acceptodds
Withdrawn from ICLR 2027

Bridging Efficiency–Accuracy Gap in Spiking Neural Networks via Operation-Constrained Quantization and Low-Spike Suppression

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spiking neural networks constitute a promising approach to energy-efficient intelligence by leveraging binary, event-driven computation. However, binary spike signals substantially constrain the channel capacity for information delivery, thereby limiting attainable model accuracy. To jointly improve representational capacity and inherit event-driven computation, we propose a novel multi-level spiking neuron termed QuantLIF, consisting of operation-constrained quantization and low-spike suppression. Specifically, we construct a dense quantization codebook whose values can be represented by a bounded number of signed power-of-two terms. Through shift-based re-parameterization, synaptic integration is replaced by several accumulations, without introducing excessive inference delay. To counteract the deterioration in event sparsity caused by multi-level spikes, the low-spike suppression stochastically filters low-contribution spikes during training and deterministically eliminates them during inference. We theoretically analyze QuantLIF from the perspectives of quantization and information theory, respectively. Comprehensive experiments demonstrate that QuantLIF achieves a favorable trade-off among predictive accuracy, firing sparsity, and energy consumption across both static and neuromorphic benchmarks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Spiking Neural Networks, Multi-level Spiking Neuron, Quantization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_aawsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the trade-off between representational capacity and computational cost in multi-level spiking neural networks and proposes QuantLIF. It constrains spike amplitudes to sums of at most D signed power-of-two terms, allowing synaptic multiplication to be implemented through a bounded number of shifts and accumulations. Low-amplitude spikes are randomly suppressed during training and deterministically removed during inference to reduce the number of events. Classification experiments on CIFAR10/100, ImageNet-1K, and CIFAR10-DVS demonstrate promising accuracy, while estimates of synaptic arithmetic energy consumption also indicate some advantages.

Critical strengths
  1. The method is well motivated. Directly connecting the representational capacity of multi-level spikes with an operation budget addresses an important challenge. The appendix also provides theoretical proofs with rigorous reasoning.

  2. The paper is clearly written and well structured. The empirical results are strong, and the ablation studies are comprehensive.

Critical weaknesses
  1. Incomplete baseline comparisons. Table 3 reports only the 77.33% TET result on CIFAR10-DVS, omitting the 83.17% result achieved with data augmentation, despite QuantLIF also using augmentation and TET loss. EventRPG reports 85.55% with Spike-VGG11 and 10 timesteps, exceeding QuantLIF's 84.10%, although its augmentation-based approach is not a directly controlled neuron comparison. Furthermore, PIT is included in Tables 2–3 but omitted from the controlled neuron comparison in Table 1. These omissions weaken the completeness of the experimental comparisons and require clarification.

  2. Limited validation of practical energy efficiency and generalizability. Although QuantLIF demonstrates promising classification performance, its generalizability beyond classification remains untested. More importantly, Appendix F estimates synaptic arithmetic energy while accounting for the actual decomposition terms, but excludes quantization, decoding, communication, and memory-access costs. These additional costs could affect the claimed efficiency advantages. A more comprehensive energy analysis, ideally supported by hardware measurements, would strengthen the conclusions.

  3. Minor presentation and terminology issues.Some figures, particularly Figure 4, contain small text. On page 6, the phrase "same D timesteps during training" is ambiguous because DD denotes the decomposition-term budget rather than simulation timesteps TT. The "(No Title)" reference entry on page 12 should also be corrected.

Questions for authors
  1. Could the authors complete the comparisons across datasets and methods, including the latest relevant approaches where possible? After adding these comparisons, how should the scope of the paper’s SOTA claim be defined?

  2. Could the authors provide comparative results on segmentation or tracking tasks to better demonstrate the effectiveness and generalizability of QuantLIF? Regarding energy estimation, please also explain the energy consumed by processing steps outside the neural network.

  3. Why does a reference contain “(No Title)”?

Reviewer_2ZnaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes QuantLIF, a multi-level spiking neuron designed to improve the accuracy–efficiency trade-off in spiking neural networks (SNNs). The proposed method consists of two components. First, operation-constrained quantization constructs a quantization codebook whose spike amplitudes can be represented by at most DD signed power-of-two terms. This increases the representational capacity of spiking neurons while enabling multiplier-free synaptic computation through shift-based reparameterization. Second, low-spike suppression randomly removes low-amplitude spikes during training and deterministically suppresses them during inference, aiming to preserve event-driven sparsity while minimizing accuracy degradation. The authors further provide theoretical analyses of the number of quantization levels, quantization precision, and the information-theoretic motivation for spike suppression. Experiments on CIFAR10, CIFAR100, ImageNet-1K, and CIFAR10-DVS demonstrate competitive classification accuracy and estimated energy efficiency compared with existing spiking neuron models.

Critical strengths

1.The paper achieves competitive performance compared with conventional SNNs and demonstrates the effectiveness of the proposed method across multiple datasets and network architectures. 2.The paper provides theoretical analyses of the proposed quantization scheme, including the number of valid quantization levels and the maximum quantization interval under a constrained operation budget.

Critical weaknesses

1.Multi-level spike representations and power-of-two quantization have already been explored in previous studies, such as ShiftLIF. Therefore, the methodological novelty of QuantLIF appears somewhat limited. 2.Most experiments on static datasets are conducted with only one timestep (T=1), where temporal dynamics are absent. Under this setting, the proposed method closely resembles activation quantization in conventional neural networks. 3.The additional benefit of stochastic low-spike suppression appears limited when an appropriate suppression threshold is selected. For example, Table 4 shows that at (V_{\mathrm{th}}=2^{-4}), stochastic suppression improves accuracy by only 0.05 percentage points compared with training without stochastic suppression, while resulting in slightly higher estimated energy consumption.

Questions for authors

1.Why is PIT excluded from the energy consumption comparison in Table 1?

Reviewer_2QVoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper aims to address the limitations of channel capacity in spike signal delivery by proposing a multi-level spiking neuron model that employs operation-constrained quantization and low-spike suppression; so-called QuantLIF. Its key idea is to construct a quantization codebook whose values can be represented by a bounded number of signed power-of-two terms. To address the deterioration caused by multi-level spikes, the low-spike suppression randomly filters low-contribution spikes during training, as well as deterministically removes them during inference. Experiments cover multiple datasets, including the CIFAR-10, CIFAR-100, ImageNet-1K, and CIFAR10-DVS datasets. The reported results are promising, as QuantLIF leads to competitive accuracy with fewer timesteps than several existing SNN methods. However, several questions about novelty, comparison study, generalizability, and energy efficiency still remain.

Critical strengths

• The proposed QuantLIF leads to competitive accuracy results across different types of datasets (i.e., static and event-based), which shows promising performance across diverse workloads.

• The proposed QuantLIF jointly considers spike representation and computational cost to optimize both performance and efficiency through a simple quantization codebook and suppression mechanism.

Critical weaknesses

• The concept of multi-level spiking neurons is not new, and the difference between the proposed multi-level spiking neurons and the existing multi-level spiking models is the low-amplitude firing suppression, which is an incremental contribution.

• Comparisons with existing multi-level spiking neurons (i.e., Ternary Spike, INT-LIF, and MLF) are only performed considering the CIFAR10 dataset in Table 1, while other comparisons (in Table 2 and Table 3) only compare against the Ternary Spike (from Guo et al., 2024) as the representative multi-level spiking neuron. This is important to show improvements relative to prior multi-level neurons.

• The overhead costs of QuantLIF in terms of processing latency/throughput, memory, and power/energy consumption during the training and inference phases are not comprehensively discussed.

Questions for authors

• Can the authors clarify why stochastic removal is only used in the training stage and not in the inference stage? What happens if the stochastic removal approach is used during the inference stage?

• It is suggested to clarify the applicability of QuantLIF for different SNN model developments, such as ANN-to-SNN conversion and direct model development in SNN domain. What differences are required to apply QuantLIF to different SNN model development approaches?

• It is recommended to discuss the applicability of QuantLIF for different types of neural coding schemes, such as rate, time-to-first spike, rank-order, etc.

• Comparisons with existing multi-level spiking neurons (i.e., Ternary Spike, INT-LIF, and MLF) are only performed considering the CIFAR10 dataset (in Table 1), while other comparisons (Table 2 and Table 3) only compare against the Ternary Spike (from Guo et al., 2024) as the representative multi-level spiking neuron. Therefore, it is suggested to perform comparisons to all multi-level spiking neurons (i.e., Ternary Spike, INT-LIF, and MLF) for all datasets to further show the benefits of QuantLIF.

• Can the authors clarify whether the results for the comparison partners come from the existing papers or from running new experiments to obtain new results with the same training mechanism and settings as QuantLIF?

• Can the authors clarify the overhead costs of QuantLIF in terms of processing latency/throughput, memory, and power/energy consumption during the training and inference phases?

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

Related papers

Discussion (0)

Sign in to comment.