acceptodds
Under review as a conference paper at ICLR 2027

SpikCTU: Compact Adaptive Reasoning for Event-Based Speech Command Recognition

Abstract

Spiking Neural Networks offer event-driven, energy-efficient computation for speech command recognition, yet most existing methods follow fixed computation paths, allocating redundant computation to easy samples while providing limited insight into how evidence evolves during inference. To address these challenges, we introduce SpikCTU, a compact spiking network with adaptive recurrent computation for speech command recognition. Specifically, a hybrid spiking embedding module preserves continuous input information alongside spike features to construct a fixed temporal memory. A shared Continuous Thinking Unit (CTU) repeatedly reads this memory using keys derived from spikes and continuous values, and updates a binary recurrent state with lightweight stage adapters. Furthermore, combining a fixed base prediction with accumulated corrections at each step enables learned aggregation over all steps or confidence-based early exit using only the predictions available so far. Experiments on three speech command recognition benchmarks show that SpikCTU achieves competitive accuracy with only 0.16–0.17M parameters, and the early exit strategy can substantially further reduce computational cost without compromising accuracy. In addition, trajectory and attention analyses reveal that recurrent steps enable error correction and progressively refine predictions. Code is available at https://anonymous.4open.science/r/spikctu-5022/.

Primary area:
applications to neuroscience & cognitive science
Keywords:
spiking neural networks, speech command recognition, adaptive computation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_2rkdOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SpikCTU is a compact hybrid spiking architecture for speech command recognition. It constructs a fixed temporal memory containing spike-derived keys and continuous values, then repeatedly accesses this memory through a shared recurrent unit with binary states and lightweight stage adapters. Intermediate predictions combine fixed base evidence with accumulated corrections, supporting either full-depth aggregation or confidence-based early exit.

Experiments on SHD, SSC, and GSC demonstrate competitive accuracy with approximately 0.16–0.17M parameters. The paper includes component ablations, five-seed results, recurrent-trajectory analyses, temporal occlusion experiments, and operation-based energy estimates.

Critical strengths

Compact architecture with competitive performance. The parameter counts are substantially lower than several strong SSC and GSC baselines. Separating fixed memory construction from shared recurrent computation provides a coherent approach to controlling model size.

Useful evidence for adaptive recurrent computation. At matched average recurrent depth, the selected early-exit policies outperform interpolated truncation baselines by 0.81 and 1.21 percentage points on SSC and GSC. The paper appropriately distinguishes truncating one trained model from independently training models at different depths.

Substantial diagnostic analysis and transparent limitations. The paper examines both corrections and degradations across recurrent steps. Temporal occlusion provides stronger evidence of input relevance than attention visualization alone. The appendix also reports seed variability and clearly separates recurrent-step savings from total proxy-energy savings.

Critical weaknesses

W1. The practical efficiency benefit of adaptive recurrence is limited by the fixed computation. Section 5.4; Appendix D, Tables 11–12.

The reported 59% and 67% reductions concern recurrent evaluations. Under the paper’s own proxy, reducing inference to one recurrent step lowers total energy by only 0.76%, 4.90%, and 1.50% on SHD, SSC, and GSC, respectively. Thus, substantial savings in recurrent depth do not establish substantial savings in total inference cost.

The appendix acknowledges this limitation, but the selected early-exit policies are not accompanied by their own total-operation or energy results. Moreover, cross-model energy comparisons use different accounting conventions, and the proxy excludes memory movement, normalization, nonlinearities, and control overhead. This limits conclusions about deployment efficiency.

The paper should report total costs for the actual selected policies and use consistent accounting for the most relevant comparisons. A measured latency result on an available platform would further clarify whether adaptive execution provides a practical benefit; specialized neuromorphic hardware is not essential.

W2. The experiments do not sufficiently isolate the contribution of the spiking recurrent design. Sections 3 and 5.2; Tables 2 and 7.

The component ablations and independently trained depth sweep establish that several design choices matter within SpikCTU. However, they do not establish whether binary recurrent dynamics provide an advantage over a comparably sized continuous recurrent model using the same encoder, memory, supervision, and stopping rule. The “continuous K/V memory” ablation changes the memory representation rather than testing this alternative.

A base-evidence-only control is also missing, although every intermediate prediction is anchored to that predictor and most computation occurs before recurrence. These controls would help separate the contributions of the encoder, shared parameterization, recurrent refinement, and spiking dynamics.

Confidence-based early exit is already established in SNNs, including SEENN. The paper recognizes this literature and distinguishes recurrent depth from sensory timesteps, but the empirical case for the specific proposed combination remains incomplete.

W3. Important implementation details require clarification for reproducibility. Section 3.1; Appendix A.1.

The GSC configuration specifies an 8-kHz sample rate together with (f_{\max}=14{,}000) Hz. The paper explicitly retains this inherited setting, but 14 kHz exceeds the 4-kHz Nyquist frequency. The effective filterbank behavior—including clipping, empty filters, or any difference between the actual and stated sample rate—needs to be documented because it affects the representation and comparison with other preprocessing pipelines.

There is also an apparent ambiguity between the token-averaging head used to construct base evidence in Section 3.1 and the statement in Appendix A.1 that an auxiliary head averaging token predictions is disabled. These may be different modules, but their distinction and active configuration should be explicit.

Questions for authors

How do the selected early-exit policies perform across seeds in accuracy, total operations, and total proxy energy?

What performance is obtained from the base predictor alone and from a parameter-matched continuous recurrent counterpart under the same training and evaluation protocol?

What is the effective GSC sampling/filterbank configuration, and how does the disabled auxiliary head differ from the base-evidence head?

4 comments

We thank you for the detailed review and for recognizing the compact architecture, adaptive computation, and diagnostic analysis. We are glad to address these questions you raised.

Actual policies and variation across seeds. We evaluated all 15 original paper checkpoints using the Table 3 thresholds unchanged across the five seeds, h = 0.01/0.84/0.91 for SHD/SSC/GSC. New CUDA FP32 evaluations on an NVIDIA A800 preserve original input order and padding. Values are mean ± sample standard deviation.

TaskFull accuracy (%)Exit accuracy (%)Mean exit steps
SHD95.51 ± 0.7194.86 ± 0.901.000 ± 0.000
SSC81.39 ± 0.2281.13 ± 0.193.271 ± 0.037
GSC95.78 ± 0.1395.71 ± 0.181.316 ± 0.010

Arrows below denote full → exit. SOPs/MACs are millions per sample, averaged across seeds.

TaskSOPs (M)MACs (M)Proxy energy (mJ)Energy saving (%)
SHD1.354 → 1.3362.786 → 2.7670.014036 → 0.0139300.756 ± 0.006
SSC1.614 → 1.5113.122 → 3.0280.015814 → 0.0152873.332 ± 0.024
GSC0.619 → 0.5514.895 → 4.8410.023075 → 0.0227651.345 ± 0.007

Exit-energy standard deviations are 0.00001069/0.00001278/0.00000775 mJ. We count fixed encoding and memory work plus each sample's recurrent prefix, verified with online single-sample probes. The counts represent independent stopping with the original padding. Fixed computation dominates total energy, explaining the smaller savings relative to recurrent-step reductions.

The shared proxy uses 0.9 pJ/SOP and 4.6 pJ/MAC and excludes memory traffic, normalization, nonlinearities, and control overhead. The binary and continuous controls below use this same counter, providing a consistent cost comparison.

Separate A800-SXM4-40GB batch-1 measurements give mean model-forward latency of 6.10 → 4.62, 15.39 → 7.08, and 9.00 → 4.83 ms on SHD/SSC/GSC. They cover 236 examples after warm-up and exclude feature extraction, transfer, and external reset. Thus, this implementation shows a larger latency reduction than the arithmetic-energy proxy reduction.

Continuous counterpart and base prediction. Historical controls use the same encoder, fixed hybrid memory, supervision, task training recipe, and exact parameter count. The three recurrent PLIF activations are replaced by scalar PReLUs. We evaluate both versions using the same GPU protocol and stopping thresholds. For seed 312:

TaskParameters, eachBinary full / exit (%)Continuous full / exit (%)Base branch only (%)
SHD164,22696.60 / 95.6395.23 / 94.704.95
SSC174,04281.74 / 81.4181.13 / 81.153.65
GSC169,42895.96 / 95.9995.87 / 95.893.86
TaskFull SOPs (M), binary / continuousFull MACs (M), binary / continuousFull proxy energy (mJ), binary / continuous
SHD1.360 / 1.3242.786 / 2.9220.014041 / 0.014635
SSC1.606 / 1.4273.122 / 3.6780.015807 / 0.018205
GSC0.616 / 0.5304.895 / 5.1730.023073 / 0.024274

The binary design uses less proxy energy with higher SHD/SSC accuracy and similar GSC accuracy in these paired runs. Across five SHD seeds, full accuracy is 95.51 ± 0.71% for binary and 95.14 ± 0.64% for continuous. The replacement changes both output domain and activation persistence, so it evaluates that combined design. The base-only values are branch readouts from the jointly trained model, with recurrent corrections removed at evaluation. This branch receives supervision through the combined predictions and has no separate classification loss.

Relation to SEENN. Confidence stopping is established in SNNs. Here it controls updates after sensory encoding and fixed memory construction. The contribution is the compact memory-reading recurrent architecture and its intermediate outputs. In the submitted paper, adaptive allocation improves accuracy over interpolated truncation of the same trained model by 0.81/1.21 points on SSC/GSC at equal average recurrent depth.

GSC preprocessing and heads. Audio is resampled from 16 to 8 kHz before a 256-point FFT, giving 129 bins over 0–4 kHz. The inherited 14-kHz upper mel setting is applied to the 0–4-kHz spectrum and produces 53 empty filters out of 140. Setting the upper frequency to 4 kHz leaves 7 empty filters because of finite FFT resolution. Separate follow-up training runs with the same recipe reach 95.75% and 95.87%, respectively, one run per setting.

The active base head classifies composed tokens, averages their class probabilities over valid tokens, and takes the logarithm. The disabled auxiliary token head is a separate optional branch. Its removal does not disable base evidence or intermediate recurrent supervision.

Thank you for the detailed response and the additional experiments. The results across five seeds, measured latency, continuous counterpart, and preprocessing clarification substantially address several of my original concerns. I have three remaining points.

  1. Efficiency claims and timing protocol. The new measurements provide useful evidence of reduced batch-1 model-forward latency on the evaluated GPU. I encourage the authors to emphasize this measured benefit while clearly distinguishing it from the modest total proxy-energy savings and reporting the associated accuracy changes. Please clarify the timing protocol, including CUDA synchronization, whether confidence computation and early-exit decisions are included, and how the 236 examples were selected.
  2. An independently trained non-recurrent baseline. The base-branch results show that the jointly trained model relies on recurrent corrections. Since this branch receives no independent classification loss, its low standalone accuracy leaves open how well a non-recurrent classifier could perform when trained directly. A baseline using the same encoder and training budget, with an independently supervised classification head and no recurrent refinement, would help establish the benefit of recurrence.
  3. Stability of the binary-versus-continuous comparison. The parameter-matched continuous counterpart is a useful addition. Paired multi-seed comparisons on SSC and GSC would further clarify the consistency of the relatively small accuracy differences. Please also distinguish the combined effect of binary outputs and persistent neuronal dynamics from the contribution of binarization alone.

I recognize the limited time remaining in the discussion period. Additional experiments would strengthen the evidence, but I would also consider a careful analysis of the existing results, a clear explanation of the remaining limitations, and appropriately scoped claims sufficient grounds to reconsider my current rating upward. My reassessment will reflect the overall contribution and how convincingly these remaining points are addressed.

We sincerely thank you for your follow-up. We tried to address your remaining questions with additional experiments.

1. Efficiency claims and timing protocol.

  • Efficiency claims. As shown in the following table, early exit substantially reduces measured batch-1 model-forward latency, with the largest gains on SSC and GSC, while preserving accuracy close to full aggregation. In the revised version, we will clarify this point.
TaskAccuracy (%)SOPs (M)MACs (M)Proxy energy (mJ)Mean latency (ms), B=1
SHD96.60 → 95.631.360 → 1.3412.786 → 2.7670.0140 → 0.01396.10 → 4.62
SSC81.74 → 81.411.606 → 1.5053.122 → 3.0280.0158 → 0.015315.39 → 7.08
GSC95.96 → 95.990.616 → 0.5484.895 → 4.8410.0231 → 0.02289.00 → 4.83
  • Timing protocol. Latency was measured with synchronized CUDA events and includes confidence computation and the exit decision. We used the first 256 test examples from the data loader, discarded 20 warm-up examples, and used the remaining 236 examples to compute the average latency.

2. An independently trained non-recurrent baseline.

Following your comment, we additionally conducted an experiment to evaluate the effectiveness of recurrent refinement. We trained a non-recurrent classifier with the same encoder, a 300-epoch budget, an independently supervised classification head, and a similar parameter count, as shown in the following table.

TaskSpikCTU accuracy / parametersNon-recurrent accuracy / parameters
SSC81.74% / 0.174 M73.13% / 0.167 M
GSC95.96% / 0.169 M90.54% / 0.167 M

At nearly the same parameter count, SpikCTU improves accuracy by 8.61 points on SSC and 5.42 points on GSC. These results demonstrate the effectiveness of recurrent CTU refinement beyond the encoder and classification head.

3. Paired multi-seed comparison with the continuous counterpart.

We added a multi-seed evaluation with five seeds on the SSC benchmark to compare SpikCTU with a continuous counterpart. The results are provided in the following table and show comparable performance between the two models. The continuous counterpart only replaces the three recurrent PLIF activations with scalar PReLU activations. However, the SNN model has higher energy efficiency.

EvaluationSNNContinuous
SSC accuracy (%)81.39 ± 0.2281.46 ± 0.12
Energy cost (mJ)0.0158070.018205

The multi-seed evaluation on GSC is ongoing.

4. Claims and limitations.

From these results, SpikCTU achieves competitive accuracy and energy efficiency compared with strong existing baselines, while early exit further reduces inference latency. However, despite the relatively small model size of 0.16–0.17 M parameters, the energy efficiency remains constrained by the hybrid spiking embedding module, namely the hSEM module. Consequently, the energy savings brought by early exit are relatively modest. We will clarify these limitations in the revised manuscript.

Once again, we appreciate the reviewer’s helpful feedback.

We have completed the GSC comparison with five paired seeds. The continuous counterpart replaces the three recurrent PLIF activations with scalar PReLU activations. The table reports mean ± standard deviation at full recurrent depth.

Evaluation on GSCSpikCTU (hybrid)Continuous
Accuracy (%)95.78 ± 0.1395.83 ± 0.18
Proxy energy (mJ/sample)0.023075 ± 0.0000080.024283 ± 0.000012

The mean accuracy differs by 0.04 percentage points, while the continuous counterpart uses 5.24% more proxy energy. Thank you for suggesting this comparison.

Reviewer_K91GOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SpikCTU uses a hybrid spiking embedding to preserve continuous input information, together with a fixed memory and a shared Continuous Thinking Unit to update a binary spike state over multiple recurrent steps. The model produces a prediction at every step and can either aggregate all predictions or stop early when its confidence is high enough. Experiments on SHD, SSC, and GSC show that the model uses only about 0.16--0.17 million parameters while maintaining competitive accuracy. Adaptive early exit further reduces the number of recurrent computations.

Critical strengths

The model is very compact and supports sample-specific computation through adaptive early exit.
The prediction-trajectory analysis shows how recurrent steps can correct or change the initial prediction, which makes the model’s behavior easier to interpret.

Critical weaknesses

The method performs below the strongest baselines on SSC and GSC, so its main advantage is model size and computational efficiency rather than absolute accuracy.
The early-exit threshold is calibrated after training instead of being jointly learned under an explicit computation budget, which may limit its transferability across tasks and hardware platforms.

Questions for authors

The method performs below the strongest baselines on SSC and GSC, so its main advantage is model size and computational efficiency rather than absolute accuracy.
The early-exit threshold is calibrated after training instead of being jointly learned under an explicit computation budget, which may limit its transferability across tasks and hardware platforms.

1 comment

We thank you for the careful review and for recognizing the compact model and adaptive computation. We are glad to clarify the accuracy and model size tradeoff and to test threshold transfer directly.

Accuracy and model size. SpikCTU targets competitive recognition with a small parameter budget. The submitted Table 1 gives the following comparisons with its strongest listed baselines.

TaskSpikCTU accuracy / parameters / energyBaseline accuracy / parameters / energyAccuracy gapParameter reduction
SSC, versus SpikCommander81.74% / 0.17 M / 0.0158 mJ83.49% / 2.13 M / 0.0180 mJ1.75 ppabout 12.5×
GSC, versus SpikCommander (SNN)95.96% / 0.17 M / 0.0231 mJ96.92% / 2.13 M / 0.0420 mJ0.96 ppabout 12.5×
GSC, versus KW-MLP (ANN)95.96% / 0.17 M / 0.0231 mJ97.56% / 0.42 M / 0.2070 mJ1.60 ppabout 2.5×

These are published point estimates under the respective source protocols. The energy values follow the 45-nm proxy conventions reported in the submitted Table 1. The main benefit is compactness with competitive accuracy. Adaptive execution provides an additional way to reduce recurrent work.

Calibration and transfer. Training supervises intermediate predictions. A threshold is then selected on SSC/GSC validation data and frozen for testing, allowing the computation level to change without retraining. The 0.1-point tolerance applies to validation accuracy. The stopping policy is calibrated after training.

We directly transferred the paper's three threshold values to each task using its archived checkpoint, without tuning on test labels. Each entry below gives test accuracy (%) / mean recurrent steps in the new GPU evaluation; latency entries use full → selected policy.

Target taskh = 0.01, from SHDh = 0.84, from SSCh = 0.91, from GSCSelected-policy latency (ms), B=1, full → exit
SHD95.63 / 1.0096.51 / 1.0896.51 / 1.126.10 → 4.62
SSC75.22 / 1.0081.41 / 3.2881.64 / 4.0215.39 → 7.08
GSC94.38 / 1.0095.86 / 1.2195.99 / 1.329.00 → 4.83

The SSC/GSC thresholds remain useful across these tasks, whereas applying the very low SHD threshold forces immediate exit and substantially reduces SSC accuracy. Locally, moving the SSC threshold from 0.82 to 0.86 changes accuracy from 81.32% to 81.56% and mean steps from 3.08 to 3.47. This demonstrates a tunable tradeoff and supports calibration on the target task.

The selected-policy latency column in Table 2 shows 54.01% and 46.38% reductions in mean model-forward latency on SSC and GSC, measured at batch size 1 on an A800-SXM4-40GB. This provides a direct execution benefit alongside the accuracy and model-size tradeoff. We will continue improving the current architecture toward higher accuracy while retaining its compact recurrent design and adaptive computation.

Reviewer_3DmeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SpikCTU, a compact spiking neural network for speech command recognition with adaptive recurrent computation. The proposed architecture combines a hybrid spiking embedding module, a shared Continuous Thinking Unit (CTU), and an adaptive prediction mechanism supporting intermediate predictions, learned aggregation, and confidence-based early exit. The method is evaluated on SHD, SSC, and GSC. SpikCTU achieves competitive accuracy with approximately 0.16–0.17M parameters. The authors also analyze recurrent prediction trajectories, state dynamics, and attention distributions to understand how predictions evolve across computation steps.

Critical strengths
  1. SpikCTU uses approximately 0.17 million parameters on SSC and GSC, substantially fewer than several high-performing SNN models. This parameter efficiency is particularly valuable for memory-constrained speech recognition applications.
  2. The experimental analysis in this paper is very comprehensive. Beyond conventional accuracy comparisons and component ablations, the authors analyze prediction corrections and degradations, spiking state activity, attention migration, and temporal occlusion. Such analysis strengthens the empirical contribution of the work.
Critical weaknesses
  1. The proposed framework combines several established techniques, including weight-shared recurrence, low-rank adapters, and confidence-based early exiting. Although the ablation studies demonstrate the usefulness of individual components, comparisons with simpler recurrent SNN architectures under matched parameter and computation budgets are needed to better establish the novelty of the proposed CTU.
  2. Table 3 reports 59% and 67% reductions in recurrent evaluations on SSC and GSC. However, Appendix D shows that even reducing computation to a single recurrent step yields only 4.90% and 1.50% theoretical total energy savings, respectively. Although reducing recurrent evaluations is useful, the corresponding improvements in end-to-end computational cost and energy appear considerably smaller. This raises concerns about the practical efficiency benefits of the proposed adaptive computation mechanism.
Questions for authors
  1. Could the authors compare SpikCTU with a simpler weight-shared recurrent SNN under matched parameter and computation budgets to better demonstrate the specific advantages of the proposed CTU?
  2. Given that recurrent computation contributes only a small fraction of the total estimated energy (Appendix D), could the authors report the total SOPs, MACs, and energy savings of the actual early-exit policies in Table 3?
  3. Table 1 reports 96.60% accuracy on SHD, while the five-seed mean is 95.51 ± 0.71%. Could the authors clarify whether the baseline results represent best runs or averages? Does the claimed performance advantage still hold under a consistent reporting protocol?
3 comments

We thank you for the time spent reviewing our work and for recognizing its compact size and diagnostic analysis. We are glad to address the recurrent baseline, practical costs, and reporting protocol.

Simple shared recurrent control. We completed a control using the same encoder, composed memory, training recipe, seed 312, and recurrent depths of 2/8/4. The shared spiking cell reads mean-pooled memory instead of state-dependent attention, with no stage adapters or learned aggregation. Its active hidden width is chosen to match the parameter count. Both models are evaluated with the same CUDA FP32 pipeline on an NVIDIA A800. Each pair below is SpikCTU / simple shared SNN.

TaskParametersTotal SOPs + MACs (M)Accuracy (%)Gain (pp)
SHD164,226 / 164,2684.146 / 4.00396.60 / 95.19+1.41
SSC174,042 / 174,1234.728 / 4.43381.74 / 79.53+2.21
GSC169,428 / 169,4835.512 / 5.19095.96 / 95.27+0.69

Parameter counts differ by less than 0.05%, and the control uses 3.5–6.2% fewer total operations. At these similar budgets, SpikCTU improves accuracy by 1.41/2.21/0.69 points on SHD/SSC/GSC. This comparison supports the value of the CTU design beyond shared recurrence alone.

Actual early-exit costs. We evaluated the original paper checkpoints using the fixed Table 3 thresholds, 0.01/0.84/0.91. The following new GPU evaluation reports accuracy, operation counts, proxy energy, and latency together. Arrows denote full aggregation → early exit.

TaskAccuracy (%)SOPs (M)MACs (M)Proxy energy (mJ)Energy savingMean latency (ms), B=1
SHD96.60 → 95.631.360 → 1.3412.786 → 2.7670.014041 → 0.0139340.76%6.10 → 4.62
SSC81.74 → 81.411.606 → 1.5053.122 → 3.0280.015807 → 0.0152823.32%15.39 → 7.08
GSC95.96 → 95.990.616 → 0.5484.895 → 4.8410.023073 → 0.0227631.34%9.00 → 4.83

We preserve the original batch padding and count, for each sample, the encoder, memory construction, and recurrent operations up to its exit. The counts represent independent stopping and are verified with online single-sample probes. The proxy uses 0.9 pJ/SOP and 4.6 pJ/MAC, excluding memory traffic, normalization, nonlinearities, and control overhead. Because encoding and memory construction are fixed costs, they dominate total operations and proxy energy. Early exit mainly removes recurrent steps, so its clearest benefit appears as lower batch-1 latency, while the total proxy-energy reduction remains modest. The latency column uses separate A800-SXM4-40GB batch-1 measurements, excluding feature extraction, input transfer, and external reset.

Reporting protocol. The 96.60% SHD result is the best of our five runs, whose mean is 95.51 ± 0.71%. The DCLS-Delays and d-cAdLIF entries in Table 1 are published means. Using the submitted five-seed means, the comparison is:

MethodSHD (%)SSC (%)GSC (%)
SpikCTU95.51 ± 0.7181.39 ± 0.2295.78 ± 0.13
DCLS-Delays95.0780.6995.35
d-cAdLIF94.8580.2395.69

SpikCTU exceeds these two baselines by 0.44/0.66 points on SHD, 0.70/1.16 on SSC, and 0.43/0.09 on GSC, respectively. Thus, the mean-accuracy advantage over these two methods is retained. The broader contribution is competitive recognition with a compact model. Table 1's best-run result and Appendix A's five-seed mean describe different statistics.

Thank you for the author's response. However, these additional explanations still haven't dispelled my doubts about the paper. The reasons are as follows:

  1. The author has provided experimental results for SOPs, MACs, theoretical energy consumption, and GPU inference latency. Nevertheless, the new experiments indicate that the total theoretical energy consumption has only decreased by 0.76%-3.32%, which still has a certain gap compared to the self-adaptive computing benefits highlighted in the paper.
  2. The average accuracy of five experiments for SpikCTU is 95.51±0.71%, which is lower than the accuracy reported by some of the other methods in Table 1. Therefore, the current experimental results are not sufficient to fully prove that SpikCTU has performance advantages over other advanced methods. In conclusion, I recognize the author's efforts in the rebuttal stage. However, considering the overall energy consumption benefits brought by the current methods and the limited performance advantages, I believe the current response is not sufficient to completely address my main concerns. Therefore, I decide to maintain the original score.

We thank you for the follow-up assessment. We agree that we need a more precise statement of the paper’s claims.

1. Energy and adaptive computation. The early exit policy is primarily intended to reduce recurrent computation and inference latency, rather than to provide a large reduction in total energy. As the following table shows, it reduces inference latency by 24.3%, 54.0%, and 46.3% on SHD, SSC, and GSC, respectively. The fixed hybrid spiking embedding module, i.e., hSEM, accounts for most of the total cost, so removing recurrent steps has only a limited effect on total energy. The full SpikCTU model nevertheless has the lowest listed energy on SSC and GSC, respectively. We will revise the manuscript to distinguish the model’s overall energy efficiency from the latency benefit of early exit.

TaskMean latency (ms)
SHD6.10 → 4.62
SSC15.39 → 7.08
GSC9.00 → 4.83

2. Accuracy claim. SpikCTU achieves a competitive performance compared with strong baselines while using only 0.16–0.17 M parameters. Our claim is compact recognition with low total energy and adaptive latency reduction, rather than uniformly highest accuracy across all methods.

Thank you again for helping us clarify the scope of our claims. We will make these distinctions explicit in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.