a constant 1.74 M), so in both it eventually dominates. The bridge design sets only the slope:

  • Flatten-then-project costs C⋅d/16=2048C\cdot d/16 = 2048 parameters per input pixel and holds the majority from ≈36².
  • SpikMamba's per-patch positional table costs 16 parameters per pixel, exactly 128=C128 = C times less. It crosses 50% at ≈330².

The general rule: resolution-dependent parameters come from trainable tensors indexed by the spatial grid (a flattened dense projection, a learned positional table). Our bridge has none: shared projections and a coordinate-based positional encoding, so its slope is zero.

Revisions. The abstract, thesis and contributions are now scoped to flatten-then-project bridges; the introduction, Discussion and Table 7 report the second profile.

W2: cross-protocol comparisons. The primary comparison is controlled and reproducible. The supplement ships Mamba-Spike's official implementation (https://github.com/ecnu-cross-innovation-lab/mamba-spike) with two scripts: its own protocol (‡) and the protocol identical to ours (†). Parameters come from its default configuration:

OursMamba-Spike † / ‡GainParams
CIFAR10-DVS81.4065.30 / 48.90+16.10 / +32.500.87 M vs 36.25 M (41×)
DVS-Gesture98.4895.83 / 94.32+2.65 / +4.160.88 M vs 36.25 M (41×)
N-MNIST99.5799.49 / 99.36+0.08 / +0.210.87 M vs 1.03 M

Under either protocol we are more accurate (also on five-seed means: 81.28 / 98.18 / 99.55) and up to 41× smaller. Our protocol also raises the baseline by 16.4 pp on CIFAR10-DVS. The interface swap (App. C.2) isolates the cause: changing only input_proj gives 7.6× fewer parameters and +4.40 pp.

The published 92.50 is the exception: its configuration and parameter count are not stated, so we could not reconcile it with the official code (App. C.3). Against it we are smaller, not more accurate, and §6 now says so. Against the literature we claim frontier position, not accuracy dominance (§6).

Q1: sigmoid-gated control. Yes, and the result supports the design. We trained the parameter-matched control A=σ(QK⊤/C)V/NA = \sigma(QK^\top/\sqrt{C})V/N on N-Caltech101. It shares DS-MPA's WQ,WK,WVW_Q, W_K, W_V and γ\gamma, with 1/N1/N as its normaliser, so the arms differ only in the kernel:

KernelN-Caltech101
DS-MPA (5 seeds)85.49 ± 0.18
Sigmoid gate (3 seeds)83.92 ± 0.18 (−1.57 pp, Welch p=2×10−4p = 2\times10^{-4})
No attention (A1)82.38

A sigmoid gate recovers only half of DS-MPA's gain over no attention. Both outputs are bounded (the control's weights σ/N\sigma/N also sum to ≤1\le 1), so the difference lies in DS-MPA's normalized, factorized affinity, which is also cheaper: O(NC2)O(NC^2) against O(N2C)O(N^2C) for pairwise sigmoid attention. Gradients are descriptive: equal total norm (1.04 vs. 1.03) and a slightly smaller stage-1 share under sigmoid (10.7% vs. 11.8%), consistent with the saturation argument. §4.3 and App. B.5 now report it.

Q3: token budget. Every reported number uses the full grid, with no pruning: all benchmarks, ablations and seeds. The --top-k defaults in our released scripts (128/128/64) are leftovers from an earlier pruned variant. We have set them to 0 (full grid) in the code and README; thank you for catching this.

Why we removed pruning. Table 14 shows pruning leaves accuracy unchanged but buys almost nothing:

  • Parameters: none saved. The Mamba blocks are O(d2)O(d^2) and independent of sequence length, so the model is 0.875 M at every kk.
  • Compute: marginal. Halving the tokens saves 5.7% of FLOPs, and keeping an eighth saves 9.9% (Table 21), because the convolutional front-end dominates.
  • Complexity: pruning added a non-differentiable top-k selection and per-sample positional encoding.

So the deployed model scans every token.

We thank the reviewer again for the insightful review. Please let us know if you have any further comments or questions. we look forward to a fruitful discussion.

I thank the authors for a substantial rebuttal. The scope revision and the second profile answer my questions of fact. My reservation is now about the size of what remains.

The second profile shows that at the resolutions this paper evaluates, a hybrid whose interface is a per-patch positional table spends 13.9% and 32.0% of its parameters at the interface, and crosses 50% only near 330². SpikMamba is a fixed-resolution model, so those instantiations demonstrate the mechanism rather than an operating regime in which the cost binds. What remains is a standard remedy for a known inefficiency, applied to one released baseline and packaged with three components whose ablations contribute between one and three points each. The model leads on no benchmark, and the 41× comparison is drawn against a reproduction that sits 27 points below the published baseline.

AuthorsOpenReview ↗

Response to Reviewer oekq's Follow-up

We thank the reviewer for engaging with the rebuttal and are glad the questions of fact are resolved. We respond briefly to the remaining points.

1. Where the cost binds. We agree that SpikMamba's profile shows the mechanism. The operating regime where the cost binds is the native resolution of current event sensors, which is where a deployed model would run. Using the per-pixel slopes measured by instantiation (exact up to 640²; now in App. C.1):

Sensor (native)Mamba-Spike interfaceSpikMamba interfaceOurs (total)
DAVIS346 (346×260)184 M (98.5%)1.46 M (45%)0.870 M
DVXplorer / Prophesee Gen3 (640×480)629 M (99.6%)4.9 M (74%)0.870 M
Prophesee IMX636 (1280×720)1.89 B (99.9%)14.8 M (89%)0.870 M

For flatten-then-project designs the cost already binds at every resolution we evaluate (92.6% at 128², 97.4% at 224²). At sensor resolution it binds for both bridge families. A fixed-width bridge removes it at all of them.

2. A standard remedy for a known inefficiency. We agree that pooling to a fixed width is standard, and we do not claim it as new. What was not established is that this layer is the binding cost of the model family: the official Mamba-Spike implementation places 92.6% of its parameters there, and replacing only that layer gives 7.6× fewer parameters and +4.40 pp under the same protocol. MR-S3A also goes beyond the standard remedy. The textbook version (pool all stages, concatenate, one shared projection) loses 1.49 pp to it at a matched parameter count (Welch p=3×10−4p = 3\times10^{-4}, App. C.15), and MR-S3A adds 2.96 pp over a stage-3 projection even with DS-MPA, RepConv and TDM all removed (p=6×10−5p = 6\times10^{-5}). The gain is free at deployment, since MR-S3A folds exactly into one convolution at inference.

3. Component effects of one to three points. On N-Caltech101, where the five-seed standard deviation is 0.18 pp, these effects are 8–18 seed standard deviations. The other components also beat their natural alternatives at a matched parameter count: a sigmoid kernel in place of DS-MPA loses 1.57 pp and a ConvGRU front-end 3.15 pp (Welch p≤2×10−4p \le 2\times10^{-4}).

4. The 41× comparison and the published 92.50. The 41× is a parameter ratio measured on Mamba-Spike's official implementation; it does not depend on accuracy. Mamba-Spike's paper attributes its published figures to a larger configuration whose size it does not state, so against the published model the ratio can only grow.

Where the published accuracy is reproducible, we exceed it. On DVS-Gesture our 98.48% (98.18 ± 0.31 over five seeds) is above Mamba-Spike's published 97.80%, and its official code reaches 95.83% under our protocol.

CIFAR10-DVS is the exception. The official code gives 48.90 under its own protocol and 65.30 under ours. The published 92.50 would exceed the best published spiking result we know (QKFormer, 84.00) by 8.5 pp, and the baselines in that paper also sit above those methods' own published values, which points to a different protocol (App. C.3). We therefore make no accuracy claim against 92.50; there our claim is parameter efficiency under a controlled comparison.

We agree the contribution is accuracy per parameter rather than absolute accuracy. On that axis the model is Pareto-optimal on three benchmarks and gives the best spiking result on DailyDVS-200, by 8.80 pp.

We thank the reviewer again for the constructive discussion and would be glad to clarify anything further, we would be grateful if the reviewer could please consider revisiting the evaluation in light of the revision.

AuthorsOpenReview ↗

Feedback Request to Reviewer oekq

We thank the reviewer again for the insightful review. Please let us know if you have any further queries; we look forward to a fruitful discussion.

Reviewer_vi5TOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents RIPE-MambaSpike. A fixed-width multiscale bridge connects reparameterized spiking stages with temporal modulation and dual-stream membrane-potential attention to a hierarchical bidirectional Mamba backbone.

Critical strengths

None

Critical weaknesses
  1. The shared projection and pooling used in MR-S3A are existing methods, and the supporting modules mainly use existing techniques like re-parameterization, temporal filtering, and linear attention. So far, the paper hasn’t fully explained the unique advantages of this design. It would be good to do an ablation study comparing it with just using shared projection and pooling to show what benefits these extra designs actually bring.

  2. Being able to handle temporal information doesn’t automatically mean it’s better than an RNN encoder. Since the authors haven’t claimed any energy or latency advantages on regular hardware, what’s the main reason for choosing an SNN front-end? Also, the energy estimation in the paper mainly covers the convolution part. It would be useful to explain the overhead caused by dual-core execution, including weight storage, data movement, and buffering.

  3. The claim of “no need for surrogate gradients” seems to only hold when computing gradients for the TDM input. The authors should clarify the scope of this claim, and specify whether it’s limited to the TDM input, as well as how gradients for the upstream membrane potential and weights are actually calculated.

Questions for authors
  1. Shared projection and pooling are not new, and the other modules use existing techniques such as re-parameterization, temporal filtering, and linear attention. What do these additions bring to MR-S3A? Please compare the full model with a baseline using only shared projection and pooling.

  2. The ability to process temporal information does not explain why an SNN is preferable to an RNN here. What motivates the SNN front-end, given that the paper does not show energy or latency gains on conventional hardware? Also, the energy estimate mainly covers convolutions. Does the model still save energy when weight storage, data movement, and buffering for dual-core execution are included?

  3. The claim of “no need for surrogate gradients” seems to apply only to gradients with respect to the TDM input. Is this understanding correct? Please clarify how gradients are computed for the upstream membrane potentials and weights.

3 comments
AuthorsOpenReview ↗

Response to Reviewer vi5T

We thank the reviewer for the insightful remarks.

What is new. The reviewer correctly notes that the components are individually known, however our novelty lies in the efficiency gains and the new insights the proposed architectural integration provides. In particular, we present:

  1. A causal diagnosis. In Mamba-Spike's official implementation, 92–97% of parameters sit in the interface. Swapping only that layer cuts the model 7.6× and gains 4.40 pp (App. C.2).
  2. Resolution-independent parameters. The deployed count is constant at 0.870 M across a 43× range of input area.
  3. Accuracy at that size. We are Pareto-optimal on three benchmarks and give the best spiking result on DailyDVS-200.

Its three mechanisms are proven exact (Props. 1–3: fusion, bounded attention, dual-kernel equivalence).

Q1: MR-S3A vs. a simpler interface. Any fixed-width pooled projection gives resolution independence, so the question is whether MR-S3A's multi-scale fusion earns its place. As requested, we tested the simplest shared alternative: pool all stages to the stage-3 grid, concatenate, and apply one 3×3 projection. All else is unchanged; parameters match (129,088 vs. 129,219). On N-Caltech101 it is worse:

Interface (matched parameters)N-Caltech101
MR-S3A (5 seeds)85.49 ± 0.18
Shared projection (3 seeds)84.00 ± 0.19 (−1.49 pp, Welch p=3×10−4p = 3\times10^{-4})

Why. After freezing BN statistics, MR-S3A folds exactly into one convolution over the concatenated scales, so both bridges compute the same class of function at inference. The +1.49 pp is a gain of MR-S3A's training parameterization, as with RepConv.

Q2: why an SNN front-end rather than an RNN encoder. We built it: three ConvGRU stages replace ours, widths solved to match our parameter count (889k vs. 881k); all else is unchanged. The spiking front-end wins on both axes:

Front-end (matched parameters)N-Caltech101Neuromorphic est. energy: front-end / model
Spiking (ours)85.49 ± 0.180.53 / 5.39 mJ
ConvGRU (3 seeds)82.34 ± 0.14 (−3.15 pp, p=6×10−7p = 6\times10^{-7})11.1 (21×) / 12.97 mJ (2.4×)

On conventional hardware, where spikes run as dense MACs, it wins on accuracy at equal cost: +3.15 pp at slightly lower whole-model compute (12.48 vs. 12.97 mJ all-MAC, Table 22). On neuromorphic hardware the advantage becomes pronounced, because only spikes have an accumulate-only form: 21× lower front-end and 2.4× lower whole-model energy. Q3 shows it running on SpiNNaker and fitting one chip.

Q3: whole-model energy, storage, data movement. We agree the earlier estimate was partial. NeuroBench now measures the whole model against a trained Mamba-Spike under identical metrics (App. C.14, Table 22; DVS-Gesture, per sample):

OursMamba-Spike
Weights (incl. dual kernel)1.16 M36.25 M
Activation sparsity89.4%98.5%
Effective MACs / ACs0.80 G / 0.36 G0.05 G / 0.04 G
Est. compute energy (45 nm)6.25 mJ0.27 mJ

Dense components (DS-MPA, bridge, Mamba) are 89% of our compute; the dual-kernel backbone is 10%.

Dual-kernel costs. Prop. 3 folds TDM's subtraction into the weights, so each spiking convolution runs as two accumulate-only kernels over binary spikes: Wa=W⋆diag(1+λ)W_a = W^\star\mathrm{diag}(1+\lambda) on the current input xtx_t and Wb=W⋆diag(λ)W_b = W^\star\mathrm{diag}(\lambda) on the previous input xt−1x_{t-1} (W⋆W^\star the fused kernel, λ\lambda TDM's per-channel weights). The extra cost is WbW_b (285,696 weights, +32%, included above) and a 52 KiB buffer for the binary xt−1x_{t-1}. The dual-kernel graph reproduces the fused model on all 264 test clips; NeuroBench classifies its convolutions as AC. On SpiNNaker, one such trained layer, replaying its own spikes, matches PyTorch at r=0.9999998r = 0.9999998 (App. C.14).

Data movement. Mamba-Spike's very sparse front-end computes less, and we make no energy claim over it. Our advantage is weight storage: 41× fewer weights (31× with dual kernels). At 8 bits our weights (1.2 MB) fit one Loihi / Loihi 2 (16 MB) or SpiNNaker 2 (19 MB) chip. Mamba-Spike's weights (36.3 MB) need 2–3 chips at 128² and ≥119 at HD (Table 4). Our model thus meets the prerequisite for single-chip deployment on every chip listed; Mamba-Spike on none.

Q4: scope of the surrogate-free path. We agree and have narrowed the claim. It concerns TDM's local Jacobian: holding xtx_t fixed, ∂x~t/∂xt−1=−diag(λ)\partial \tilde x_t/\partial x_{t-1} = -\mathrm{diag}(\lambda). This direct temporal path bypasses the recurrence of the neuron that emitted xt−1x_{t-1}. Loss gradients still reach it through downstream surrogates, and upstream weights need the upstream surrogate; training is not surrogate-free end to end. §4.2 and App. B.4 now say this.

We thank the reviewer again for the insightful review. Please let us know if you have any further comments or questions. we look forward to a fruitful discussion.

AuthorsOpenReview ↗

Feedback Request to Reviewer vi5T

We thank the reviewer again for the insightful review. Please let us know if you have any further queries; we look forward to a fruitful discussion.

Thank you for the response. However, my main concerns remain insufficiently addressed. In particular, the response to Question 3 is still unclear: although the proposed method has fewer parameters than Mamba-Spike, its energy consumption is substantially higher (6.25 vs. 0.27). What, then, is its practical advantage in terms of efficiency? I therefore maintain my original score.

Reviewer_18ieOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RIPE-MambaSpike, a compact event-vision architecture that replaces the resolution-dependent projection in Mamba-Spike with a fixed-width multi-resolution interface. It further combines reparameterized spiking stages, temporal modulation, membrane-potential attention, and hierarchical Mamba blocks.

Critical strengths
  1. MR-S3A replaces the flatten-then-project interface with a fixed-width multi-resolution representation.

  2. Evaluation includes five event-based datasets with different resolutions, class counts, and task difficulty. The method is evaluated on five event-based datasets, including the more challenging 200-class DailyDVS-200 benchmark.

  3. The paper separately examines MR-S3A, DS-MPA, TDM, RepConv, model width, temporal steps, and interface alternatives.

Critical weaknesses
  1. The claimed 41× reduction and accuracy improvement do not compare against the same published model configuration. Most other Table 1 comparisons are also cross-protocol.

  2. Several baseline results use different protocols. Published numbers and reproduced results may differ in preprocessing, augmentation, temporal steps, and training schedules.

  3. The claimed novelty is primarily architectural integration. Most components adapt established ideas rather than introduce a fundamentally new learning mechanism.

  4. Accuracy improvements are inconsistent. The method does not achieve the highest accuracy on CIFAR10-DVS, N-Caltech101, DVS-Gesture, or DailyDVS-200 overall.

  5. Hardware-efficiency claims lack direct validation. The accumulate-only interpretation is not supported by experiments on neuromorphic or edge hardware.

Questions for authors
  1. Which Mamba-Spike accuracy and parameter count correspond to the same architecture, checkpoint, and evaluation protocol? Can the authors provide a fully controlled Mamba-Spike comparison using identical preprocessing, augmentation, resolution, and temporal steps?

  2. Can the authors replace the current sigma analysis with a statistical comparison incorporating variance from both full and ablated models? Does the parameter count remain constant when temporal steps or channel widths change, rather than only when input resolution changes?

  3. Can the authors provide measurements on neuromorphic or edge hardware to substantiate the claimed deployment efficiency?

  4. How much improvement comes specifically from MR-S3A when all other RIPE components are removed?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 18ie

We thank the reviewer for the careful reading and for noting our five-benchmark, component-level evaluation.

W1 / Q1: same configuration, fully controlled. The 41× and the accuracy gains are measured against one model: Mamba-Spike's official implementation (https://github.com/ecnu-cross-innovation-lab/mamba-spike) in its default configuration, trained by us. Its accuracy and 36.25 M parameters are both measured on that model; we never pair 41× with the published 92.50, which states no parameter count. The run marked † (code in the supplement) uses identical preprocessed event tensors and split, 128² input, T=10T = 10, augmentation, batch, learning rate, schedule, EMA and 300 epochs. Only model-specific settings differ: weight decay (0.03 vs. 0.06) and the loss (CE vs. per-timestep TET). Under that protocol:

Ours (5-seed mean)Mamba-Spike †Params
CIFAR10-DVS81.28 ± 0.1165.300.87 M vs. 36.25 M
DVS-Gesture98.18 ± 0.3195.830.88 M vs. 36.25 M
N-MNIST99.55 ± 0.0299.490.87 M vs. 1.03 M

Our protocol also raises the baseline by 16.4 pp on CIFAR10-DVS. The introduction now says "under an identical protocol", and App. C.3 lists the shared settings.

W2: other baselines. Agreed. They are quoted (∗) and used only to place our model on the published accuracy–parameter frontier, never as controlled comparisons.

W3: novelty. The reviewer is correct that the components are individually known, however our novelty lies in the efficiency gains and the new insights the proposed architectural integration provides. In particular, we provide (i) a causal diagnosis of where a hybrid's parameters go (92–97% in Mamba-Spike's interface; App. C.2), (ii) parameters constant at 0.870 M across a 43× range of input area, and (iii) exact identities for fusion, bounded attention and dual-kernel execution.

W4: accuracy. Agreed; our goal is the best accuracy per parameter, not the highest accuracy (§6: "it leads on accuracy nowhere"). We are Pareto-optimal on three benchmarks and give the best spiking result on the challenging, larger DailyDVS-200, by 8.80 pp at 8.04 M.

Q2: statistics and parameter scaling. The original removal ablations have one run each, so their prediction-interval tests now state that they assume equal seed variance; under it, all eight CIFAR10-DVS / N-Caltech101 removals pass Holm. Where both arms are replicated, the four new controls (sigmoid kernel, shared projection, ConvGRU front-end, stripped MR-S3A; 3 vs. 5 seeds) are all significant by Holm-corrected Welch tests (adjusted p≤4×10−4p \le 4\times10^{-4}).

Parameters are not constant in TT or width, nor do we claim so. From T=4T = 4 to 32, parameters go 0.864 → 0.890 M (+3% for 8×). Width scales quadratically: 3.36 M at 2× and 13.17 M at 4× (Table 21). Both are design choices; resolution is fixed by the sensor, which is why the paper addresses it.

W5 / Q3: hardware. We now test the accumulate-only interpretation on neuromorphic hardware.

  • On SpiNNaker. Stage-3 conv2 of the trained model runs as Prop. 3's two accumulate-only synaptic projections, replaying its own spikes from 4 DVS-Gesture test clips (1.71 M synaptic events). Its summed input matches PyTorch at r=0.9999998r = 0.9999998 (relative RMSE 5.8×10−45.8\times10^{-4}), with no lost packets (App. C.14, Table 24).
  • Whole network, standardized. NeuroBench on the full test set: 1.16 M weights, 89.4% activation sparsity, 0.36 G ACs and 0.80 G MACs per sample (Table 22); it classifies all ten dual-kernel convolutions as accumulate-only, and the dual-kernel graph reproduces the fused model on all 264 test clips.
  • Storage feasibility. At 8 bits our weights (1.2 MB) fit the on-chip memory of one Loihi / Loihi 2 (16 MB) or SpiNNaker 2 (19 MB) chip with >90% to spare. Mamba-Spike's weights (36.3 MB) need 2–3 chips at 128² and ≥119 at HD (Table 4). Our model meets the prerequisite for single-chip deployment on every chip listed; Mamba-Spike on none.

Q4: MR-S3A alone. We removed DS-MPA, RepConv and TDM together and varied only the interface (N-Caltech101, 3 seeds each):

Stripped model (no DS-MPA, RepConv, TDM)N-Caltech101
with MR-S3A83.72 ± 0.21
with stage-3 projection only80.76 ± 0.19
MR-S3A effect+2.96 pp (Welch p=6×10−5p = 6\times10^{-5})

The benefit persists without the other components and is larger there (+2.96 vs. +1.41 pp). Two other controls rule out the remaining explanations. The gain is not extra parameters: an equal-size shared projection loses 1.49 pp (App. C.15). Nor is the 41× bought with accuracy: replacing only Mamba-Spike's interface with a pooled one gives 7.6× fewer parameters and +4.40 pp (App. C.2). The gain and the saving both come from the interface design.

We thank the reviewer again for the insightful review. Please let us know if you have any further comments or questions; we look forward to a fruitful discussion.

AuthorsOpenReview ↗

Feedback Request to Reviewer 18ie

We thank the reviewer again for the insightful review. Please let us know if you have any further queries; we look forward to a fruitful discussion.

Thank you for the detailed response and the substantially expanded revision. My main concern regarding the strength and scope of the contribution remains.

The central contribution is still primarily an integration of established components around a fixed-width multi-scale interface. The propositions mainly formalize reparameterization, convex-combination boundedness, and the commutation of linear temporal filtering with convolution; they do not provide guarantees concerning learning, generalization, or the claimed accuracy-efficiency trade-off. The primary empirical comparison also remains difficult to interpret. The published CIFAR10-DVS accuracy of Mamba-Spike is 92.50%, whereas the official-code reproductions obtain 48.90% and 65.30%.
Finally, the SpiNNaker experiment validates only part of one convolutional layer rather than complete-model deployment. Quantized full-network mapping, measured latency, and energy remain unverified, while the reported whole-model estimates show substantially higher compute energy and activation memory than Mamba-Spike.

I appreciate the considerable revision effort, but I still view the work as an incremental architectural integration whose theoretical and deployment evidence is insufficient for acceptance. I therefore maintain my score of 2.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers; their questions led to nine new experiments and analyses (§5, §7; App. C) and four scoped claims. In short: a 0.87 M-parameter, resolution-independent Spiking–Mamba model, 41× smaller than Mamba-Spike's official configuration and more accurate under an identical protocol, with weights that fit one neuromorphic chip. It is Pareto-optimal on CIFAR10-DVS, N-Caltech101 and DailyDVS-200 among models with reported parameter counts (Fig. 3), and the best spiking method on DailyDVS-200 (+8.80 pp).

1. Three components beat their natural alternatives at matched parameters (N-Caltech101; Welch vs. five-seed full model, 85.49 ± 0.18; Holm p≤4×10−4p \le 4\times10^{-4}):

ReplacedAccuracyΔ
DS-MPA → sigmoid gate83.92 ± 0.18−1.57 pp, p=2×10−4p = 2\times10^{-4}
MR-S3A → shared concat projection84.00 ± 0.19−1.49 pp, p=3×10−4p = 3\times10^{-4}
Spiking front-end → ConvGRU82.34 ± 0.14−3.15 pp, p=6×10−7p = 6\times10^{-7}

Even with DS-MPA, RepConv and TDM removed, MR-S3A adds +2.96 pp over a stage-3 projection (p=6×10−5p = 6\times10^{-5}), and folds into one convolution at inference. The spiking front-end is 3.15 pp more accurate at near-equal estimated cost on conventional hardware (12.48 vs. 12.97 mJ) and an estimated 2.4× cheaper on neuromorphic hardware.

2. The interface diagnosis holds on a second model. In SpikMamba (Chen et al., 2024) the interface is likewise the sole resolution-dependent part, with a 128× smaller slope. At native sensor resolution (DAVIS346 to 1280×720) it is 45–89% of SpikMamba's parameters and 98.5–99.9% of Mamba-Spike's; ours stays at 0.87 M (App. C.1).

3. The Mamba-Spike comparison is controlled. The 41× and our gains are against its official implementation (https://github.com/ecnu-cross-innovation-lab/mamba-spike, default configuration), trained under our protocol except for model-specific weight decay and loss. We are more accurate on every shared benchmark, and our protocol raises the baseline everywhere (16.4 pp on CIFAR10-DVS). On DVS-Gesture we also exceed its published 97.80% (98.48%; five-seed mean 98.18). Its published figures use a larger, unstated configuration; its CIFAR10-DVS 92.50 would exceed the best published spiking result we know (84.00) by 8.5 pp, so we claim no accuracy advantage there.

4. Statistics. All eight single-run CIFAR10-DVS / N-Caltech101 ablations pass Holm-corrected prediction-interval tests (equal variance stated, App. C.8); every new control has replicated seeds.

5. Accumulate-only execution is verified (DVS-Gesture):

  • Verified computation. Prop. 3 runs each spiking convolution as two accumulate-only kernels, WaW_a on the current and WbW_b on the previous binary input. It reproduces the fused model's predictions on all 264 test clips; NeuroBench classifies all 10 dual-kernel convolutions as accumulate-only (Table 22).
  • SpiNNaker. A trained dual-kernel layer replaying 1.71 M of its own spike events matches PyTorch at r=0.9999998r = 0.9999998 with no lost packets (Table 24).

6. Storage feasibility (Table 4). Mamba-Spike computes less per inference; our advantage is storage, which decides deployability. Our 8-bit weights (1.2 MB) fit one Loihi / Loihi 2 / SpiNNaker 2 chip (Akida at 4 bits) with >90% to spare. Mamba-Spike's (36.3 MB) need 2–3 chips at 128² and ≥119 at HD.

Scoped claims: the parameter diagnosis applies to flatten-then-project bridges; "surrogate-free" refers to TDM's local Jacobian; energy figures are estimates; the competitive claim is accuracy per parameter. --top-k defaults are fixed; new scripts are in the supplement.

We hope the new evidence and scoped claims address the reviewers' concerns, and would be grateful if they would consider revisiting their evaluations in light of the revision. We thank all reviewers again for their insightful reviews. Please let us know if you have any further comments or questions; we look forward to a fruitful discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.