acceptodds
Under review as a conference paper at ICLR 2027

SNN-FRFT: ADAPTIVE FRACTIONAL REPRESENTATIONS FOR EFFICIENT SPIKING FORECASTING

Abstract

A fixed orthogonal Fourier basis spreads chirp-like signals across coefficients, limiting sparse spiking forecasting. We introduce a spiking neural network with an adaptive fractional Fourier transform (SNN-FrFT). A spiking selector estimates the order from the observed history; the forecasting head predicts residual dynamics, and the same order reconstructs the future phase. Signed Dual-Region Log-Temporal Coding (DR-LTC) covers strong coefficients logarithmically and maps sufficiently weak coefficients to zero-event states. The compact head projects once and reuses current across leaky integrate-and-fire steps. On a three-seed repartitioned NeLoRa benchmark, SNN-FrFT achieves 79.28% spectral peak agreement at spreading factor 10 versus 73.67% for Spiking Fourier Network (SpikF), with 97.23% fewer core multiply-accumulate operations. Synthetic factorization connects the prediction gain to pairing the adaptive input representation with fu￾ture reconstruction. A matched three-epoch coding study improves NeLoRa peak agreement over linear time-to-first-spike coding while leaving 74.06% of terminal rails silent. On the same HM100, the compiled width-256 cached head uses 38.3% less mean board-level dynamic energy than SpikF’s decoder.

Primary area:
learning on time series and dynamical systems
Keywords:
Spiking Neural Networks, Fractional Fourier Transform, Temporal Coding, Neuromorphic Edge Computing

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_YvLROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper presents SNN-FrFT, an efficient spiking forecasting framework that couples history-conditioned fractional Fourier representations with matched future-phase reconstruction. A lightweight spiking selector estimates the transform order, while dual-region temporal coding suppresses weak coefficients and a compact forecasting head reuses projected current across simulation steps. On NeLoRa SF10, SNN-FrFT improves spectral peak agreement from 73.67% to 79.28% with 97.23% fewer core multiply-accumulate operations; its cached head also reduces mean dynamic energy by 38.3% on the HM100 neuromorphic platform.

Critical strengths
  • Well-motivated representation–reconstruction coupling. The method addresses a concrete limitation of fixed Fourier representations for chirp-like signals. Using the same history-estimated order for input adaptation and future-phase restoration gives the architecture a clear physical interpretation. The controlled 2 × 2 ablation is particularly useful: neither component alone substantially improves forecasting, whereas their combination reduces synthetic-chirp MAE by 20.32%.
  • Effective and inexpensive order selection. A-SAS estimates an input-dependent order without future observations or spreading-factor metadata. Table 2 compares it against analytic estimation, fixed and globally learned orders, and exhaustive concentration-based search. Its performance is close to exhaustive search while requiring only 2,496 selector MACs per window, supporting the practical value of the learned selector.
  • Strong computational efficiency with useful task-level gains. On NeLoRa SF10, the method improves spectral peak agreement from 73.67% to 79.28% while reducing core matrix MACs by 97.23%. The paper also includes fixed-transform arithmetic in a broader cost estimate and separates the compact head’s efficiency from the adaptive representation’s accuracy contribution. This accounting makes the efficiency claim easier to assess.
  • Informative coding and sparsity analysis. DR-LTC explicitly balances strong-coefficient coverage, near-zero resolution, and event suppression. The matched coding comparisons and gate-off control help explain this tradeoff. The additional threshold sweep across ten frozen heads provides useful evidence that substantial terminal-event reductions can coexist with stable peak agreement over a range of thresholds.
  • Valuable downstream and hardware validation. The held-out SF7 demodulation experiment extends the evaluation beyond within-symbol forecasting and includes wrapped payload symbols. With the same spiking classifier architecture, SNN-FrFT outperforms SpikF features at every reported SNR. The HM100 profiling further demonstrates practical deployment potential, with 38.3% lower mean dynamic energy and 2.07× higher sustained throughput for the width-256 cached head.
Critical weaknesses
  • The accuracy advantage is strongest for spectral localization and chirp-related tasks. Appendix D reports substantially higher I/Q waveform errors than SpikF, and the main width-256 model also has higher MAE on the three conventional time-series datasets. The paper would benefit from a clearer statement of which forecasting objectives best match the proposed representation.
  • Some evaluation boundaries limit the breadth of the conclusions. The main NeLoRa test population was previously involved in method development, although the subsequent head-training protocol uses disjoint packets. Hardware measurements cover cached heads and exclude both frontends. These limitations are disclosed clearly, and the fresh SF7 experiment provides helpful additional evidence.
Questions for authors

See the Weaknesses.

1 comment

We thank Reviewer yvlr for the review and for the close reading of our evaluation protocol.

W1. Which forecasting objectives match the method.

Appendix H (Table 19) answers the point on the conventional datasets. Series without chirp structure give a per-window order nothing to adapt to, so they use the matching configuration of the same model: a fixed order of 0.981 (the median selector output on training histories) and a width-1024 compact head. With it, SNN-FrFT matches or improves on SpikF on four of the five datasets, with 97.4–98.5% fewer learned MACs:

DatasetMAE vs SpikFMSE vs SpikFLearned MACs vs SpikF
ETTh1 (3 seeds)−0.02%−1.27%−98.47%
ETTm1 (3 seeds)+0.18%−0.34%−98.47%
Weather (1 seed)+5.57%+9.05%−97.44%
Beijing PRSA (new to both methods)−1.07%−2.06%−98.46%
UCI HPC (new to both methods)−2.38%−1.07%−98.47%

PRSA and HPC were new to both methods, and their test splits were read once, after model selection was locked. The width-256 adaptive model in the main tables is the configuration for chirp signals.

The method is built for objectives defined by where a chirp's spectral energy lies: the dechirped peak bin, the symbol decision, or a frequency and chirp-rate track. On NeLoRa, SNN-FrFT reaches 92.775% peak agreement against 91.363% (SF10: 79.279% against 73.674%). On held-out SF7 packets it decodes symbols more accurately at every SNR (64.03% against 25.12% at −16 dB, 92.93% against 49.03% at 0 dB). Sample-wise waveform fidelity is a different objective. SpikF's dense decoder regresses every I/Q sample and reaches a lower I/Q MAE (0.000944 against 0.004333, Appendix D). For applications that use the forecast through a spectral decision, such as LoRa demodulation, chirp tracking and range-Doppler processing, peak agreement and decoding accuracy are the relevant measures.

W2. Scope of the evaluation.

As described in Appendix A.1, we used the original NeLoRa development pool to develop the frontend before repartitioning it. Under the revised protocol, we retrained all prediction heads from scratch, selected their checkpoints using validation packets disjoint from the forecasting test packets, and read the test tensor once per locked model. The SF7 demodulation study in Appendix E provides independent evidence on a downstream task, using recordings that were never used in development: SNN-FrFT is more accurate than SpikF at every SNR from −22 to 0 dB (92.93% against 49.03% at 0 dB).

Table 4 compares the two cached prediction heads on the same HM100 at batch size 32, with both frontends outside the measurement boundary. The 38.3% reduction in dynamic energy applies to that comparison. Appendix C separately reports operation counts for the full pipelines, including both frontends: SpikF's frontend uses 9,011,200 learned MACs per window, while ours uses a 2,496-MAC selector plus 8,448 complex multiplications for the FrFT, and the full pipeline needs ≈691,904 multiply-adds against 23,756,800 (−97.09%).

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

The paper proposes SNN-FrFT for forecasting complex chirp signals. A history-conditioned spiking selector predicts a fractional Fourier order from phase-curvature features. A pre-chirp/FFT/post-chirp frontend uses this order, signed dual-region logarithmic temporal coding (DR-LTC) quantizes intermediate rails, and a compact spiking head reuses one projected current across 16 LIF steps. The selected order is also used to restore the future quadratic phase. On the measured NeLoRa SF10 test, peak agreement is 79.28% versus 73.67% for SpikF over three head seeds.

Critical strengths

Controlled evidence for the representation/restoration coupling. The matched synthetic 2-by-2 comparison (Sec. 4.3; Table 7) reduces I/Q MAE from 0.645774 (fixed FFT, no restoration) to 0.514547 when both adaptive input and matching restoration are enabled. More importantly, the compact fixed-FFT NeLoRa control reaches 75.48% SF10 agreement versus 79.28% for the adaptive method at almost identical head MAC cost (Sec. 4.2). This is a useful test of the adaptive representation beyond the comparison to a much larger SpikF execution graph. The synthetic result is one paired initialization and should not be overgeneralized.

Practical measurement and transparent limitations. The authors report three head-training seeds, matched coding comparisons, CFO perturbations, and real hardware profiling (Tables 1-3). DR-LTC reduces terminal events from 1111.4 (gate off) to 531.2 per window while SF10 agreement changes only from 75.08% to 74.47% in the short coding study. The reported HM100 head measurements establish a real compiled-implementation comparison.

Critical weaknesses

The primary benefit is not adequately separated from simple chirp modeling and compact-head design. Phase-curvature-based chirp-rate estimation followed by analytic dechirping/extrapolation is a natural low-cost alternative, but no such baseline is reported. A non-spiking head with the same adaptive representation and a compact current-reuse control would help assess which parts genuinely need spiking dynamics. The 97.23% MAC reduction versus SpikF mostly compares different execution architectures: the proposed head projects once while SpikF processes temporally expanded features; the proposed model also has more parameters (658,977 versus 471,613). The matched FFT-head comparison is more interpretable and should be foregrounded. Finally, Table 1 emphasizes raw 256-point FFT peak agreement rather than direct I/Q forecasting loss or end-task LoRa demodulation. Matched held-out NeLoRa I/Q MAE/MSE for both full models are not reported there, while the historical compiled-head diagnostic set in Table 6 shows lower SpikF MAE (0.00244391) than SNN-FrFT-256 (0.00798814). Those are not matched final test results, but they make the missing forecasting comparison important.

Event sparsity does not yet establish full-system neuromorphic savings. Section 3.3 states that software reconstructs coded values and performs structured arithmetic, counting spike events separately from dense operations; a 74.06% silent terminal-rail fraction is therefore not a 74% reduction in actual frontend operations or energy. The 38.3% HM100 dynamic-energy reduction in Table 3 is for cached heads only: both frontends are excluded, and the packed transfer interfaces differ greatly (32,772 versus 2,050 float32 values/window). More seriously, all three compiled HM100 heads fail the prespecified software-equivalence criterion (Appendix A.4, Table 6). Consequently the hardware data support a restricted comparison of two different compiled heads, not yet a validated end-to-end energy advantage for equivalent forecasting functionality. The lower SF10 peak-bin agreement on CFO in error-bin terms (0.6180 versus 0.5299), and 7-10% worse MAE on ETTh1/ETTm1/Weather, also constrain the generality of the conclusions.

Questions for authors

Could the authors report matched NeLoRa test-set I/Q MAE and MSE for SpikF, fixed-FFT compact head, and SNN-FrFT, ideally by SF, alongside peak agreement? The historical Table 6 MAEs are not a replacement for the held-out benchmark.

How does A-SAS compare with a direct phase-curvature/chirp-rate estimator and with selecting the FrFT order by minimizing k90 using the observed history only? This comparison is important to determine the contribution of the learned spiking selector.

At matched representation, head width, parameter budget and execution boundary, what is the benefit of the LIF current-reuse head over a compact ANN/MLP readout or equivalent quantized nonlinear head? Please separate architectural caching gains from representation gains in the efficiency claim.

How should the HM100 energy comparison be interpreted given the failed output-equivalence tests and very different feature-transfer sizes? Can the authors report task metrics of the compiled outputs, quantify any output discrepancy by model, and clarify how much energy/time is attributable to transfers versus accelerator execution? A full-pipeline neuromorphic deployment need not be completed during rebuttal.

The full DR-LTC model attains 74.47% SF10 peak agreement versus 75.18% for single-region log coding and 75.08% for DR-LTC without the terminal gate (Sec. 4.5). Is the intended claim principally a sparsity/accuracy Pareto improvement rather than a peak-accuracy gain over prior log coding? A small threshold sweep would help substantiate this.

Appendix A.1 notes that the new NeLoRa test population comes from a source development pool used during earlier method development. What decisions were made using that population, and how might this affect test-set independence?

5 comments

We thank Reviewer akyk for the thorough and constructive questions on metrics, order selection, coding and test independence; each prompted a new analysis, reported below.

Q1. Matched NeLoRa test-set I/Q MAE and MSE by SF, alongside peak agreement.

Same 1,015 held-out test windows and final models, mean over 3 seeds; I/Q errors are averaged over the real and imaginary components.

SF (windows)MethodPeak agreementI/Q MAEI/Q MSE
all (1,015)SpikF91.363%0.0009441.58e-6
all (1,015)SNN-FrFT92.775%0.0043334.55e-5
8 (251)SpikF100.000%0.0006176.21e-7
8 (251)SNN-FrFT98.406%0.0023341.09e-5
9 (431)SpikF100.000%0.0010761.88e-6
9 (431)SNN-FrFT99.923%0.0048285.04e-5
10 (333)SpikF73.674%0.0010201.91e-6
10 (333)SNN-FrFT79.279%0.0052006.53e-5

The I/Q MAE is 0.004333 for SNN-FrFT and 0.000944 for SpikF over all windows, and 0.005200 versus 0.001020 at SF10. SNN-FrFT remains ahead of SpikF in spectral-peak agreement, 92.775% versus 91.363% over all windows and 79.279% versus 73.674% at SF10, and in compute, 658,112 versus 23,756,800 learned multiply-adds per window (−97.23%; −97.09% with the fixed FrFT arithmetic included, see G3). The revision adds this table as Table 13 (Appendix D). The demodulation study (G1) adds an end-task measure: with the same spiking head, the FrFT representation is more accurate than SpikF features at every tested SNR, e.g. 92.93% versus 49.03% at 0 dB.

Q2. A-SAS versus a direct phase-curvature (chirp-rate) estimator and versus history-only k90 minimization.

We ran both, together with fixed and learned global orders; the full table is G2. The closed-form estimate, computed from a lag-16 phase-curvature statistic of the same history, reaches 70.971% SF10 peak agreement, 8.3 pp below A-SAS (79.279%). Minimizing k90 over the history is the strongest non-learned option at 78.979%; A-SAS matches it with one transform per window instead of ≈1,700. The learned spiking selector thus turns an expensive search into a single 2,496-MAC inference. With SF metadata, a per-SF lookup (81.181%) is the natural choice, and Sec. 4.2 of the revision describes A-SAS as the estimator for windows whose chirp rate is not known in advance.

Q3. Separating architectural (caching) gains from representation gains in the efficiency claim.

The 97.230% figure is the reduction in core learned-matrix multiply-adds per window. Module-level counts split it as follows:

ComponentSpikF MACs / windowSNN-FrFT MACs / windowShare of the total reduction
Forecasting-head matrices14,745,600655,61661.00%
Non-head matrices9,011,2002,49639.00%
Core matrices, total23,756,800658,11297.230% total reduction

SpikF's head matrices are dense1 and dense2; its non-head matrices are the patch projector, the time-to-frequency layer and the two convolutions of the SFS block. SNN-FrFT's head matrices are the two linear layers of the compact LIF head, and its only non-head matrix is the A-SAS selector. Fixed transform arithmetic, temporal coding and membrane updates are counted separately in G3. The revision adds this breakdown as Table 11 (Appendix C), next to the parameter counts in Table 10 (658,977 for SNN-FrFT, 471,613 for SpikF).

(continued)

Q4. Interpreting the HM100 energy comparison given failed output equivalence and very different transfer sizes.

The compiled heads are equivalent to their software references at the task level. On the 768 SF10 windows of the hardware population, the board-versus-PC difference in peak agreement is 0.000 pp for each model (Table 17 in Appendix G; also in our reply to vkbu Q3); the strict pointwise test fails because 0.147% (w128) and 0.155% (w256) of fc1+LIF spike decisions flip near threshold.

Measured on HM100 at batch 32, in ms per window:

ComponentSpikFSNN-FrFT
Host → device transfer0.104150.01292
Synchronous accelerator call0.067770.01968
Device → host transfer0.009500.00716
Host coordinate restoration0.017340.04989
Other host work0.005240.00532
Cached prediction segment (sum)0.203990.09496

SNN-FrFT sends 2,050 float32 values per window to the device, against 32,772 for SpikF (8,200 versus 131,088 bytes, 16.0× smaller). This makes host-to-device transfer 8.1× faster and accounts for most of the 2.15× segment speed-up; accelerator execution alone is 3.44× faster, while our host-side phase restoration takes 0.0499 ms against 0.0173 ms. The 38.3% dynamic-energy reduction (0.1140 versus 0.0703 mJ/window) applies to this cached segment, with both frontends excluded. The revision presents it as a comparison of two compiled cached heads with equal task output: both tables are in Appendix G, cited from Sec. 4.6, and the full-pipeline count is in Appendix C.

(continued)

Q5. Whether the DR-LTC claim is a sparsity/accuracy trade-off rather than a peak-accuracy gain, and a threshold sweep.

The intended claim is a sparsity/accuracy Pareto improvement, and Sec. 4.5 of the revision now says so. In the coding study, with separately trained heads:

Code (separately trained heads, 3 seeds)Terminal events / windowSF10 peak agreement
Signed linear TTFS331.670.47%
Signed single-region log1,023.875.18%
DR-LTC without terminal gate1,111.475.08%
DR-LTC (gate 0.03 U)531.274.47%

DR-LTC halves terminal events relative to both log codes, and the 0.6–0.7 pp difference is within the across-head SD of each code (1.5–1.7 pp). To isolate the gate, we froze the 10 independently trained main heads and varied only the inference-time threshold from 0 to 0.9 U, with no head, selector or upstream stage retrained. The threshold was preset at 0.03 U before training (U = training-set maximum terminal magnitude ÷ 8 = 4.0023, so 0.120 absolute). Coefficients entering the gate are already DR-LTC-quantized with a smallest nonzero magnitude of 0.0219 U, so thresholds below that value equal gate-off and the gate acts in discrete steps.

Threshold / UTerminal events / windowEvents vs. presetPeak agreement % (95% CI)Δ peak vs. preset, pp (95% CI)SF10 peak %Δ I/Q MAE vs. preset
0 (gate off; ≡ 0.01, 0.02)1111.43+109.2%92.325 [91.889, 92.761]−0.158 [−0.487, +0.172]78.769+7.72%
0.03 (preset; ≡ 0.025)531.23092.483 [92.203, 92.763]078.8290
0.035 / 0.04214.84−59.6%92.512 [92.236, 92.788]+0.030 [−0.273, +0.332]79.009−3.63%
0.045158.91−70.1%92.660 [92.303, 93.017]+0.177 [−0.134, +0.489]79.399−4.70%
0.05 / 0.0685.75−83.9%92.493 [92.225, 92.760]+0.010 [−0.301, +0.321]79.219−5.01%
0.0766.76−87.4%92.433 [92.166, 92.701]−0.049 [−0.337, +0.239]79.129−5.13%
0.0852.91−90.0%92.335 [92.053, 92.616]−0.148 [−0.422, +0.127]78.949−5.17%
0.0945.22−91.5%92.473 [92.268, 92.678]−0.010 [−0.306, +0.286]79.219−5.20%
0.136.95−93.0%92.177 [91.924, 92.431]−0.305 [−0.522, −0.089]79.069−5.01%
0.1229.55−94.4%91.754 [91.375, 92.133]−0.729 [−1.023, −0.435]78.619−4.34%
0.220.16−96.2%90.975 [90.381, 91.570]−1.507 [−2.075, −0.940]77.447−0.74%
0.316.58−96.9%89.537 [89.137, 89.937]−2.946 [−3.332, −2.559]75.946+4.55%
0.513.38−97.5%89.596 [89.174, 90.018]−2.887 [−3.469, −2.305]76.246+20.20%
0.711.78−97.8%88.177 [87.574, 88.781]−4.305 [−5.041, −3.570]77.207+34.72%
0.910.76−98.0%86.660 [85.958, 87.362]−5.823 [−6.673, −4.972]76.637+49.28%

All 1,015 test windows, averaged over 10 heads; CIs are two-sided Student-t intervals across heads and Δ values are paired per head. Figure 9 (Appendix F) plots the full 23-point sweep.

On the same heads, moving from gate-off to the preset halves terminal events (1,111 → 531) with no loss in peak agreement (+0.16 pp, CI includes zero; SF10 +0.06 pp) and lowers I/Q MAE by 7.2%. From 0.035 to 0.09 U, events fall a further 60–92% with no significant change in peak agreement. A small loss appears at 0.1–0.12 U (−0.31 / −0.73 pp), and larger degradation starts at 0.2 U (−1.51 pp). The preset therefore lies inside a broad plateau, at its conservative end; validation shows the same large-drop thresholds (−1.42, −2.55 and −3.38 pp at 0.12, 0.2 and 0.3 U) and slightly earlier small drops (−0.22 to −0.40 pp at 0.045–0.09 U). The scan is descriptive and does not reselect the operating point. Values at the preset (92.483% all windows, 78.829% SF10) agree within 0.03 pp with earlier evaluations of the same heads (92.473%, 78.799%); the difference comes from one-window flips between CPU and GPU recomputation. The 74.06% terminal silent-rail fraction at the preset (1 − 531/2,048) is reported as an event count, consistent with G3.

Reviewer_xZU2OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes SNN-FrFT, a spiking forecaster for chirp-like signals. It predicts 256 future I/Q samples from a 1024-sample history. A spiking order selector (A-SAS) estimates a fractional Fourier transform (FrFT) order from each history window. The FrFT, implemented as a pre-chirp, a radix-2 FFT and a post-chirp, uses this order to concentrate chirp energy. Transform coefficients are encoded with a dual-region logarithmic temporal code (DR-LTC). A compact LIF head projects once and reuses the resulting current over 16 steps to predict residual dynamics. The same order is then used to restore the quadratic phase over the forecast horizon. On a repartitioned NeLoRa benchmark, the method reports 79.28% SF10 spectral-peak agreement versus 73.67% for SpikF, with 97.23% fewer core MACs. It also reports a 38.3% reduction in head-level dynamic energy on a Lynxi HM100 board. Additional experiments include a synthetic-chirp factorization, a DR-LTC coding ablation, and ETTh1/ETTm1/Weather.

Critical strengths
  1. Physically grounded representation for chirps. Removing an estimated quadratic phase with the FrFT and restoring it over the forecast horizon is well motivated for chirp signals. The paper quantifies the concentration effect: history k90 falls from 29.46 to 3.83 on NeLoRa and from 257.4 to 12.3 on synthetic chirps. This is the main reason the idea is worth considering. It offers a principled, signal-aware alternative to learning rapidly changing phase in a fixed Fourier basis.
  2. Careful measurement and honest reporting. The hardware protocol is unusually detailed. It covers idle-power subtraction before and after each load, interleaved measurement orders, warm-up, and numerical-equivalence checks. The authors also do not overclaim. They report results that work against the method: larger peak-bin error under CFO, worse MAE than SpikF on ETT, failed HM100 equivalence checks, and a test split that is not historically untouched. This makes the reported numbers more trustworthy and easier to interpret.
Critical weaknesses
  1. The evidence does not show that the gains come from adaptive fractional representation or spiking computation.

Most of the efficiency gain, and part of the accuracy gain, comes from replacing SpikF with a much smaller head. The fixed-FFT variant with the same compact head already reaches 75.48% SF10 agreement, above SpikF’s 73.67%. It uses 655,616 core MACs, versus 658,112 for the full model. The 97.23% MAC reduction is therefore almost entirely due to the head, not to the FrFT or DR-LTC. The HM100 energy comparison likewise covers only the cached heads, with both frontends excluded. The need for input-adaptive order selection on LoRa is unclear. In LoRa CSS, the chirp slope is BW²/2^SF, which is fixed for a given spreading factor and sampling rate. The ideal order on NeLoRa should therefore depend only on SF, which takes three values here. The claim that “a fixed or globally learned order cannot follow changes between windows” is not tested against a per-SF fixed order, or against an order computed analytically from the phase-curvature estimate. Adaptive selection hurts on non-chirp data. On ETTh1, SNN-FrFT obtains MAE 0.430, compared with 0.390 for the fixed-FFT head, a degradation of about 10%. The paper frames this section as evidence of head efficiency, but the result also shows that A-SAS is harmful when no frequency sweep is present.

  1. The evaluation is too narrow to support the claims, and the significance of the task is not established.

SpikF is the only external baseline in the entire paper. All other comparisons are internal variants. There is no classical signal-processing baseline, such as estimating chirp parameters by dechirping and an FFT and then extrapolating. For signals that are close to pure chirps within a symbol, such a baseline could be very strong. There are also no lightweight learned forecasters such as DLinear, FITS or a small MLP, even though DLinear and FITS are cited. On NeLoRa, SpikF is already at 100% on SF8 and SF9, so the comparison effectively rests on 333 SF10 test windows. It is also not explained why forecasting LoRa baseband is a meaningful task, since NeLoRa-Bench was designed for demodulation. The paper does not explain how raw-FFT peak agreement before dechirping relates to downstream decoding.

Questions for authors
  1. How does a per-SF fixed order, or an order computed analytically from the phase-curvature estimate without a learned selector, compare with A-SAS on NeLoRa? If they perform similarly, what is the role of the spiking selector?
  2. How do classical chirp-parameter estimation followed by extrapolation, and lightweight forecasters (DLinear, FITS, a small MLP), perform on NeLoRa under the same protocol?
  3. Excluding the head change, how much of the improvement over SpikF is attributable to the FrFT and phase restoration? Can this decomposition be reported across seeds?
  4. Are NeLoRa windows aligned to symbol boundaries? Do any windows contain a symbol transition or a frequency wrap, and how does a single order handle them?
5 comments

We thank Reviewer xZU2 for recognizing the physically grounded chirp representation and the care of our measurements, and for the LoRa-specific questions, in particular how our forecasting metric relates to downstream decoding, which motivated the new real-data SF7 demodulation study (G1).

Q1. Per-SF fixed order and analytic order versus A-SAS, and the role of the spiking selector.

We ran the requested comparisons; the full table is G2. The reviewer's physical argument holds in our data. With the SF supplied, a per-SF fixed order reaches 81.181% (shared head) and 80.080% (SF-specific heads) on SF10, against 79.279% for A-SAS. On real SF7 recordings, the concentration-optimal order (median 1.0790 over the 8,109 training symbols, identical on validation) coincides with the analytic value from SF, BW and Fs (1.0792).

The spiking selector addresses the case where the SF is not given. There, A-SAS is the most accurate option measured: 8.3 pp above the direct phase-curvature estimate from the same history (70.971%), 7.3 pp above a single fixed order (71.972%), consistent with one order having to compromise between the SF8, SF9 and SF10 chirp rates, and 4.8 pp above a learned global order (74.474%). It matches exhaustive k90 search with one transform per window instead of ≈1,700, at 2,496 MACs (0.38% of the model). The revised introduction replaces the sentence that "a fixed or globally learned order cannot follow changes between windows" with this scoped statement, and Table 2 in Sec. 4.2 reports all order sources. Whenever the chirp configuration is known, a per-SF order is the natural choice.

Q2. Classical chirp estimation plus extrapolation, and lightweight forecasters, under the same protocol.

We added all four requested baselines with the same 1,024-sample history, 256-sample horizon, test windows and metrics as the main table. The classical baseline has no learned parameters: it estimates the chirp rate from the history's phase curvature, dechirps the history, locates the residual tone by FFT with sub-bin least-squares refinement and fits its complex amplitude, then extrapolates the chirp analytically. "Analytic + residual" adds a trained head on the residual. DLinear and FITS use their reference designs on I/Q.

ForecasterSF10 peak agreement (333 windows)I/Q MAE (all 1,015 windows)
Analytic chirp extrapolation (curvature + dechirp + FFT fit)57.357%0.006057
Analytic extrapolation + trained residual70.571%0.002620
DLinear36.537%0.004622
FITS34.034%0.007295
SpikF73.674%0.000944
SNN-FrFT79.279%0.004333

SNN-FrFT gives the highest SF10 peak agreement of all six forecasters, with 97.23% fewer learned multiply-adds than SpikF. The classical estimator reaches 57.4%, and 70.6% with a learned residual. The per-SF I/Q errors are in our reply to Reviewer akyk (Q1). The revision adds this table as Table 12 (Appendix D), referenced from Sec. 4.1.

(continued)

Q3 and W1. How much of the gain over SpikF comes from the FrFT and phase restoration, with the head held fixed.

The two claims have different sources, and the revision states them separately (Sec. 4.2 and Appendix C). The 97.23% reduction in learned MACs comes from the compact execution architecture: 61.0% of it from the project-once LIF head and 39.0% from removing SpikF's non-head matrices. The FrFT itself adds only 2,496 selector MACs and O(N log N) fixed arithmetic. The accuracy gain over SpikF comes from both the head and the representation:

Step (NeLoRa SF10, 333 windows)SF10 peak agreementGain
SpikF73.674%
Fixed FFT + compact LIF head (head swap only)75.475%+1.80 pp (head)
+ adaptive FrFT input + matched phase restoration (SNN-FrFT)79.279%+3.80 pp (representation)

Retraining the fixed-FFT and SNN-FrFT heads from 10 independent paired initializations gives a representation gain of +3.06 pp (sample SD 2.08 pp; paired t(9)=4.67, two-sided p=0.0012), positive in 9 of 10 pairs and zero in one:

Seed100101102103104105106107108109
SNN-FrFT − fixed FFT (pp)+3.00+2.10+6.310.00+1.50+4.50+5.11+5.11+1.20+1.80

On real demodulation (G1), where the classifier is also held fixed and only the frontend changes, the FrFT frontend beats SpikF's spiking backbone at all 8 SNRs, by +10.79 pp at −22 dB and +43.90 pp at 0 dB (92.93% versus 49.03%). Sec. 4.2 of the revision now reports this 10-seed paired comparison.

Remaining points of W1. On the frontends excluded from the HM100 measurement, the full-pipeline count (G3), which includes the fixed FrFT arithmetic, still gives a 97.09% reduction in multiply-adds and 6,144 versus 149,568 membrane updates. On ETTh1, per-window adaptation is designed for frequency sweeps; with the order fixed at 0.981 and a width-1024 compact head, ETTh1 MAE is 0.02% lower than SpikF's at 98.47% fewer learned MACs.

Q4. Symbol alignment, transitions and frequency wraps in the NeLoRa windows, and how a single order handles a wrap.

We audited every window from its file and offset metadata:

ItemValue
Windows (train / val / test)8,192 / 1,033 / 1,015 (10,240)
Source symbols1,551, from symbol positions 0–7 of each packet (the preamble)
Symbol value of every window0 (unmodulated preamble up-chirp)
Windows crossing a nominal symbol boundary0
Windows containing a nominal frequency wrap0
Window startsoffsets k·128 samples inside one symbol; 622 start at the symbol start, 594 end exactly at the symbol end
History / horizon1,024 / 256 samples, always within a single symbol

The windows slide in 128-sample steps inside one synchronized preamble up-chirp, so none contains a symbol transition or a nominal wrap (the released dataset defines wraps in symbol coordinates). The revision adds this audit as Table 14 (Appendix D), and Sec. 4.1 now describes the forecasting benchmark as within-symbol continuation of preamble chirps.

Wraps are covered by the new demodulation study (G1), which uses all payload symbols, values 0–127. A wrap is a frequency jump at a constant chirp rate, so the order matched to the chirp rate also matches both segments, and the FrFT maps a wrapped symbol to two peaks exactly 2^SF bins apart; Figure 7 (Appendix D) shows symbol 40 with peaks at bins 552 and 424. Since every nonzero LoRa symbol contains exactly one wrap, 127 of the 128 test classes are wrapped, and the spiking classifier with A-SAS orders decodes 92.93% of all test symbols at 0 dB (SpikF with the same head: 49.03%).

(continued)

W2. Significance of the task, the narrow SF10 comparison, and the link to downstream decoding.

NeLoRa-Bench was designed for demodulation, and raw-FFT peak agreement before dechirping does not by itself show a decoding benefit. We therefore evaluated SNN-FrFT on demodulation itself; the full setup, per-SNR table and figure are in G1. With the same spiking classification head, SNN-FrFT decodes real held-out LoRa symbols more accurately than SpikF at every SNR, 92.93% versus 49.03% at 0 dB and 73.38% versus 31.07% at −14 dB, at 3.4% of SpikF's multiply-adds. The revision adds this experiment as Appendix E, referenced from Sec. 4.2, describes the forecasting benchmark as within-symbol chirp continuation (Sec. 4.1), and states the SF8/SF9 ceiling in Sec. 4.2. The SF10 comparison is also supported across 10 paired seeds (+3.06 pp over fixed FFT, p=0.0012; see Q3).

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

This paper proposes SNN-FrFT, a spiking forecasting architecture that couples a history-adaptive fractional Fourier transform (FrFT) representation with matched future-phase reconstruction. A lightweight spiking order selector (A-SAS) estimates the fractional order α from the observed history, while a compact forecasting head predicts residual dynamics in the fractional domain and restores future phase using the same α. A new Signed Dual-Region Log-Temporal Coding (DR-LTC) encodes transform coefficients with a coarse logarithmic and a fine near-zero region plus a terminal magnitude gate to induce sparsity. On NeLoRa chirp baseband, SNN-FrFT improves spectral peak localization (notably at SF10) with large reductions in core matrix MACs versus SpikF, and shows measured board-level energy savings on HM100 for the cached head.

Critical strengths

1. Well-motivated technical design.

The paper combines adaptive FrFT representation with matched future-phase reconstruction, allowing the model to better handle chirp-like signals. The proposed DR-LTC coding also provides a reasonable balance between near-zero precision and spike sparsity. The component ablations help explain why these designs are effective.

2. Solid experimental and hardware validation.

The paper provides useful ablations on synthetic chirps and evaluations on real NeLoRa signals. The improvement on SF10 is encouraging, and the substantial reduction in core MACs demonstrates the efficiency of the proposed head. The HM100 measurements further support its practical value by showing lower energy consumption and higher throughput for the cached forecasting head.

Critical weaknesses

1. Limited end-to-end efficiency evaluation.

The reported 97.23% MAC reduction excludes the fixed transform and coding costs, while the HM100 energy measurements only cover the cached forecasting heads. Although these results are useful, it remains unclear how much efficiency improvement can be achieved by the complete pipeline. A full-pipeline comparison with SpikF would make the practical benefits more convincing.

2. Limited comparisons for adaptive order selection.

The paper demonstrates the benefit of combining adaptive FrFT with future-phase reconstruction, but the effectiveness of the A-SAS selector itself is less thoroughly evaluated. Comparisons with alternative order-selection strategies, such as grid-search or learned FrFT orders, would help clarify how much of the improvement comes from the proposed selector rather than the representation and reconstruction design.

Questions for authors
  1. End-to-end efficiency: Could the authors clarify the overall computational cost when including the FrFT, temporal coding, and membrane updates? The reported MAC reduction and HM100 energy measurements focus on different parts of the pipeline, so an end-to-end comparison with SpikF would help assess the practical efficiency gains.

  2. Adaptive order selection: How does A-SAS compare with alternative order-selection strategies, such as grid-searched or learned α? Such a comparison would help determine how much of the improvement comes from the proposed selector itself.

  3. Hardware consistency: The paper reports that the compiled HM100 heads do not achieve numerical equivalence with their software counterparts. Could the authors clarify the main causes of these discrepancies and whether they affect forecasting accuracy on the deployed hardware?

1 comment

We thank Reviewer vkbu for the careful reading and for asking us to account for the full pipeline and the hardware results; the complete operation count and task-level hardware checks below make our efficiency claims more precise.

Q1. End-to-end efficiency, including the FrFT, temporal coding and membrane updates.

The full operation count is in G3 of our general response. Including the FrFT and phase arithmetic, SNN-FrFT needs ≈691,904 multiply-adds per window against 23,756,800 for SpikF (−97.09%, versus −97.23% for learned matrices alone), and 6,144 membrane updates against 149,568. The O(N log N) radix-2 FrFT contributes 8,448 complex multiplications. The same holds on real-data demodulation (G1): the complete classifier, selector and FrFT included, costs ≈326,336 multiply-adds per symbol against 9,568,256 for SpikF with the same head (−96.6%), and it is more accurate at every tested SNR (92.93% versus 49.03% at 0 dB).

The board energy figure (0.1140 versus 0.0703 mJ/window at batch 32, −38.3%) is measured on the cached heads, because the frontend currently runs as structured dense arithmetic on the host. We report the rest of the pipeline as operation counts: the revision adds the G3 table as Appendix C and cites the full-pipeline count in Sec. 4.1.

Q2. A-SAS versus grid-searched and learned orders.

G2 compares A-SAS with six alternatives on the same SF10 test windows. Without SF metadata, A-SAS (79.279%) is 7.3 pp above a single grid-searched order (71.972%) and 4.8 pp above an end-to-end learned global order (74.474%), and it matches the exhaustive history-only k90 search (78.979%) with 2,496 MACs and one transform instead of ≈1,700. With SF metadata, a per-SF lookup reaches 80.080–81.181%.

Per-window selection therefore pays off where the chirp rate varies across windows and the SF is not given (mixed SF8–10), reaching the accuracy of an expensive search at a small fraction of its cost. The revised introduction states this setting precisely in place of the sentence that "a fixed or globally learned order cannot follow changes between windows", and the G2 table is Table 2 in Sec. 4.2.

Q3. Why the compiled HM100 heads are not numerically equivalent, and whether accuracy is affected.

The board matches the software reference at the task level for all three compiled heads. On the 768 SF10 windows of the hardware population, SF10 peak agreement changes by exactly 0.000 pp and I/Q MAE by less than 1e-6:

HeadBoard SF10 peak %Board − PC (pp)Pointwise max absNRMSE %PC → board I/Q MAE
SpikF68.6200.0001.42e-40.0810.0024439 → 0.0024431
SNN-FrFT w12879.9480.0001.06e-30.3820.0095727 → 0.0095725
SNN-FrFT w25678.6460.0008.36e-40.3460.0079881 → 0.0079877

The strict pointwise check fails because the compiler maps the FP32 graph to the accelerator's native precision, and small arithmetic differences flip LIF neurons whose membrane potential sits near threshold: 0.147% (w128) and 0.155% (w256) of fc1+LIF neuron-steps. The archived compiler artifacts do not export the internal quantizer mapping, so we characterize the residual by where it appears. As a PC-side sensitivity check, symmetric int16 quantization of weights, and of weights plus activations, moves all-window peak agreement by at most 0.49 pp for SNN-FrFT and 0.39 pp for SpikF over 3 seeds. The revision reports this table in Appendix G (Table 17), cited from Sec. 4.6, as task-level equivalence on the historical hardware population.

Comments on the paper

3 comments

We thank the four reviewers for detailed and technically specific reviews. Several questions overlap, so we first report the three new results that answer them together. Each reviewer-specific reply is self-contained and points back here only for the full tables. We have uploaded a revised PDF; the new material is in Appendices C to H, and the figures mentioned in our replies are there.

The core of SNN-FrFT is unchanged: a fractional-Fourier representation whose order matches the chirp rate, read by a compact project-once LIF head. The new experiments test it on the task NeLoRa-Bench was built for (symbol demodulation), against every order-selection strategy the reviewers proposed, and with every operation of the pipeline counted.

G1. Real-data symbol demodulation on held-out NeLoRa SF7 packets (xZU2 W2, akyk Q1/Q6, vkbu Q1, yvlr W2)

This experiment tests the adaptive fractional representation, the core of SNN-FrFT, on a second task. The frontend (A-SAS, FrFT and DR-LTC coding) is the same as in forecasting and is trained without symbol labels; only the head is replaced. We replaced the forecasting head with a spiking classification head, Linear(d→256), 16 LIF steps, Linear(256→128), and decoded all SF7 symbols of NeLoRa-Bench (128 classes, 1,024 samples per symbol, no resampling). The 13,714 symbols that pass the official clean dechirp check are split by packet into 110 / 37 / 37 packets for train / validation / test; the 2,804 test symbols cover all 128 classes. Complex AWGN follows the official NeLoRa procedure and SNR definition, training SNR is drawn from U[−30, 0] dB, and every method sees identical noisy test samples.

SNN-FrFT feeds the head the magnitude of its DR-LTC-coded FrFT coefficients, with the order chosen per window by A-SAS (retrained on SF7 training histories without symbol labels). SpikF feeds the same head with its original spiking backbone, spike features averaged over 16 steps, trained jointly with the head. The learning rate is chosen on validation from {1e-3, 5e-4, 2e-4}, with 3 seeds and at most 30 epochs. Both configurations were locked on validation before the test packets were read, and the test packets were evaluated once.

Test symbol accuracy (%), mean ± SD over 3 seeds:

Method−22 dB−18 dB−16 dB−14 dB−12 dB−10 dB−5 dB0 dB
SNN-FrFT (A-SAS) + spiking head24.44±1.1947.19±0.9364.03±1.4073.38±1.6779.78±1.5282.93±2.0490.31±1.9092.93±2.15
SpikF + same spiking head13.65±0.2019.46±1.3525.12±1.4931.07±2.0036.21±2.0740.41±1.8647.67±3.5249.03±4.01

With the classifier held fixed, the FrFT frontend is more accurate at every SNR. The margin is +10.79 pp at −22 dB and between +38.9 and +43.9 pp from −16 dB to 0 dB, where SpikF features stay below 50%. Figure 8 (Appendix E of the revised PDF) plots both curves. The complete SNN-FrFT classifier costs ≈326,336 multiply-adds per symbol (head 295,168, selector 2,496, FrFT 7,168 complex multiplications), against 9,568,256 for SpikF with the same head (−96.6%).

G2. A-SAS against every proposed order-selection strategy (vkbu Q2, xZU2 Q1, akyk Q2)

All rows use the same NeLoRa SF10 test windows (333 windows, 3 head seeds) and a retrained head of the same architecture.

Order sourceSF metadataOrder cost per windowSF10 peak agreement
Direct phase-curvature (chirp-rate) estimate, analytic ordernoclosed form70.971%
Global fixed order, grid-searched on training datano071.972%
Learned global scalar (STE, end-to-end)no074.474%
History-only k90 grid searchno1,699.6 FrFT evaluations78.979%
A-SAS (spiking selector)no2,496 MACs79.279%
Per-SF lookup, SF-specific headsyeslookup80.080%
Per-SF lookup, shared headyeslookup81.181%

Without SF metadata, A-SAS is the most accurate strategy. It is 8.3 pp above the closed-form curvature estimate from the same history, 7.3 pp above a single grid-searched order, and 4.8 pp above an end-to-end learned global order. It matches exhaustive k90 search with one transform per window instead of ≈1,700. When the SF is supplied as metadata, a per-SF lookup gives 80.1–81.2% and is the natural choice for that setting. A-SAS fills the complementary role of SF-agnostic order estimation at 0.38% of the model's MACs. This table is Table 2 in Sec. 4.2 of the revision.

(continued)

G3. Cost of the complete forecasting pipeline (vkbu Q1, akyk Q3)

One NeLoRa forecasting window (1,024-sample history, 256-sample horizon):

Cost per windowSpikFSNN-FrFT
Learned matrix MACs (head + selector)23,756,800658,112
Fixed spectral arithmetic (pre/post chirp, radix-2 FFT, output-phase removal, future-phase restoration)none8,448 complex mult. (≈33,792 real mult.)
Total multiply-adds (real-multiplication equivalent)23,756,800≈691,904 (−97.09%)
Membrane updates (neurons × time steps)149,5686,144 (head 256×16 + selector 64×32)
Temporal-code encode/decode (DR-LTC, 13 stages)none26,624 rail codings; 21,349 events, 531 terminal
Parameters471,613658,977

Counting the fixed transform moves the reduction from 97.23% to 97.09%. The radix-2 FrFT is O(N log N) and needs 8,448 complex multiplications, against about 1.05M for a dense N-point transform. SpikF's membrane updates are measured with forward hooks on its LIF nodes; ours follow directly from the head and selector dimensions. This table is Table 10 (Appendix C) of the revision.

Area_Chair_tCArOpenReview ↗

reminder

Dear reviewers,

This is a gentle reminder to read the authors’ rebuttals and respond in the discussion.

Thank you for your time and effort throughout the review process, and for your contributions to the community.

Best Regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.