acceptodds
Under review as a conference paper at ICLR 2027

Runtime-Adaptive Temporal Integration Improves Sparsity and Accuracy in Spiking Neural Networks

Abstract

Spiking neural networks inherently operate in an event-driven manner, yet their neurons typically rely on a fixed temporal integration scale during inference, regardless of changes in input conditions or their own activity states. Although some neuron models learn task-dependent membrane time constants during training, these time constants are typically kept fixed at inference, limiting the ability of SNNs to adapt their temporal integration to runtime conditions. Motivated by this observation, we study runtime-adaptive temporal integration as a computational perspective in which each neuron can adjust its temporal integration online according to local input conditions and recent firing activity. To instantiate this perspective, we develop RATI-LIF, a lightweight LIF-based neuron that provides a practical realization of runtime-adaptive temporal integration. RATI-LIF learns a task-dependent reference integration timescale through layer-wise base time constants and further modulates this timescale at runtime through neuron-level input-driven and output-spike-driven dynamics. Experimental results show that RATI-LIF substantially reduces neuronal firing rates and estimated energy consumption while improving classification accuracy. Taken together, these results suggest that runtime-adaptive temporal integration offers a promising direction for building sparser and more effective spiking neural networks, with RATI-LIF providing a lightweight and practical realization of this idea.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Networks (SNNs), Brain-Inspired Computing

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

In this work, the authors tackle runtime adaptation of temporal integration in spiking neural networks. Their design centers on a LIF neuron whose learned reference timescale is locally modulated by both input currents and recent firing activity. The paper also provides a theoretical analysis detailing when this formulation reduces spike counts under matched inputs. Evaluated across neuromorphic vision, physiological data, and event-based speech benchmarks, the proposed neuron improves both accuracy and firing sparsity while offering clear potential for energy savings.

Critical strengths

Careful theoretical treatment of firing sparsity under runtime modulation. The authors provide a thorough mathematical analysis of how dynamic temporal modulation shapes spike sparsity, working out an exact finite-horizon firing-rate balance alongside sufficient conditions for firing suppression. I especially like that the paper clearly separates what is guaranteed at the single matched-neuron level from what actually carries over to trained multi-layer networks. (As a side note, the AI Use Statement mentions using AI tools to help develop and verify the mathematical proofs, though it leaves the exact scope of that assistance somewhat vague.)

Critical weaknesses
  1. My main concern is that the proposed neuron appears to offer an incremental modification of existing neuronal dynamics. In RATI-LIF, recurrent input and output-spike traces are used to modulate τ(t)\tau(t) and the membrane retention factor, which is really just another way of doing history-dependent gating of membrane memory. We already have several closely related mechanisms in the SNN literature, including compartment coupling in LM-H [1] and TC-LIF [2], dynamic feedback modulation in STC-LIF [3], and state-dependent auxiliary dynamics in CLIF [4]. I realize these models are not identical on paper, but the authors do not adequately explain what their specific formulation achieves that these prior neurons cannot. Highlighting firing sparsity is a nice angle, though it does not by itself make the neuron design novel.
  2. The practical advantages of RATI-LIF are not sufficiently established. First, Appendix A.6 already shows quite a bit of sensitivity to λ\lambda and η\eta on a single task, and there is no ablation on the clipping bounds in Eq. (9). In fact, I could not even find the values of those bounds in the experimental setup, which makes me wonder how much the model relies on hand-tuned clipping to stay stable. Second, RATI-LIF carries extra state variables and has to recompute its decay factor at every timestep. Without seeing actual training time and inference latency numbers measured on the same GPU against existing neuron models, I am skeptical of the claim that this is a lightweight practical realization.

[1] Hao, Z., et al. A progressive training framework for spiking neural networks with learnable multi-hierarchical model. ICLR 2024.

[2] Zhang, S., et al. Tc-lif: A two-compartment spiking neuron model for long-term sequential modelling. AAAI 2024.

[3] Lihao Wang, Zhaofei Yu. Autaptic synaptic circuit enhances spatio-temporal predictive learning of spiking neural networks. ICML 2024.

[4] Huang, Y. et al. CLIF: Complementary Leaky Integrate-and-Fire Neuron for Spiking Neural Network. ICML 2024.

Questions for authors
  1. What actually sets RATI-LIF apart from the mechanisms already used in LM-H, TC-LIF, STC-LIF, and CLIF? A side-by-side comparison would really help clarify whether the core contribution comes from a genuinely different class of neuronal dynamics, a cleaner implementation, or simply a better empirical trade-off between accuracy and firing sparsity.
  2. What exact values did you use for τmin⁡\tau_{\min} and τmax⁡\tau_{\max}, and how were they chosen? Also, how often does τ(t)\tau(t) actually hit or saturate at either bound during training and inference? Please show whether the gains in accuracy and spike sparsity still hold up when you vary the modulation coefficients and clipping bounds across more than just one dataset.
  3. Could you provide the per-iteration training time and inference latency measured on the same hardware against LIF, PLIF, and representative adaptive neuron baselines? To keep the comparison fair, all runs should share the exact same backbone, number of timesteps, and level of kernel optimization. Reporting peak GPU memory usage would also be very helpful for quantifying the overhead of tracking those extra neuronal states.
2 comments

We sincerely thank the reviewer for the insightful comments on the methodological distinction, clipping behavior, and practical efficiency of RATI-LIF.

Q1. Distinction from existing adaptive neuron models

Our primary motivation is to investigate how runtime-adaptive temporal integration affects firing sparsity, rather than to introduce a fundamentally new class of neuronal dynamics. RATI-LIF serves as a lightweight instantiation of this research perspective, with theoretical characterization provided in Sec. 2.3 and Appendix A.2.

Mechanistically, LM-H and TC-LIF employ coupled membrane potentials, STC-LIF modulates membrane retention and input through learnable feedback connections, and CLIF regulates post-spike reset through a complementary potential. RATI-LIF instead directly modulates the membrane time constant through local input and output-spiking histories, providing a simple formulation for investigating the relationship between runtime-adaptive integration and firing sparsity.

Q2. Clipping bounds and hyperparameter sensitivity

As described in Appendix A.3, the clipping bounds are introduced primarily to ensure numerical stability and maintain a physically meaningful membrane decay factor. Specifically, we set τmin⁡=0.01\tau_{\min}=0.01 and τmax⁡=10.1\tau_{\max}=10.1 across all experiments. The lower bound ensures that the effective membrane time constant remains strictly positive, preventing zero or negative values, while the upper bound prevents excessively large time constants that would cause the membrane decay factor ρm(t)=exp⁡(−Δt/τ(t))\rho_m(t)=\exp(-\Delta t/\tau(t)) to approach 1. Together, these bounds constrain the decay factor within (0,1), maintaining stable membrane integration. The theoretical effects of clipping are explicitly considered in Appendix A.2.5 (Eqs. 84–89).

To further investigate the practical impact of clipping, we additionally measured the frequency at which the adaptive membrane time constant reaches the lower and upper bounds during training and inference across different datasets. The results are summarized in Table R6.

Table R6: Frequency (%) of the adaptive membrane time constant reaching the clipping bounds during training and inference.

DatasetTraining Lower (%)Training Upper (%)Inference Lower (%)Inference Upper (%)
DVS-Gesture13.88013.740
CIFAR10-DVS0.504400.42550
SHD0.186000.23290

As shown in Table R6, no upper-bound saturation is observed in the evaluated datasets. Lower-bound saturation is below 1% on CIFAR10-DVS and SHD, while it occurs in approximately 14% of updates on DVS-Gesture during both training and inference.

The absence of upper-bound saturation is consistent with the fact that the learnable reference time constant τ0\tau_0 is initialized below the upper bound and remains below it throughout training, while the observed runtime modulation does not push the effective time constant beyond τmax⁡\tau_{\max}. These results indicate that the adaptive time constant operates within the unclipped range for the majority of neuronal updates, although the extent of lower-bound saturation varies across datasets.

Regarding the modulation coefficients, the values of λ\lambda and η\eta used across different datasets are reported in Table 5 of the manuscript, while their sensitivity is analyzed on DEAP in Appendix A.6 (Table 7). The results demonstrate that RATI-LIF maintains competitive performance across multiple coefficient configurations within a reasonable range, although certain extreme settings can lead to performance degradation. This indicates that the model does not depend exclusively on a single hyperparameter configuration, while appropriate parameter selection remains important.

Due to the limited time available during the rebuttal period, a comprehensive sensitivity analysis of the modulation coefficients and clipping bounds across all datasets and architectures could not be completed. Nevertheless, the existing hyperparameter ablation and the additional clipping-frequency statistics provide further evidence that the proposed runtime adaptation mechanism operates predominantly within the unclipped range in the evaluated settings. We will incorporate these clarifications and additional results into the revised manuscript.

Q3. Training and inference efficiency

To assess the practical computational overhead of RATI-LIF, we have conducted additional experiments comparing standard LIF and RATI-LIF on the SHD dataset under controlled experimental conditions. Both models employ the same two-hidden-layer recurrent fully connected network with 128 neurons per hidden layer, trained for 150 epochs with a batch size of 128 and 100 simulation timesteps. All experiments are conducted on the same NVIDIA RTX PRO 6000 GPU using identical training configurations and the same procedure for measuring peak GPU memory usage.

Table R7: Comparison of total training time, test time, and peak GPU memory usage on SHD.

NeuronTotal Training Time (s)Test Time (s)Peak GPU Memory (MB)
LIF3970.410.9131026
RATI-LIF4687.231.1971126

As shown in Table R7, RATI-LIF increases total training time, test time, and peak GPU memory usage compared with standard LIF. This additional overhead primarily arises from maintaining the two recursive modulation states and dynamically updating the membrane time constant. These results suggest that RATI-LIF can realize runtime-adaptive temporal integration with relatively limited computational and memory overhead under the evaluated configuration.

Due to the limited time available during the rebuttal period, we have not completed per-iteration training-time measurements, finer-grained inference latency measurements, or comparisons with PLIF and other adaptive neuron baselines. We therefore provide the controlled SHD comparison as an initial assessment of the practical overhead introduced by RATI-LIF. We will incorporate these results and clarify the scope of the efficiency evaluation in the revised manuscript.

We thank the reviewer again for the valuable suggestions. We will incorporate the additional results and clarifications into the revised manuscript and explicitly acknowledge the remaining limitations of our evaluation.

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

This paper proposes Runtime-Adaptive Temporal Integration LIF (RATI-LIF), a spiking neuron model that dynamically adjusts its membrane time constant based on its input and output activities. The authors also provide a theoretical analysis of the conditions under which adaptive temporal integration can reduce firing rates. Experiments across neuromorphic vision, affective states, and speech datasets using multiple SNN architectures show improved classification accuracy in most settings and reduced firing rates compared to Parametric LIF (PLIF) .

Critical strengths
  1. Extensive evaluations across diverse tasks and network architectures. The comparisons with existing methods are clear.
  2. The ablation studies justify the design choices of RATI-LIF and clarify the contribution of each component.
  3. Clear and easy-to-follow presentation.
Critical weaknesses
  1. Maintaining two additional neuronal states τinput\tau_{\text{input}} and τoutput\tau_{\text{output}} may introduce non-negligible computational and memory cost during training, which is not evaluated.
  2. The energy analysis in Section 3.2 lacks detailed accounting for the additional state updates and memory accesses. These operations may incur non-negligible energy costs, potentially offsetting the benefits of reduced firing activity.
  3. The sparsity improvements over standard LIF are inconsistent. While RATI-LIF achieves a 72.6% firing-rate reduction on DEAP, it exhibits higher firing rates than LIF in some settings in Table 1. This raises concerns about the generality of the claimed sparsity benefits.
Questions for authors
  1. (Weakness 1) Please provide detailed comparison of training speed and memory cost.
  2. (Weakness 2) Please clarify whether the energy model accounts for additional state updates and memory accesses.
  3. (Weakness 3) Consider extending the energy estimation beyond DEAP, particularly to settings where RATI-LIF exhibits higher firing rates than LIF.
  4. How does RATI-LIF perform on static image recognition tasks like CIFAR-100?
  5. Minor: Equation (6) expresses τnl(t)\tau_n^l(t) as a function of the current input and previous spike, whereas Equations (7)–(8) introduce dependencies on historical activities. Consider clarifying this stateful dependence.
  6. Minor: in Table 1, "Spikingformer" is used to describe the architectures of Spikformer and QKFormer. However, Spikingformer refers to a distinct architecture [1]. "Spiking Transformer" would be more appropriate here.

[1] Zhou et. al., Spikingformer: A Key Foundation Model for Spiking Neural Networks, AAAI 2026

4 comments

We sincerely thank the reviewer for the thoughtful comments and suggestions, particularly regarding computational efficiency, energy estimation, and the applicability of RATI-LIF.

Q1. Training time and memory overhead

To evaluate the computational overhead introduced by RATI-LIF, we have conducted an additional comparison between standard LIF and RATI-LIF on the SHD dataset. Both models employ a two-hidden-layer recurrent fully connected network with 128 neurons per hidden layer, trained for 150 epochs with a batch size of 128 and 100 simulation timesteps. All experiments are conducted on the same NVIDIA RTX PRO 6000 GPU under identical training configurations, and peak GPU memory usage is measured using the same procedure for both models.

Table R2: Comparison of training time and peak GPU memory usage on SHD.

NeuronTotal Training Time (s)Peak GPU Memory (MB)
LIF3970.411026
RATI-LIF4687.231126

As shown in Table R2, RATI-LIF incurs a modest increase in training time and peak GPU memory usage compared with standard LIF. This additional overhead primarily arises from maintaining the two recursive modulation states and dynamically updating the membrane time constant. Overall, these results suggest that RATI-LIF achieves runtime-adaptive temporal integration with relatively limited computational and memory overhead.

Due to the limited time available during the rebuttal period, conducting comprehensive training-time and memory profiling across all datasets and architectures is not feasible. We therefore focus on SHD as a representative temporal benchmark to provide a controlled assessment of the computational overhead associated with runtime-adaptive temporal integration. We will incorporate these additional experimental results and the corresponding discussion into the revised manuscript.

Q2. Scope of the energy estimation

Following prior SNN studies, our energy analysis focuses on operation-level estimation rather than hardware-specific energy consumption (Sec. 3.2). The computational costs of the additional state updates are included in the operation-level formulation described below (Q3), whereas memory-access energy is not explicitly accounted for.

The two auxiliary states in RATI-LIF are maintained locally at each neuron, and their memory-access costs depend on the underlying hardware architecture and implementation. We will clarify the scope and limitations of this estimation in the revised manuscript.

Q3. Extended energy analysis on DEAP and SHD

We agree that extending the energy analysis beyond DEAP would provide a more comprehensive evaluation of RATI-LIF's energy efficiency.

As clarified in Sec. 3.1, PLIF provides a more direct static reference for isolating the effect of runtime adaptation. Compared with PLIF, RATI-LIF consistently reduces firing rates across all nine configurations in Table 1 of the manuscript. Nevertheless, we acknowledge that RATI-LIF does not universally reduce firing rates or energy consumption relative to standard LIF.

To further evaluate energy consumption, we adopt the commonly used operation-level energy estimation approach in SNN research, which estimates computational energy based on the number of accumulation (AC) and multiply-accumulate (MAC) operations. The corresponding computational cost formulations for LIF, PLIF, and RATI-LIF are summarized in Table R3.

Table R3: Computational cost formulations for LIF, PLIF, and RATI-LIF.

Neuron ModelDynamicsTotal Cost
LIF/PLIFIt=Wm,nXtmI_t=W^{m,n}X_t^m; Ut=βUt−1+It−vthSt−1nU_t=\beta U_{t-1}+I_t-v_{\mathrm{th}}S_{t-1}^nmnFinrEAC+nFoutrEAC+nEMACmnF_{\mathrm{in}}^rE_{\mathrm{AC}}+nF_{\mathrm{out}}^rE_{\mathrm{AC}}+nE_{\mathrm{MAC}}
LIF/PLIF (Rec)It=Wm,nXtm+Wn,nSt−1nI_t=W^{m,n}X_t^m+W^{n,n}S_{t-1}^n; Ut=βUt−1+It−vthSt−1nU_t=\beta U_{t-1}+I_t-v_{\mathrm{th}}S_{t-1}^nmnFinrEAC+(n2+n)FoutrEAC+nEMACmnF_{\mathrm{in}}^rE_{\mathrm{AC}}+(n^2+n)F_{\mathrm{out}}^rE_{\mathrm{AC}}+nE_{\mathrm{MAC}}
RATI-LIFEqs. (7–11) in the manuscriptmnFinrEAC+(n2+2n)FoutrEAC+4nEMACmnF_{\mathrm{in}}^rE_{\mathrm{AC}}+(n^2+2n)F_{\mathrm{out}}^rE_{\mathrm{AC}}+4nE_{\mathrm{MAC}}

While this approach provides a consistent basis for estimating computational energy, directly applying these simplified formulations to complex architectures such as VGG and Spiking Transformers may not accurately capture their diverse computational operations without additional module-level analysis. Therefore, we focus our extended energy evaluation on fully connected networks, where computational operations can be more explicitly characterized.

Specifically, we measure the average firing rate (FR), input firing rate (FR_in), and output firing rate (FR_out) for LIF, PLIF, and RATI-LIF, and estimate their computational energy consumption using the AC/MAC-based formulations summarized above. For DEAP, we evaluate the 512-to-3 fully connected layer, with input and output dimensions m = 512 and n = 3, respectively. For SHD, we evaluate the second hidden layer of a two-layer recurrent fully connected network, with m = n = 128. The results are summarized in Table R4.

Table R4: Comparison of firing rates and estimated computational energy on DEAP and SHD.

DatasetNeuronFR (%)FR_in (%)FR_out (%)Estimated Energy (pJ/step)
DEAPLIF43.0743.1627.67611.19
PLIF35.4935.5327.91505.72
RATI-LIF22.9622.8934.40373.49
SHDLIF15.1813.2117.155085.32
PLIF14.219.7018.724801.06
RATI-LIF14.0410.1617.936538.55

As shown in Table R4, RATI-LIF achieves the lowest average firing rate and estimated computational energy on DEAP, demonstrating that runtime-adaptive temporal integration can effectively promote firing sparsity while reducing computational energy. On SHD, RATI-LIF also achieves the lowest average firing rate among the three neuron models, although this reduction does not translate into lower estimated energy consumption. This is primarily because the additional computations required for runtime modulation can offset the energy savings from reduced spike activity.

These results further indicate that the energy efficiency of RATI-LIF depends on the balance between firing sparsity and neuronal computation overhead, rather than being universally guaranteed across all network configurations.

We will incorporate these additional results and clarify the scope of the operation-level energy estimation in the revised manuscript.

Q4. Static image classification on CIFAR-100

To further evaluate the applicability of RATI-LIF to static image recognition, we have conducted additional experiments on the CIFAR-100 dataset. Specifically, we adopt three representative SNN frameworks previously employed in our neuromorphic vision experiments, namely SLTT (ResNet-18), Spikformer, and QKFormer, and extend their evaluation to static image classification. We compare the classification accuracy of standard LIF and our proposed RATI-LIF under each framework.

Table R5: Classification accuracy (%) on CIFAR-100.

MethodArchitectureLIFRATI-LIF
SLTTResNet-1874.3174.90
SpikformerSpiking Transformer77.6378.21
QKFormerSpiking Transformer80.5181.25

As shown in Table R5, RATI-LIF consistently improves classification accuracy over standard LIF across all three frameworks, with improvements ranging from 0.58 to 0.74 percentage points. These results suggest that the benefits of runtime-adaptive temporal integration are not limited to event-based temporal data but also extend to conventional static image recognition tasks. This further demonstrates the applicability of RATI-LIF across different data modalities and network architectures.

We will incorporate these additional CIFAR-100 results into the revised manuscript.

Q5. Stateful dependence in Eq. (6)

We would like to clarify that Eq. (6) is intended as a conceptual formulation to illustrate that the membrane time constant is dynamically modulated by local input and output-spiking activities at each timestep, rather than to explicitly specify its complete state dependence. Eqs. (7)–(8) provide the concrete implementation, where two recursive modulation states incorporate historical input currents and output spikes through exponentially decaying updates. Thus, historical activities influence the current time constant implicitly through these states, without requiring explicit storage of the full activity histories.

To avoid potential ambiguity, we will add the following clarification between Eq. (6) and Eqs. (7)–(8) in the revised manuscript:

“Eq. (6) provides a conceptual description of runtime modulation by local input and output activities, while the concrete implementation in Eqs. (7)–(8) incorporates their historical dependence through recursively updated modulation states.”

Q6. Terminology in Table 1

We agree that “Spiking Transformer” is a more appropriate term to describe the architectures of Spikformer and QKFormer, as “Spikingformer” refers to a distinct architecture. We will correct this terminology in Table 1 of the revised manuscript to avoid confusion.

We thank the reviewer again for the constructive feedback, which has helped us improve the experimental evaluation and clarify the scope and limitations of our work.

Thanks for your additional efforts and detailed response. My questions on training overhead and image classification accuracies have been addressed. However, my concerns on the energy consumption estimation remains. As stated in your response to Q3, the firing rate of RATI-LIF is close to that of LIF or PLIF, and the estimated operational energy cost is even higher. This indicates that the energy efficiency of RATI-LIF claimed in the original submission cannot generalize to a broader range of scenarios. Thus, I will keep my original rating.

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

This paper proposes RATI-LIF, which modulates a learnable, layer-wise reference time constant at runtime using current input and recent output spikes. The model is evaluated on vision, EEG, and event-based speech tasks, with analyses of accuracy, firing sparsity, and estimated energy. The central claim is that runtime-adaptive temporal integration can reduce firing activity while maintaining task performance.

Critical strengths
  • The problem is clearly stated: a fixed membrane time constant cannot respond to local input or neuronal activity during inference, motivating runtime adaptation.
  • The method is lightweight and explicit. It uses local recurrent states rather than an additional control network, and comparisons with fixed and learnable-but-static time constants help isolate runtime adaptation.
  • The ablations examine both the reference timescale and runtime adaptation, including separate input-side and output-side variants on DEAP.
Critical weaknesses
Questions for authors
1 comment

We sincerely thank the reviewer for the constructive comments and valuable suggestions. We appreciate the opportunity to clarify our contribution and provide additional experimental evidence.

Q1. Relation to prior adaptive time-constant models

Our central contribution is not the introduction of adaptive membrane time constants, but the systematic investigation of their relationship with firing sparsity. RATI-LIF serves as a concrete instantiation of this research perspective, supported by theoretical characterization (Sec. 2.3 and Appendix A.2) and extensive empirical evidence.

Fang et al. (2021) learn time constants that remain fixed during inference, while Hasani et al. (2021) and Zhang et al. (2025) explore runtime adaptation primarily for temporal information processing. Our work instead focuses on how and under what conditions runtime adaptation promotes firing sparsity. Empirical comparisons with PLIF are provided in Table 1 of the manuscript. Since the original LTC is a non-spiking model, a direct comparison in terms of firing sparsity is not applicable.

Q2. Motivation and effects of runtime modulation

As motivated in Appendix A.1 (Eqs. 21–26), increased membrane conductance shortens the effective integration timescale. Our subtractive formulation provides a simple computational abstraction of this behavior without explicitly simulating conductance dynamics. The two recursive states capture input and output-spiking histories, with clipping ensuring numerical stability.

According to Eqs. (7–10), positive inputs shorten the integration timescale and strengthen membrane leakage, whereas negative inputs have the opposite modulatory effect. Recent output spikes provide additional negative feedback on the integration timescale. Appendix A.2.5 (Eqs. 81–89) explicitly characterizes these modulation effects, including clipping, while Appendix A.2.6–A.2.7 establish the sparsity guarantee under nonnegative inputs and discuss its limitations under signed inputs.

Q3. Additional CIFAR10-DVS ablation studies

We have conducted additional ablation studies on CIFAR10-DVS using two representative architectures, VGG and Spikformer. These experiments examine the individual and combined contributions of the learnable reference time constant and runtime-adaptive modulation.

Table R1: Ablation studies on CIFAR10-DVS using VGG and Spikformer architectures.

ConfigurationVGG Acc. (%)Spikformer Acc. (%)
Fixed73.778.5
Learnable75.479.4
Dynamic75.279.3
+input modulation74.778.7
+output modulation75.479.8
Full RATI-LIF76.980.2

As shown in Table R1, both the learnable reference time constant and runtime adaptation independently improve classification accuracy over the fixed-time-constant baseline across the two architectures. Furthermore, combining input-driven and output-spike-driven modulation with the learnable reference time constant consistently achieves the best performance. These results extend our existing ablation studies and provide further evidence that the proposed components contribute to performance improvements across different network architectures.

We thank the reviewer again for the helpful feedback. We will incorporate the above clarifications and additional ablation results into 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.