acceptodds
Under review as a conference paper at ICLR 2027

SpikeLite: Lightweight Spiking Neural Networks for Time-Series Forecasting

Abstract

Spiking neural networks (SNNs) offer an energy-efficient paradigm for time-series forecasting through spike-driven computation. However, recent SNN forecasters often pursue higher accuracy through increasingly complex attention mechanisms, or specialized neuronal dynamics, weakening the lightweight motivation of SNNs. We introduce SpikeLite, a spiking forecasting framework built around two modules: a Frequency-Selective Spiking Encoder (FSSE) for frequency-sensitive temporal encoding and a Sparse Spiking Channel Attention (SSCA) module for selective cross-channel interaction. FSSE exploits the low-pass filtering behavior of LIF dynamics to reorganize each input sequence into frequency-sensitive components while collectively preserving the input at the decomposition stage. SSCA then learns a binary mask from encoded channel representations and uses it to selectively exchange information within spike-driven self-attention, retaining informative cross-channel interactions while suppressing redundant ones. When explicit channel interaction is unnecessary, SpikeLite uses the lighter FSSE-only channel-independent path. Experiments under the SeqSNN and SpikF protocols cover four standard multivariate and eight long-term forecasting benchmarks. SpikeLite achieves the best aggregate performance under both protocols, with an average of 0.790 and RSE of 0.440, and lowest average MSE/MAE of 0.343/0.345 in long-term forecasting. Moreover, evaluation on the ECL dataset shows that SpikeLite achieves the lowest reported energy consumption, further demonstrating its potential for energy-efficient time-series forecasting.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Networks, Time-Series Forecasting, Frequency-Selective Encoding, Energy Efficiency

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_rK2oOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

SpikeLite combines two components for spiking time-series forecasting. The Frequency-Selective Spiking Encoder (FSSE) applies parallel LIF branches with different learnable decay factors to each variable. Differences between adjacent branch responses and a final residual sum back to the original input before projection, giving a frequency-sensitive decomposition. Sparse Spiking Channel Attention (SSCA) then derives a sample-dependent binary channel mask from FFT descriptors of the learned representations and applies it to spike-driven attention; the mask is trained with a straight-through estimator. The method can omit SSCA for a lighter channel-independent path.

The evaluation follows the SeqSNN protocol on four standard multivariate benchmarks and the SpikF protocol on eight long-term benchmarks. The reported aggregate results are R2=0.790R^2=0.790 and RSE =0.440=0.440 for the former, and MSE =0.343=0.343 and MAE =0.345=0.345 for the latter. On ECL at horizon 720, the estimated encoder energy is 95.30 µJ per sample, compared with 117.66 µJ for SpikF, using a 45-nm operation-cost model. The paper includes three-seed stability measurements, ablations on Traffic and Weather, a frequency sweep, and a detailed operation-count appendix.

Critical strengths

FSSE has a clear signal-processing motivation. Linearized LIF dynamics are low-pass filters whose response changes with the decay factor, and the successive differences plus residual form an exact reconstruction before the learned projection. The paper verifies distinct responses with a sinusoidal sweep. The nonlinear spike threshold means these are not clean frequency bands, but the proposed representation is still more principled than adding an arbitrary stack of spiking layers.

The results cover two established forecasting protocols and include ANN models as well as SNN baselines. The ablation is informative on Traffic: removing FSSE raises average MSE from 0.479 to 0.489, while removing SSCA raises it to 0.602; Weather shows the same direction, with smaller differences. The three-seed tables show small within-method variance. On ECL the model also slightly improves MSE@720 over SpikF, 0.218 versus 0.219, while reducing estimated encoder energy by about 19%. Taken together, these are credible reasons to consider the method, particularly where cross-channel structure matters.

Critical weaknesses
  • The average MSE advantage is thin: 0.343 versus 0.345 for iTransformer and 0.346 for SpikF, while Weather is tied with SpikF at 0.245. The method also loses to some alternatives on individual ETT and Exchange settings, including a weak Exchange 720 result discussed in the appendix. Since most competitor numbers are imported from earlier papers and only the new method has a three-seed stability table here, the paper does not establish that the 0.002–0.003 aggregate gap is robust to matched tuning and implementation. The large SSCA ablation on Traffic demonstrates that channel interaction helps there, but it does not show whether this particular sparse mechanism is better than a comparably sized dense or simpler channel mixer.

  • The efficiency result is a modeled estimate on one dataset/horizon, not a measured energy or latency result. Appendix F does count the FFT and mask-generation operations and uses measured spike rates, so the accounting is more careful than a simple MAC count. Still, it excludes the prediction head for all methods and cannot capture memory traffic, sparse-kernel overhead or the cost of materializing and applying the mask on actual hardware. The ECL comparison also shows 67.2K parameters for SpikeLite against 1.2K for SpikF; the proposed model is lower in estimated energy, but it is not smaller by every reasonable definition of “lightweight.” The current evidence supports a lower estimated encoder operation cost under the stated assumptions; it does not yet establish a lower deployment cost.

Questions for authors

Could the authors compare FSSE plus SSCA against FSSE plus dense SSA and against a simple, parameter-matched channel mixer under the same training budget? Per-dataset mask density, actual skipped operations and latency would help determine whether sparsity, rather than merely channel mixing, drives the claimed efficiency gain. This is especially relevant given the large Traffic ablation gap.

For the close long-term forecasting margins, can the authors rerun the strongest imported baselines under one code and tuning protocol, or report paired uncertainty from matched runs? Could they also provide end-to-end latency or measured energy on a concrete target, including the prediction head? I would weigh the paper more favorably if the small accuracy improvement and the estimated energy saving held under those comparisons.

2 comments

We sincerely thank the reviewer for the constructive and detailed evaluation. We provide a point-by-point response to each of your valuable comments to ensure clarity and address all aspects thoroughly.

Q1. Effectiveness of SSCA

The central distinction is the role of the channel mask. SSCA uses a binary mask to gate pairwise scores in no-softmax attention over binary, spike-generated Q/K/V. DUET uses channel relationships to select interactions in its normalized continuous-attention formulation. SSCA shares the use of rFFT descriptors, learnable distances, reciprocal affinities, and row-wise normalization with DUET. We use rFFT because the FSSE representation retains frequency-selective structure. Our design then forms a low-rank relation metric and applies deterministic thresholding for spike-driven channel routing.

We evaluate three DUET-style mask variants within the same SpikeLite forecasting setup. The raw-versus-latent comparison tests descriptor source, and the latent stochastic-versus-threshold comparison tests discretization. The detailed MSE/MAE table and operation-model energy estimates are provided in our responses to Reviewer FzWT’s Q1 and Q2. In summary, SSCA is competitive with dense SSA, outperforms the same-density random-mask control in 11 of 12 settings, and the global mixer has higher forecasting errors across the reported settings. On Traffic, the mask-aware operation estimate is 321.85 μJ/sample for SSCA versus 429.12 μJ/sample for dense SSA.

Q2. Details of Our Results

First we tested SpikF, the strongest baseline, and its results are consistent with those reported in the table. In addition, we supplemented the evaluation with results for SpikHRFSSM, as shown below.

SpikeLite obtains average MSE/MAE of 0.3434/0.3446, compared with 0.3458/0.3456 for SpikF, and 0.3492/0.3470 for SpikHRFSSM. The results show that SpikF remains the best-performing method, this improvement is both stable and clear.

MethodECL MSEECL MAEWeather MSEWeather MAEETTh1 MSEETTh1 MAEETTh2 MSEETTh2 MAEETTm1 MSEETTm1 MAEETTm2 MSEETTm2 MAETraffic MSETraffic MAEExchange MSEExchange MAEAvg. MSEAvg. MAE
SpikF0.1830.2750.2450.2650.440⁎0.428⁎0.3720.3940.3880.385⁎0.2810.3200.4970.2960.360⁎0.402†​0.34580.3456†
SpikeLite0.175⁎0.263⁎0.2450.265†​0.443†​0.429†​0.3740.394†​0.387†​0.389†​0.2780.319†​0.479⁎0.294†​0.366†​0.4040.3434⁎0.3446⁎
SpikHRFSSM0.1950.2750.242†​0.264⁎0.443†​0.4310.368†​0.393⁎0.3900.389†​0.274⁎0.316⁎0.4940.291⁎0.3880.4170.34930.3470

Then we report our energy consumption on ECL-720 and Weather-720. Compared with SpikF, currently the most energy-efficient module, we achieve better energy efficiency both with and without the prediction head. The advantage is especially pronounced in the prediction head: SpikeLite uses a single-layer linear readout, whereas SpikF uses an MLP decoder applied across temporal groups, so according to the code-shape computation, the prediction head of SpikF incurs a higher cost.

All values are in µJ per sample under the 45-nm operation-cost model (MAC = 4.6 pJ; AC = 0.9 pJ).

Dataset / methodEncoder estimatePrediction head / decoderTotal
ECL@720 SpikeLite95.30136.08231.38
ECL@720 SpikF117.661,259.871,377.35
Weather@720 SpikeLite23.5525.0448.59
Weather@720 SpikF112.571,255.271,367.84

Table 2. Weather@720 component breakdown

MethodComponentOperation type / assumptionEnergy
SpikeLiteFSSE LIF branch updatesAC, r=0.16340.0027
SpikeLiteFour frequency-band projectionsMAC13.3540
SpikeLiteIdentity projectionMAC3.3385
SpikeLiteSSCA input LIFAC, r=0.16340.0089
SpikeLiteSSCA Q/K/V projectionsAC, r=0.16344.8029
SpikeLiteSSCA QKᵀ and AVDense-attention upper bound0.1868
SpikeLiteSSCA output projectionAC, r=0.16341.6010
SpikeLiteMask generationMAC, once per input window0.4352

Table 3. Traffic @720 component breakdown

FSSE componentOperation countAccountingEnergy (µJ/sample)
LIF branch updates0.324 MAC, r=0.15030.0438
Three frequency-band projections20.708 MDense MAC95.2584
FSSE encoder total95.3022
Reviewer_FzWTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work propose SpikeLite for temporal forecasting tasks in SNNs. The wholw model contains two main parts: one is the Frequency-Selective Spiking Encoder for freq.-selective encoding, and the other is Sparse Spiking Channel Attention for cross-channel attention. FSSE uses parallel LIF branches with learnable decay factors to produce frequency-sensitive responses. Successive differences between branch responses, together with an input residual, form components that collectively reconstruct the input before projection. These components are independently projected and aggregated into compact channel representations. SSCA selectively models cross-channel dependencies through a learned binary mask applied to spike-driven self-attention. Its mask-generation pipeline closely resembles DUET’s frequency-domain channel modeling, including spectral descriptors, learned distances, inverse-distance affinities, and row-wise normalization. The evaluation covers four standard multivariate forecasting benchmarks under the SeqSNN protocol and eight long-term forecasting benchmarks under the SpikF protocol, comparing against ANN and SNN baselines.

Critical strengths
  1. Under both the SeqSNN and SpikF evaluation protocols, SpikeLite achieves the best overall results among all compared methods, performing exceptionally well on electricity forecasting datasets. Across various datasets and different forecasting horizons, it demonstrates superior competitiveness compared to artificial neural networks (ANNs) and current SOTA spiking neural network (SNN) baseline models for time-series tasks. (importance 8/10)
Critical weaknesses
  1. SSCA shares a core processing framework with DUET, encompassing the mask generation pipeline, Fourier magnitude descriptors, a learnable Mahalanobis-like distance, affinity based on the reciprocal of distance, and row-wise normalization. Although SpikeLite introduces modifications to input representation, mask discretization, and the implementation of the attention mechanism, the paper fails to explicitly articulate any substantive methodological improvements over this closely related design. Furthermore, the study lacks a direct comparison with DUET under identical evaluation conditions, making it difficult to assess the individual contributions of the core components proposed in this work. (importance 9/10)
  2. Section 5.4 and Appendix D evaluate the impact of removing FSSE or SSCA, yet these comparisons do not clarify why their respective designs are necessary. Removing SSCA is equivalent to eliminating the entire channel-interaction module; consequently, the resulting performance degradation does not prove that "learned sparse masking" is superior to a conventional "dense spiking attention" mechanism. Similarly, the evaluation of FSSE lacks a comparison against a "non-spiking filter bank" that employs an equivalent projection approach and parameter scale. (importance 7/10)
Questions for authors

Clarification requests:

  1. What are the differences between SSCA’s mechanism and DUET’s channel clustering module? Can the author isolate the benefits of using encoded representations and deterministic thresholding?
  2. Should SSCA be compared against similar spiking attention modules that employ either a fully connected mask or a random mask with the same level of sparsity—and does SSCA demonstrate any advantages in such a comparison?
  3. Can you provide the energy breakdown across neuronal updates, temporal projections, FFT/mask generation, and attention process, together with their possible implementation scheme on neuromorphic hardware?
4 comments
Reviewer_FzWTOpenReview ↗

Response

Since the authors did not submit a rebuttal before the deadline, the concerns raised in my review remain unaddressed. I will therefore keep my original score.

We sincerely thank you for your careful review. Below is our detailed responses to your concerns.

As this is our first ICLR submission, we apologize for the delayed rebuttal response. We have since completed the supplementary experiments and obtained concrete results.

Q1&W1. Design and Effectiveness of the SSCA

First, we explain the design rationale of SSCA. The purpose of SSCA is to identify which channel-to-channel interactions should participate in spike-driven aggregation. SSCA converts the estimated channel relationships into a binary connectivity mask and applies it directly to the attention scores computed from binary Q/K spikes. The masked scores then aggregate binary V spikes without introducing a softmax. The mask therefore controls event-driven communication between channels.

DUET uses channel relationships to construct a mask for its channel-attention operation, where the mask determines which channel pairs participate in normalized attention. The two masks serve different computational roles: SSCA gates spike-driven, no-softmax aggregation, whereas DUET’s mask organizes channel interaction in its continuous attention formulation.

We agree that SSCA shares several steps with DUET, including rFFT-magnitude descriptors, a learnable distance, reciprocal affinities, and row-wise normalization. However, our motivation for adopting these steps is fundamentally different: we use them to estimate channel relationships from the frequency-selective representation produced by FSSE. Specifically, FSSE first separates the input into components with different frequency responses, and the rFFT descriptor then summarizes the frequency structure of the resulting representation for relation estimation. SSCA subsequently uses a low-rank metric and deterministic thresholding to form a binary mask.

Next, we designed three DUET-style mask variants within the same SpikeLite forecasting model, evaluated on three long-horizon datasets: Traffic, Weather, and ECL (horizons 96, 192, 336, 720). Among them, DUET-raw and DUET-latent use the same full-rank metric and stochastic discretization, while changing the descriptor source from the normalized input to the FSSE representation. DUET-latent and DUET-latent-threshold use the same latent representation and full-rank metric, while changing stochastic discretization to deterministic thresholding. These comparisons isolate the effects of representation source and discretization in the mask generator.

We first estimate the energy of the channel-interaction module using the same 45-nm operation model as in the paper, with 4.6 pJ per dense MAC and 0.9 pJ per accumulation. The estimate includes relation generation, Q/K/V projections, attention, and output projection. On Traffic, the estimated module energy is 463.87 μJ/sample for SSCA, compared with 629.22 for DUET-raw, 1,676.45 for DUET-latent, and 1,646.55 for DUET-latent threshold under the dense-operation accounting. Under an ideal mask-aware implementation that skips zero connections, the corresponding estimates are 321.85, 502.35, 1,566.66, and 1,603.34 μJ/sample. Thus, the mask-aware estimates are lower for SSCA in this setting, while the comparison also shows that mask-generation cost depends strongly on the descriptor and metric configuration.

Dataset (Avg.)SSCADUET-rawDUET-latentDUET-latent threshold
ECL178.77 / 157.64 μJ (20.3%)203.84 / 187.88 μJ (46.2%)549.77 / 540.10 (67.4%)548.35 / 544.58 (87.3%)
Traffic463.87 / 321.85 μJ (11.0%)629.22 / 502.35 μJ (27.8%)1,676.45 / 1,566.66 μJ (35.2%)1,646.55 / 1,603.34 μJ (71.8%)

Next Table compares forecasting accuracy. Across the three benchmarks, SSCA remains close to the DUET-style variants in MSE and MAE, while its low-rank deterministic mask provides the lower-cost relation mechanism in the profiled settings.

DatasetHorizonSpikeLite (SSCA) MSE / MAEDUET-raw MSE / MAEDUET-latent MSE / MAEDUET-latent threshold MSE / MAE
ECL96.145 / .236.143 / .233.143 / .231.145 / .235
ECL192.161 / .250.160 / .248.163 / .250.161 / .248
ECL336.175 / .264.172 / .262.174 / .263.172 / .261
ECL720.218 / .300.213 / .294.216 / .297.215 / .296
Weather96.158 / .196.158 / .196.157 / .196.157 / .197
Weather192.209 / .243.209 / .244.208 / .243.209 / .243
Weather336.266 / .284.264 / .285.266 / .285.267 / .285
Weather720.346 / .337.343 / .337.346 / .338.346 / .338
Traffic96.453 / .285.438 / .275.439 / .278.438 / .276
Traffic192.468 / .287.455 / .278.464 / .281.456 / .278
Traffic336.481 / .293.471 / .286.479 / .290.471 / .285
Traffic720.514 / .311.508 / .309.510 / .311.506 / .308

Q2&W2. Effectiveness of SSCA We compare SSCA with three channel-interaction controls using the same FSSE forecasting setup and training protocol. Dense SSA removes the channel mask while retaining spiking Q/K/V and attention. The random-mask control preserves each SSCA row’s number of selected channels while randomizing neighbor identities, testing whether learned channel selection helps beyond sparsity alone. The global mixer replaces pairwise attention with a continuous low-rank global-context gate.

The global mixer’s parameter budget matches the SSCA mask generator, rather than the complete SSCA module. Its interaction module has 2.88K parameters on ECL, 2.16K on Traffic, and 1.44K on Weather. We therefore use it as a low-parameter continuous-mixing baseline.

Next table shows that SSCA outperforms the same-density random-mask control on 11 of 12 settings; ECL at horizon 720 is the exception. SSCA and dense SSA have broadly comparable forecasting accuracy, while dense SSA has lower errors on Traffic in the reported results. The global mixer has higher MSE and MAE than SSCA on all 12 settings. We use the dense and random-mask comparisons to evaluate the accuracy–sparsity trade-off, rather than claiming that SSCA uniformly improves accuracy over dense SSA.

DatasetHorizonSpikeLite SSCADense SSARandom maskGlobal mixer
ECL96.145 / .236.144 / .236.150 / .238.157 / .243
ECL192.161 / .250.163 / .253.166 / .254.171 / .260
ECL336.175 / .264.178 / .267.181 / .268.184 / .273
ECL720.218 / .300.216 / .301.216 / .298.225 / .304
Weather96.158 / .196.158 / .197.167 / .209.167 / .205
Weather192.209 / .243.208 / .243.216 / .257.217 / .253
Weather336.266 / .284.266 / .285.270 / .289.273 / .291
Weather720.346 / .337.346 / .339.349 / .340.352 / .344
Traffic96.453 / .285.447 / .279.466 / .289.484 / .294
Traffic192.468 / .287.458 / .232.474 / .292.491 / .299
Traffic336.481 / .293.479 / .290.484 / .297.519 / .312
Traffic720.514 / .311.510 / .309.522 / .315.543 / .320

For energy, we use the paper’s 45-nm operation model, with 4.6 pJ per dense MAC and 0.9 pJ per accumulation. On Traffic, the SSCA module’s operation-model estimate is 463.87 μJ/sample on the dense execution path.

Interaction moduleOff-diagonal mask densityDense-path estimate (μJ/sample)Mask-aware estimate (μJ/sample)
SSCA11.0%463.87321.85
Dense SSA100%429.12429.12
Random mask11.2%502.71341.10
Global mixer—≈2.57≈2.57

Q3. Detailed Energy Analysis

We report our energy consumption on ECL-720 and Weather-720. Compared with SpikF, currently the most energy-efficient module, we achieve better energy efficiency both with and without the prediction head. The advantage is especially pronounced in the prediction head: SpikeLite uses a single-layer linear readout, whereas SpikF uses an MLP decoder applied across temporal groups, so according to the code-shape computation, the prediction head of SpikF incurs a higher cost.

Table 1. Energy estimates

All values are in µJ per sample under the 45-nm operation-cost model (MAC = 4.6 pJ; AC = 0.9 pJ).

Dataset / methodEncoder estimatePrediction head / decoderTotal
ECL@720 SpikeLite95.30136.08231.38
ECL@720 SpikF117.661,259.871,377.35
Weather@720 SpikeLite23.5525.0448.59
Weather@720 SpikF112.571,255.271,367.84

Table 2. Weather@720 component breakdown

MethodComponentOperation type / assumptionEnergy
SpikeLiteFSSE LIF branch updatesAC, r=0.16340.0027
SpikeLiteFour frequency-band projectionsMAC13.3540
SpikeLiteIdentity projectionMAC3.3385
SpikeLiteSSCA input LIFAC, r=0.16340.0089
SpikeLiteSSCA Q/K/V projectionsAC, r=0.16344.8029
SpikeLiteSSCA QKᵀ and AVDense-attention upper bound0.1868
SpikeLiteSSCA output projectionAC, r=0.16341.6010
SpikeLiteMask generationMAC, once per input window0.4352

Table 3. Traffic @720 component breakdown

FSSE componentOperation countAccountingEnergy (µJ/sample)
LIF branch updates0.324 MAC, r=0.15030.0438
Three frequency-band projections20.708 MDense MAC95.2584
FSSE encoder total95.3022 ≈ 95.30
Reviewer_coLAOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work proposes a lightweight SNN model, unlike recent designs, that implements a (1) Frequency-Selective Spiking Encoder and (2) Sparse Spiking Channel Attention (SSCA). The model is benchmarked on both standard and long-horizon forecasting tasks and is compared with recent models such as SpikF.

Critical strengths

Strengths

  1. The goal of designing lightweight models is well motivated
  2. The Frequency-Selective Spiking Encoder (FSSE) utilises the LIF dynamics and SCA, which appears novel.
  3. The authors use a well-established experimental setup to evaluate the model against a wide range of ANN and SNN baselines across multiple diverse benchmark datasets.
Critical weaknesses

Weaknesses

  1. Limited gain over STOA methods like SpikF. Whereas SpikF provides a relatively detailed, still theoretical, estimate of the efficiency breakdown.
  2. It appears the advantage of SCA is very dataset-dependent.
  3. The FSSE requires significant handcrafting, especially during initialisation.
Questions for authors
  1. Please report the end-to-end energy, wall-clock time and memory consumption of your model. Also, include the energy consumption breakdown, such as for data movement, etc.
  2. Although it's described as lightweight, this model hasn't been implemented on any neuromorphic hardware, making it difficult to clearly understand its real benefits. As of now, the benefits of this model relative to previous models are limited; please comment on the specific benefits of this approach.
  3. Could you also evaluate your work on the TIME benchmark [1]?
  4. It appears this work misses some recent key works, such as Wang et al. [2] and Agrawal et al. [3]. Please compare your work with respect to these recent works.

References

[1] https://github.com/zqiao11/TIME

[2] S. Wang et al., "TimeMixer++: A General Time Series Pattern Machine for Universal Predictive Analysis," in The Thirteenth International Conference on Learning Representations (ICLR 2025)

[3] K. Agrawal et al., "A Spiking Heterogeneous Harmonic Resonate-and-Fire State Space Model for Time Series," in Proceedings of the 43rd International Conference on Machine Learning (ICML 2026)


4 comments

We sincerely thank the reviewer for the thorough and constructive feedback. Our responses are provided below.

Q3. Evaluation Scope and TIME We thank the reviewer for pointing us to TIME. SpikeLite is designed for supervised forecasting, whereas TIME evaluates zero-shot generalization of pretrained time-series foundation models. These protocols address different evaluation questions, so applying SpikeLite’s current training procedure to TIME tasks would not constitute a valid TIME evaluation.

Our evaluation already covers two complementary forecasting protocols: the SeqSNN protocol includes four multivariate datasets (METR-LA, PEMS-BAY, Solar, and Electricity); the SpikF protocol includes eight long-term forecasting datasets across traffic, electricity, weather, transformer, and exchange-rate domains. This amounts to 12 datasets and 48 dataset–horizon settings under established supervised protocols. Our evaluation datasets essentially cover all dataset types under both conventional time-series forecasting and long-horizon forecasting tasks.

Q4. Recent Baselines and Performance Claim We thank the reviewer for highlighting these recent methods. The official TimeMixer repository provides the original TimeMixer implementation, but we could not find an author-maintained implementation of TimeMixer++ there. We therefore compare against the reproducible original TimeMixer and SpikHRFSSM baselines under the long-term forecasting protocol, and recompute their averages over the eight datasets. SpikeLite obtains average MSE/MAE of 0.3434/0.3446, compared with 0.3455/0.3476 for TimeMixer, 0.3458/0.3456 for SpikF, and 0.3492/0.3470 for SpikHRFSSM. The results show that SpikF remains the best-performing method, and as demonstrated in response to Q2, this improvement is both stable and clear.

MethodECL MSEECL MAEWeather MSEWeather MAEETTh1 MSEETTh1 MAEETTh2 MSEETTh2 MAEETTm1 MSEETTm1 MAEETTm2 MSEETTm2 MAETraffic MSETraffic MAEExchange MSEExchange MAEAvg. MSEAvg. MAE
SpikF0.1830.2750.2450.2650.440⁎0.428⁎0.3720.3940.3880.385⁎0.2810.3200.4970.2960.360⁎0.402†​0.34580.3456†
SpikeLite0.175⁎0.263⁎0.2450.265†​0.443†​0.429†​0.3740.394†​0.387†​0.389†​0.2780.319†​0.479⁎0.294†​0.366†​0.4040.3434⁎0.3446⁎
TimeMixer0.182†​0.272†​0.240⁎0.2710.4470.4400.364⁎0.3950.381⁎0.3950.275†​0.3230.484†​0.2970.3910.388⁎0.3455†​0.3476
SpikHRFSSM0.1950.2750.242†​0.264⁎0.443†​0.4310.368†​0.393⁎0.3900.389†​0.274⁎0.316⁎0.4940.291⁎0.3880.4170.34930.3470

W2. SSCA Benefits Across Datasets As shown by our ablation results, SSCA improves forecasting performance on datasets with both many and few channels. SSCA is designed to enable efficient cross-channel information exchange. Its gains are larger on Traffic, where modeling relationships among a large number of sensors provides greater benefits. On Weather, the improvement is smaller because FSSE already captures much of the useful temporal information, leaving less room for additional gains from cross-channel interaction. We view this variation in gain magnitude not as a weakness, but as evidence that SSCA provides consistent, complementary benefits across datasets with different channel structures.

W3. Claim of FSSE Design The FSSE initialization is simple: we assign distinct initial decay factors to a small number of parallel LIF branches to provide different temporal responses. These values are learnable and are optimized jointly with the forecasting objective; we do not hand-design dataset-specific filters or impose a fixed ordering after initialization. Appendix E, Fig. 5(e–f), already reports the initialized and trained decay factors, together with branch-order changes and the relationship between decay factors and spectral centroids. The results show that the learned branches retain frequency-sensitive responses while adapting their parameters during training. Thus, FSSE uses a lightweight initialization heuristic rather than extensive manual filter design.

Q1. Detailed Energy Analysis

We report our energy consumption on ECL-720 and Weather-720. Compared with SpikF, currently the most energy-efficient module, we achieve better energy efficiency both with and without the prediction head. The advantage is especially pronounced in the prediction head: SpikeLite uses a single-layer linear readout, whereas SpikF uses an MLP decoder applied across temporal groups, so according to the code-shape computation, the prediction head of SpikF incurs a higher cost.

Table 1. Energy estimates

All values are in µJ per sample under the 45-nm operation-cost model (MAC = 4.6 pJ; AC = 0.9 pJ).

Dataset / methodEncoder estimatePrediction head / decoderTotal
ECL@720 SpikeLite95.30136.08231.38
ECL@720 SpikF117.661,259.871,377.35
Weather@720 SpikeLite23.5525.0448.59
Weather@720 SpikF112.571,255.271,367.84

Table 2. Weather@720 component breakdown

MethodComponentOperation type / assumptionEnergy
SpikeLiteFSSE LIF branch updatesAC, r=0.16340.0027
SpikeLiteFour frequency-band projectionsMAC13.3540
SpikeLiteIdentity projectionMAC3.3385
SpikeLiteSSCA input LIFAC, r=0.16340.0089
SpikeLiteSSCA Q/K/V projectionsAC, r=0.16344.8029
SpikeLiteSSCA QKᵀ and AVDense-attention upper bound0.1868
SpikeLiteSSCA output projectionAC, r=0.16341.6010
SpikeLiteMask generationMAC, once per input window0.4352

Table 3. Traffic @720 component breakdown

FSSE componentOperation countAccountingEnergy (µJ/sample)
LIF branch updates0.324 MAC, r=0.15030.0438
Three frequency-band projections20.708 MDense MAC95.2584
FSSE encoder total95.3022 ≈ 95.30

Q2&W1. SpikeLite's Performance Claim

Our mean ± standard deviation results show low variability across repeated runs. This small run-to-run variance does not mean that the performance gains are negligible: SpikeLite achieves the best or second-best results in most evaluation settings and obtains the best average performance, demonstrating both stable and strong forecasting performance.

Specifically,across the 16 SeqSNN settings, SpikeLite outperforms SpikF on 13 settings in R² and on 12 settings in RSE, with one RSE tie. Its mean ± standard deviation results also show low run-to-run variability. Thus, the small variance is accompanied by consistently strong performance: SpikeLite achieves the best or second-best results in most settings and the best aggregate scores.

SpikeLite has not yet been implemented on neuromorphic hardware, so we do not claim measured hardware-level energy or latency benefits. Its demonstrated benefit is forecasting performance across two complementary supervised protocols. In the long-term benchmark, SpikeLite achieves average MSE/MAE of 0.3434/0.3446, compared with 0.3458/0.3456 for SpikF. We additionally evaluated SpikF on the SeqSNN benchmark, which covers four standard multivariate datasets at four horizons. Across these 16 tasks, SpikeLite obtains higher average R² and lower average RSE than SpikF, with better results on 13/16 tasks for R² and 12/16 tasks for RSE (one tie). Thus, SpikF is less competitive with SpikeLite on this complementary benchmark.

MethodMetricMETR-LA-6METR-LA-24METR-LA-48METR-LA-96PEMS-BAY-6PEMS-BAY-24PEMS-BAY-48PEMS-BAY-96Solar-6Solar-24Solar-48Solar-96Electricity-6Electricity-24Electricity-48Electricity-96Average
SpikFR² ↑0.836†0.601†0.464⁎0.330⁎0.885†0.754†0.713†0.6669⁎0.954†0.871†0.807†0.753†0.991†0.987†0.985†0.981†0.7862†
SpikFRSE ↓0.429†0.657†0.766⁎0.859⁎0.364†0.526†0.581†0.593⁎0.204†0.349†0.446⁎†0.498⁎0.198†0.197†0.216†0.246†0.4456†
SpikeLiteR² ↑0.847⁎0.622⁎0.452†0.316†0.889⁎0.781⁎0.715⁎0.639†0.962⁎0.885⁎0.811⁎0.767⁎0.993⁎0.991⁎0.988⁎0.983⁎0.7901⁎
SpikeLiteRSE ↓0.414⁎0.649⁎0.781†0.873†0.361⁎0.506⁎0.577⁎0.649†0.201⁎0.348⁎0.446⁎†0.496⁎0.146⁎0.167⁎0.196⁎0.233⁎0.4402⁎

Best results are bold with an asterisk (⁎); second-best results are bold with a dagger (†). The Solar-48 RSE values are tied, so both are marked as best and second-best.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.