acceptodds
Under review as a conference paper at ICLR 2027

RFspkSSM: Spiking State Space Models for Efficient Radar Target Recognition

Abstract

Spiking neural networks and state-space models offer complementary approaches to efficient sequence processing, combining sparse event-driven computation with scalable long-range modelling without quadratic dependence on sequence length. Existing spiking state-space models focus largely on generic sequence modelling, with limited exploration in radar sensing and signal processing. We propose radar-oriented spiking state-space models (RFspkSSM), built around a second-order oscillatory spiking state-space model (OscSpikeSSM) that combines trainable neuronal parameterisations with stability-constrained oscillatory dynamics. For radar sensing, the core extends to structured point clouds, multi-view heatmaps, and radar-IMU representations spanning applications such as activity and gesture recognition, 3D human and hand pose estimation, and raw FMCW signal processing. Across 10 benchmark datasets, RFspkSSM efficiently outperform baselines. More specifically, RFspkSSM attains 96.8% accuracy in radar gesture recognition, reduces pose centre-distance error to 9.2 cm ( lower), uses up to fewer parameters for mmWave recognition, and achieves up to lower theoretical compute energy in radar signal processing. Our results show that RFspkSSM generalises across heterogeneous sequence and radar representations, combining strong task performance with efficient modelling that learns spectral processing from sequence data.

Primary area:
applications to neuroscience & cognitive science
Keywords:
SSM, Radar, Spiking Neurons, Gesture Recognition, Activity Recognition, Pose Estimation, Signal Processing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_kcY9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper explores spiking state-space models (SSMs) for radar signal analysis. This is a meaningful direction, as the temporal modeling capabilities of spiking neural networks (SNNs) remain underexplored.

Critical strengths

None

Critical weaknesses
  1. I could not tell how the proposed spiking SSM differs from existing spiking SSMs. The current description gives too little technical detail to understand its operation. A diagram explaining the computation would also help.

  2. RF neurons already exhibit resonance and have been used in sequence modeling. What does the proposed resonant SSM add beyond RF-based models? The paper needs to explain this difference and include a comparison with these models.

  3. The main method takes up less than one page, leaving several basic questions unanswered. Which parts of the design are new, what benefits do they bring, and which operations are spike-driven? I am also unsure what the radar input is: does the model receive raw waveforms, or are they first converted into image representations?

  4. Tables 1–3 extend beyond the page margins, and Figure 2 does not distinguish the proposed components from the existing ones. The use of SiLU also needs explanation. Does it introduce non-spiking computation? If so, how does this affect the computational savings expected from the SNN?

  5. Most of the SNN baselines are older models. Comparisons with recent spiking Transformers are missing, as are comparisons with existing spiking SSMs. Since the paper presents a spiking SSM, the latter comparison is needed to assess what the proposed design adds.

Questions for authors

See weakness

8 comments
AuthorsOpenReview ↗

Response to Reviewer KcY9

We would like to thank you for your feedback. Your comments have considerably helped with improving the clarity of the revised manuscript.

Response to W1 and Q1: Technical Formulation and Working Mechanism: We clarify that the proposed OscSpikeSSM combines the second-order oscillatory state-space dynamics of LinOSS with the trainable neuronal parameterisation of SiLIF/AdLIF, as described in Section 3.1. Unlike conventional first-order spiking neurons, OscSpikeSSM maintains coupled position-like and velocity-like states, allowing temporal information to evolve through damped oscillations. Unlike non-spiking oscillatory SSMs, it incorporates threshold-based binary spike communication, spike-dependent feedback, and subtractive reset. Specifically, the state evolution follows

F_t &= -k u_{t-1} + B I_t - R s_{t-1},\\\\ v_t &= \frac{v_{t-1}+\Delta t F_t} {1+\Delta t c},\\\\ u_t &= u_{t-1}-r_{\mathrm{reset}}s_{t-1} +\Delta t v_t,\\\\ s_t &= H(u_t-\theta). \end{aligned} \tag{6}$$ Here, the learnable oscillator coefficients control damping, restoring dynamics, input coupling, and spike-dependent feedback. The underdamping constraint preserves oscillatory continuous-time modes, while the damped-IMEX discretisation and its discrete-time stability conditions are detailed in Appendix A.1.2. Spike-dependent feedback modifies the subsequent velocity evolution, whereas subtractive reset regulates firing without directly resetting the velocity-like state. These mechanisms allow oscillatory state evolution to continue across spike events. The working mechanism is illustrated through the phase-space trajectories in Figure 1 and the comparison with SiLIF and AdLIF in Figure 3 of Appendix A.1.1. To further establish their contributions, we conducted targeted ablations , reported in Table 5 of this rebuttal. These ablation results are also presented and discussed in our response to Reviewer HPF1 (Q2). Removing the underdamping constraint reduces accuracy from 92.14% to 91.43%, while removing spike-dependent feedback and subtractive reset reduces accuracy to 90.98% and 90.27%, respectively. The largest degradation occurs without subtractive reset, supporting its role in regulating spike activity while retaining oscillatory state evolution. Furthermore, the architecture-matched comparisons with LinOSS and P-SpikeSSM achieve average accuracies of 85.64% and 82.94%, respectively, compared with 89.24% for OscSpikeSSM across radar datasets. These comparisons retain the surrounding architecture and training protocol with comparable model capacities, providing evidence for the contribution of the proposed temporal dynamics. The complete results are provided in Table 4 of our response to Reviewer HPF1 (Q1). We also evaluate Raw OscSpikeSSM on MMVR, where the centre-distance error decreases from 38.28 cm for the raw model to 9.20 cm for the proposed radar-specific architecture. This comparison demonstrates the additional benefit of preserving radar geometry through multidirectional scans and depth-aware cross-view fusion, rather than attributing the improvement solely to the temporal core. These results are reported in Table 2 of our response to Reviewer WzBt (Q3). Additionally, replacing OscSpikeSSM with a conventional LIF core increases the MMVR centre-distance error from 9.20 cm to 11.35 cm, as discussed in Section 4.3. Together, the component ablations and matched temporal-core comparisons provide complementary evidence supporting the proposed design. **Response to Q2:** We thank the reviewer for raising this distinction. Our primary focus is on extending LIF-family spiking models to include second-order oscillatory dynamics, which do not currently exist. Unlike conventional resonate-and-fire (R&F) neurons, OscSpikeSSM combines trainable SiLIF/AdLIF parameterisation with stability-constrained oscillations, spike-dependent feedback, and subtractive reset (Section 3.1). We acknowledge existing R&F-based models and their potential hardware limitations in Appendix A.2.2, which is also provided below - Although recent resonate-and-fire (R&F) neuron-based SSMs have achieved strong performance on long-range classification benchmarks (Agrawal et al., 2026), their hardware realisation can require larger circuit footprints and silicon area than simpler LIF-family neurons (Le et al., 2023; Auge & Mueller, 2020). This can reduce the achievable neuron density on area-constrained hardware, limiting the number of neurons that can be implemented within a given chip area. This hardware gap motivates the development of oscillatory spiking dynamics that retain the simplicity of LIF-family neuron implementations while improving their ability to model long temporal dependencies. The use of LIF neurons also potentially makes the model amenable for implementation in analog neuromorphic chips such as Pulsar (Innatera, 2026; Richter et al., 2024).
AuthorsOpenReview ↗

Response to Reviewer KcY9

Response to Q3: In this paper, in addition to proposing an oscillatory spiking SSM, we also propose multiple radar-application-specific architectures, unlike previous works (Bal & Sengupta, 2025; Shen et al., 2025; Fabre et al., 2025). The dataset-specific input representations and preprocessing requirements are summarised in Table 6, distinguishing preprocessed radar inputs from the raw FMCW signals used by RF-SpecSpkSSM. Further implementation details are provided in Appendices A.2.3, A.2.4, A.2.5, A.2.6 and A.2.7. We have included the following paragraph in the main text of the revised manuscript to clarify the radar input representations and preprocessing requirements.

Radar Input Representations: The RFspkSSM models demonstrate robust performance across a wide range of radar representations depending on the task and/or the availability of publicly available data formats. More specifically, AOPHand, mmFiT, and MMActivity use radar-derived point clouds, Google Soli uses preprocessed range-Doppler frames, MMVR uses dual-view radar heatmaps, and EgoHands uses synchronised radar and IMU sequences. In contrast, RF-SpecSpkSSM processes raw FMCW ADC samples before conventional range-Doppler transformation. Only basic signal conditioning is retained, while the range and Doppler FFTs are replaced by the proposed RF-SpecSpkSSM spectral front end, as illustrated in Figure 4 of Appendix A.2.7. The dataset-specific input representations and preprocessing requirements are further detailed in Table 11 of the Appendix.

Table 6: Radar input representations used across the RFspkSSM models. Input shapes exclude the batch dimension.

DatasetModel InputInput Preparation
AOPHandPoint cloud, 4×2564\times256Subsampling/padding to 64 points, training-set normalisation, and 16×1616\times16 reshaping.
mmFiTPoint cloud, 4×2564\times256Subsampling/padding to 64 points, training-set normalisation, and 4×8×84\times8\times8 reshaping.
MMActivityPoint cloud, 4×2564\times256Subsampling/padding to 64 points, native feature scale, and 16×1616\times16 reshaping.
Google SoliRange-Doppler, 40×102440\times1024Single-channel selection and frame-wise flattening.
MMVRDual-view heatmaps, 4×256×1284\times256\times128 per viewPrecomputed, temporally synchronised horizontal-depth and vertical-depth heatmaps.
EgoHandsRadar–IMU sequences, T=30T=30Temporally synchronised radar and IMU sequences.
Aircraft MarshallingRaw FMCW ADC, 960×512960\times512Fast- and slow-time DC removal, Hann windowing, and segmentation into 17 overlapping windows (192 chirps, hop 48).

Response to Q4: We thank the reviewer for raising this point. We have followed the simplified operation-based analyses of Bal & Sengupta (2025); Liao et al. (2026), the energy model estimates computational cost using MACs and spike-driven ACs. Another goal of this work is also to drive future hardware designs. In RFspkSSM, binary spikes enable accumulation-based computation, whereas operations on real-valued features, including convolutions and linear projections, require dense MACs. In particular, SiLU and GELU are conventional nonlinear activations and do not constitute spike-driven computation. Although SiLU introduces non-spiking computation, its element-wise complexity is O(LN)O(LN), which is lower than the O(LN2)O(LN^2) complexity of dense linear layers, consistent with the analysis of nonlinear activation costs in P-SpikeSSM (Bal & Sengupta, 2025). Compared with GELU, which involves evaluating the Gaussian cumulative distribution function, SiLU uses a simpler sigmoid-based formulation, potentially reducing activation overhead.

Our theoretical energy analysis follows the operation-based MAC/AC methodology adopted in prior SNN and hybrid ANN–SNN architectures (Horowitz, 2014; Bal & Sengupta, 2025; Liao et al., 2026; Panda et al., 2020; Aydin et al., 2024; Lee et al., 2026). Our hybrid ANN-SNN design is further motivated by emerging neuromorphic platforms such as Innatera Pulsar. The use of LIF neurons potentially makes the model amenable for implementation in analog neuromorphic chips (Innatera, 2026; Richter et al., 2024).

AuthorsOpenReview ↗

Response to Reviewer KcY9

Response to Q5: We thank the reviewer for highlighting these presentation issues. The principal contributions in Figure 2 are the OscSpikeSSM-based RF-SpikingOSS blocks, multidirectional spiking scans along the radar axes, depth-aware cross-view fusion, and the proposed query-based pose prediction and refinement architecture, as detailed in Section 4.3 and Appendix A.2.5. Our additional experiments further establish these contributions. In architecture-matched comparisons, OscSpikeSSM achieves 89.24% average accuracy across four radar benchmarks, outperforming P-SpikeSSM (82.94%), as reported in Table 4 of our response to Reviewer HPF1 (Q1). Furthermore, applying Raw OscSpikeSSM directly to MMVR yields a centre-distance error of 38.28 cm, compared with 9.20 cm for our proposed MMVR-OscSpikeSSM architecture (Table 2 of our response to Reviewer WzBt (Q3)). This substantial improvement demonstrates the importance of adapting OscSpikeSSM to the spatial and geometric structure of radar data. Thus, the proposed radar-specific architecture, designed around OscSpikeSSM constitutes an additional contribution of our work. We have revised Figure 2 caption to explicitly distinguish the proposed components from standard operations. We have also reformatted Tables 1-3 to fit within the page margins, with clearer headings, spacing, and consistent notation.

Response to Q6: As you know, one of the goals of this work is to advance the field of spiking state-space models (SSMs). Recent research has shown that SSMs are computationally simpler and can help overcome the quadratic complexity of transformers. Unlike conventional Transformers with O(L2)\mathcal{O}(L^2) self-attention complexity, efficient SSMs offer O(L)\mathcal{O}(L) sequence scaling through structured state updates (Gu et al., 2022; Gu & Dao, 2024), making them attractive for long-range temporal modelling in resource-constrained radar sensing. Moreover, given the biological closeness of these SNNs, this work could also help bridge the gap between deep learning and neuroscience.

We provide architecture-matched comparisons with LinOSS and P-SpikeSSM in Table 4 of the revised manuscript, in which only the temporal core is replaced, while the surrounding architecture and training protocol are retained. OscSpikeSSM achieves 89.24% average accuracy across four radar benchmarks, compared with 85.64% for LinOSS and 82.94% for P-SpikeSSM, under comparable model capacities. Furthermore, we compare against recent spiking Transformers, including SpikFormer and SpikingResFormer, using the benchmarks reported by Liao et al. (2026). As shown in Table 1 in Section 4.2, mm-OscSpikeSSM achieves an average accuracy of 86.72% across AOPHand, mmFiT, and MMActivity, compared with 83.57% for SpikFormer and 81.65% for SpikingResFormer, while using substantially fewer parameters. Spiking Transformers generally aim to approach ANN Transformer performance while improving computational efficiency through spike-driven processing. We also compare against RAPTR (Kato et al., 2025), a recent Transformer-based architecture (NeurIPS 2025), with no subsequent directly comparable MMVR study. Our MMVR-OscSpikeSSM achieves 3.41×3.41\times lower centre-distance error, 6.37×6.37\times lower model memory, and 1.61×1.61\times lower theoretical compute energy (Tables 5-6, Section 4.3).

Reviewer_HPF1OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes RFspkSSM, a family of radar-oriented spiking state-space models built around a second-order oscillatory spiking core, OscSpikeSSM. The authors develop task-specific architectures for diverse radar perception tasks and introduce an oscillatory spectral-processing front end to replace conventional FFT operations. Experiments across ten datasets demonstrate competitive performance and task-dependent efficiency trade-offs.

Critical strengths

The paper adapts its proposed OscSpikeSSM core to point-cloud recognition, multi-view pose estimation, radar–IMU hand pose estimation, and radar gesture recognition, alongside a related spectral-processing formulation for raw FMCW signals. This broad evaluation demonstrates the feasibility of applying oscillatory-spiking dynamics to diverse radar perception settings and provides potentially useful task-specific architectural designs.

Critical weaknesses
  1. Insufficient controlled comparisons to establish the advantages of OscSpikeSSM. The radar benchmarks primarily compare complete architectures with different encoders, backbones, fusion mechanisms, or decoders, making it difficult to attribute the reported improvements to OscSpikeSSM itself. While LinOSS and several spiking SSMs are included in the general long-sequence experiments (Table 9), comparable evaluations with matched architectures and model capacities are lacking on radar tasks. The substantial parameter reductions relative to larger baselines therefore do not establish the efficiency advantages of the proposed dynamics. These limitations weaken the evidence for OscSpikeSSM's incremental contribution over existing oscillatory and spiking SSMs.

  2. Insufficient validation of the proposed oscillatory mechanisms and radar-specific inductive bias. The MMVR LIF replacement provides useful evidence, but does not isolate the effects of the underdamping constraint, spike-dependent feedback, and subtractive reset. The proposed mechanisms are not systematically ablated against existing second-order SSM formulations. Moreover, the radar-specific motivation lacks direct evidence that the learned oscillatory dynamics capture task-relevant spectral characteristics. The four-frame point-cloud experiments provide limited evidence for long-range temporal modeling, while the illustrative spike dynamics do not establish a general sparsity advantage. Consequently, the necessity and effectiveness of the proposed mechanisms remain insufficiently demonstrated.

Questions for authors
  1. Please establish the contribution of OscSpikeSSM through controlled comparisons. The comparisons with LinOSS and existing spiking SSMs are primarily conducted on non-radar long-sequence benchmarks (Table 9), while the radar experiments compare architectures that differ in multiple components beyond the recurrent core. Could the authors provide controlled comparisons on representative radar tasks, replacing only OscSpikeSSM with existing oscillatory or spiking SSMs (e.g., LinOSS and P-SpikeSSM) while keeping the surrounding architecture, parameter/state budgets, and training protocol comparable? This would help establish whether the reported accuracy and efficiency improvements arise from the proposed dynamics rather than task-specific architectural choices.

  2. Please validate the proposed mechanisms and radar-specific inductive bias. The MMVR LIF replacement supports the usefulness of the overall oscillatory core but does not isolate the contributions of the enforced underdamping constraint, spike-dependent feedback, and subtractive reset. Could the authors provide targeted ablations distinguishing these mechanisms from existing second-order SSM dynamics? Furthermore, although the oscillatory parameters are learnable, what evidence shows that the resulting frequency responses actually adapt to radar-specific temporal or spectral characteristics? Analyses of learned frequencies or controlled frequency interventions would help substantiate the proposed inductive bias.

5 comments
AuthorsOpenReview ↗

Response to Reviewer HPF1

We would like to thank you for you positive feedback. Your detailed comments have considerably helped with improving the clarity of the revised manuscript.

Response to W1 and Q1:Controlled Comparisons of OscSpikeSSM. To isolate the contribution of OscSpikeSSM from task-specific architectural choices, we conducted additional architecture-matched comparisons with LinOSS (Rusch & Rus, 2025) and P-SpikeSSM (Bal & Sengupta, 2025) across radar benchmarks: AOPHand, mmFiT, MMActivity, and Google Soli, as reported in Table 4. These baselines represent non-spiking oscillatory SSMs and existing spiking SSMs, respectively. In these experiments, only the OscSpikeSSM temporal core is replaced, while retaining the same surrounding architecture, input representations, classification heads, dataset splits, and training protocols, with comparable parameter and state budgets. OscSpikeSSM achieves the highest accuracy on all four datasets, averaging 89.24%, compared with 85.64% for LinOSS and 82.94% for P-SpikeSSM. These improvements of 3.60 and 6.30 percentage points, respectively, provide direct evidence that the proposed oscillatory spiking dynamics contribute independently of the surrounding architecture. For the separate system-level efficiency analysis, we use SpikingResFormer (Liao et al., 2026) as a strong published spiking baseline, which achieves 81.65% average accuracy across AOPHand, mmFiT, and MMActivity, compared with 78.46% for our matched P-SpikeSSM implementation. Thus, the matched-core experiments establish the contribution of the proposed dynamics, while the published baseline comparisons assess the performance and efficiency of the complete architecture (Tables 1-2, Section 4.2).

Table 4: Comparison of raw and architecture-matched spiking state-space models on AOPHand, mmFiT, MMActivity, and Google Soli. Accuracy is reported as mean ±\pm standard deviation across training runs, and model size is measured in millions of parameters.

ModelAOPHand Acc. (%)AOPHand Params (M)mmFiT Acc. (%)mmFiT Params (M)MMActivity Acc. (%)MMActivity Params (M)Google Soli Acc. (%)Google Soli Params (M)Avg. (%)
LinOSS (matched)90.75±0.8590.75_{\pm0.85}0.19889.56±0.3489.56_{\pm0.34}0.26268.75±4.4268.75_{\pm4.42}0.14093.5±0.693.5_{\pm0.6}0.39885.64
P-SpikeSSM (matched)75.68±2.2775.68_{\pm2.27}0.20589.69±0.6389.69_{\pm0.63}0.26570.00±1.7170.00_{\pm1.71}0.14296.4±0.496.4_{\pm0.4}0.50082.94
mm-OscSpikeSSM92.14±0.3292.14_{\pm0.32}0.18290.02±0.5290.02_{\pm0.52}0.25578.00±1.8778.00_{\pm1.87}0.13896.8±0.396.8_{\pm0.3}0.33489.24
AuthorsOpenReview ↗

Response to Reviewer HPF1

Response to W2 and Q2: To isolate the contributions of the proposed mechanisms, we have conducted targeted ablations on AOPHand, as reported in Table 5 of this rebuttal and Table 14 of the revised manuscript. Removing the underdamping constraint reduces accuracy from 92.14% to 91.43%, while removing spike-dependent feedback and subtractive reset reduces accuracy to 90.98% and 90.27%, respectively. All variants retain the same surrounding architecture and training protocol. These results provide direct evidence for the individual contributions of the proposed mechanisms, beyond the MMVR LIF replacement. Furthermore, to distinguish OscSpikeSSM from existing second-order and spiking SSM formulations, we provide architecture-matched comparisons with LinOSS and P-SpikeSSM on AOPHand in Table 4 of the revised manuscript. Together, these comparisons support the effectiveness of the proposed oscillatory spiking dynamics and their constituent mechanisms. Furthermore, to establish the importance of our radar-specific architectural design, we compare Raw OscSpikeSSM with MMVR-OscSpikeSSM. Raw OscSpikeSSM yields a centre-distance error of 38.28 cm, compared with 9.20 cm for MMVR-OscSpikeSSM. This demonstrates the importance of our radar-specific architecture, incorporating multidirectional spiking scans, depth-aware cross-view fusion, and query-based pose refinement, as an additional contribution beyond the temporal core (Table 2, response to Reviewer WzBt, Q3).

We acknowledge that the four-frame radar experiments alone do not establish long-range temporal modelling. However, we have separately evaluated OscSpikeSSM on long-sequence benchmarks in Section 4.1 and Appendix A.2.2 (Table 12). Regarding spike sparsity, Figure 3 in Appendix A.1.1 compares OscSpikeSSM with SiLIF and AdLIF under identical input stimulation and firing thresholds. OscSpikeSSM exhibits fewer, more temporally separated spikes while preserving bounded oscillatory state evolution.

Regarding the radar-specific spectral inductive bias, Section 4.5 and Appendix A.2.7 demonstrate the use of OscSpikeSSM-derived frequency-selective oscillator banks for processing raw FMCW radar signals. As reported in Table 8 of the revised manuscript, replacing conventional FFT preprocessing with RF-SpecSpkSSM reduces theoretical energy from 1.1 to 0.9 mJ/sample, with a modest accuracy trade-off of 3.1 percentage points (64.5% to 61.4%). This provides empirical evidence that oscillatory state-space dynamics can extract task-relevant range-Doppler representations. We clarify that the spectral-bank frequencies are kept fixed to preserve the physical range-Doppler grid, while the remaining front-end parameters are learned during training.

Table 5: Core-component ablations of mm-OscSpikeSSM on AOPHand. Accuracy is reported as mean ±\pm standard deviation across runs. All variants retain the same surrounding architecture and training protocol. The underdamping-constraint ablation maintains stable comparison dynamics.

ConfigurationAccuracy (%)
mm-OscSpikeSSM92.14±0.3292.14_{\pm0.32}
w/o underdamping constraint91.43±0.4691.43_{\pm0.46}
w/o feedback90.98±0.7590.98_{\pm0.75}
w/o reset90.27±0.6890.27_{\pm0.68}
AuthorsOpenReview ↗

Response to Reviewer HPF1

References

  1. Bal, M., and A. Sengupta. "P-SpikeSSM: Harnessing Probabilistic Spiking State Space Models for Long-Range Dependency Tasks." ICLR 2025.
  2. Liao, Z., et al. "Frequency Matching in Spiking Neural Networks for mmWave Sensing." ICML 2026.
  3. Rusch, T. K., and D. Rus. "Oscillatory State-Space Models." ICLR 2025.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Response to Reviewer WzBt

Thank you for your thorough review and detailed feedback. Please find our clarifications and revisions below:

Response to W1 and Q1: Evaluation Protocol: We thank the reviewer for raising this concern. Our experiments strictly followed the established dataset-specific training/evaluation/testing protocols adopted by the corresponding reference models (Liao et al., 2026; Yin et al., 2021; Kato et al., 2025; Lv et al., 2025; Vermunt & Corradi, 2024) to ensure direct comparability with published baselines. This was a deliberate choice to maintain consistency with established benchmarks and avoid introducing differences in the evaluation procedures that could affect the reported comparisons. We would like to clarify that an error in the Appendix led to this misunderstanding regarding the reported results. We have therefore clarified this in our revised manuscript.

Furthermore, we used optuna for hyperparameter optimisation, creating a separate validation set from the training data to select the best hyperparameters. The test set was not used during this search. Hope this clarifies the situation.

Response to W3 and Q3: MMVR Supervision and Architectural Contributions. We thank the reviewer for raising this concern. We would like to first clarify that both RAPTR (Kato et al., 2025) and our proposed MMVR-OscSpikeSSM follow the same weak supervision based on 3D bounding boxes and projected 2D keypoints, as the MMVR dataset itself does not provide ground-truth 3D joint annotations. Both methods employ box, gravity, projected 2D keypoint, OKS, and confidence-related losses, while their geometric constraints are adapted to their respective architectures. Our complete training objective, supervision strategy, and loss configurations are as described in Section 4.3 and Appendix A.2.5 under “Weakly-Supervised Pose Decoding and Training” and “Training Configuration.” Thus, we believe that both methods use the same weakly supervised setting and follow the same MMVR evaluation protocol, although their architectures differ.

The reported 3.41-fold reduction in centre-distance error is definitely not due solely to the OscSpikeSSM core. Having established the effectiveness of our OscSpikeSSM over existing LIF-based spiking SSMs in Section 4.2 (please also check the updated manuscript), we next develop a radar-specific architecture for 3D pose estimation in Section 4.3. Our MMVR-OscSpikeSSM is a new radar-specific architecture, and its design is itself a contribution of this work. Unlike applying a generic sequence model directly to flattened radar features, our architecture preserves the two radar views and introduces multidirectional spiking state-space scans along the temporal, spatial, and depth axes, depth-aware cross-view fusion, and query-based 3D pose refinement. These components are designed to exploit the spatial and geometric structure of multi-view radar measurements. To distinguish the contribution of this complete architecture from the underlying temporal model, we include a comparison between Raw OscSpikeSSM and the proposed MMVR-OscSpikeSSM in Table 5 of our revised manuscript. Raw OscSpikeSSM directly processes radar sequences using stacked OscSpikeSSM layers and a pose prediction head. Furthermore, as reported in Section 4.3, we conduct additional architectural ablations, including replacing OscSpikeSSM with LIF, forward-only scanning, and temporal-only scanning, to examine the contributions of the oscillatory core and architecture design.

As acknowledged by RAPTR (Kato et al., 2025), MMVR lacks ground-truth 3D joint annotations; hence, we follow its evaluation protocol using enclosing 3D bounding-box metrics as indirect measures of the predicted pose's spatial localisation and extent. Beyond MMVR, our work introduces a new spiking SSM, OscSpikeSSM, supported by extensive experiments, architectural ablations, and efficiency analyses across diverse radar sensing applications (Sections 4.2-4.5).

Table 2: Comparison of raw and radar-specific OscSpikeSSM architectures on MMVR P1S1.

MethodCenter distance (cm)Edge length error (cm): (h)Edge length error (cm): (v)Edge length error (cm): (d)
Raw OscSpikeSSM38.2839.9965.1467.42
MMVR-OscSpikeSSM9.2022.264.0617.92
AuthorsOpenReview ↗

Response to Reviewer WzBt

Response to W2 and Q2: Discrete-Time Stability. We thank the reviewer for raising this distinction. We have expanded Appendix A.1.2 to include a detailed discrete-time stability analysis of the proposed OscSpikeSSM (as detailed below), including the stability condition for the learned integration step and the effect of spike-dependent feedback and subtractive reset. We have also added Table 9 in the revised manuscript, reporting the spectral radii and Jury stability margins of trained checkpoints across representative datasets. The results confirm that all evaluated oscillator modes satisfy the discrete-time stability condition. We further clarify that our theoretical result guarantees bounded state trajectories under bounded inputs and a Schur-stable transition.

The parameterisation c>0c>0 and k>c2/4k>c^2/4 guarantees stable underdamped continuous-time modes, but the learned step Δt=exp⁡(Δtlog⁡)\Delta t=\exp(\Delta t_{\log}) is constrained only to be positive. For one hidden mode, let τ=Δt\tau=\Delta t and d=1+τcd=1+\tau c. With zero input and no spikes, the damped-IMEX update becomes zt=Adzt−1\mathbf z_t=A_d\mathbf z_{t-1}, where

A_d=\frac{1}{d} \begin{bmatrix} d-\tau^2k & \tau\\\\ -\tau k & 1 \end{bmatrix}. \tag{1}$$ Its eigenvalues are: $$\lambda_{\pm}= \frac{2+\tau c-\tau^2k\pm \sqrt{(2+\tau c-\tau^2k)^2-4(1+\tau c)}}{2(1+\tau c)}. \tag{2}$$ Since $\det(A_d)=1/d$ and $\operatorname{tr}(A_d)=(d+1-\tau^2k)/d$, the three Jury conditions reduce to $\tau c/d>0$, $\tau^2k/d>0$, and $(4+2\tau c-\tau^2k)/d>0$. Thus, for $c,k,\tau>0$, the unforced discrete transition is Schur stable if and only if $\Delta t^2 k < 4 + 2\Delta t c$. **Spike feedback and reset.** With the input and spike terms retained, the implemented recurrence can be expressed as $$\mathbf z_t = A_d\mathbf z_{t-1} +\mathbf g I_t-\mathbf q s_{t-1}, \qquad s_{t-1}=H(u_{t-1}-\theta). \tag{3}$$ where $$\mathbf g=\frac{\tau B}{d} \begin{bmatrix}\tau\\\\1\end{bmatrix}, \qquad \mathbf q= \begin{bmatrix} r_{\mathrm{reset}}+\tau^2R/d\\\\ \tau R/d \end{bmatrix}. \tag{4}$$ The reset modifies the position state, whereas feedback modifies the velocity update and, through it, the position state. Although the threshold makes this recurrence nonlinear, the spike contribution is bounded since $s_t\in\\{0,1\\}$. For a mode with $\rho(A_d)<1$, there exist $M\geq1$ and $0<\eta<1$ satisfying $\\|A_d^n\\|\leq M\eta^n$. If $|I_t|\leq I_{\max}$, then $$\\|\mathbf z_t\\| \leq M\eta^t\\|\mathbf z_0\\| +\frac{M}{1-\eta} \bigl(\\|\mathbf g\\|I_{\max}+\\|\mathbf q\\|\bigr). \tag{5}$$ Hence, Schur stability of the homogeneous transition is sufficient for bounded state trajectories under bounded inputs, even with spike-dependent feedback and subtractive reset. This establishes bounded state trajectories under the stated conditions, rather than global asymptotic stability. **Stability of trained models.** Table 1 reports the discrete-time stability analysis of the trained models. We report the minimum Jury margin $m_{\min}=\min_i(4+2\Delta t_i c_i-\Delta t_i^2k_i)$ and maximum spectral radius $\rho_{\max}(A_d)=\max_i \rho(A_{d,i})$. All evaluated modes satisfy the Schur stability condition, with positive Jury margins and spectral radii below unity, confirming the stability of the homogeneous damped-IMEX transitions across the evaluated datasets. **Table 1: Discrete-time stability audit of the trained RFspkSSM models. The minimum Jury margin $m_{\min}$ and maximum spectral radius $\rho_{\max}(A_d)$ are evaluated across the learned oscillator modes in each model.** | Model | $m_{\min}$ | $\rho_{\max}(A_d)$ | | --- | ---: | ---: | | EigenWorms | 4.1038 | 0.9748 | | Google Soli | 2.2909 | 0.9828 | | MMVR | 4.0448 | 0.9889 | | Aircraft Marshalling | 4.0899 | 0.9781 |
AuthorsOpenReview ↗

Response to Reviewer WzBt

Response to Q4: Thanks for the catch. We have corrected the typographical error in Table 3 of Section 4.2, where the energy unit for Google Soli was mistakenly reported as mJ instead of μ\muJ. Table 3 (given below) presents the breakdown of dense MACs, spike-driven ACs, and their corresponding theoretical compute-energy estimates across the RFspkSSM architectures. Recent approaches, such as the Pulsar neuromorphic chip (Innatera, 2026), combine ANN and analog SNN modules to realise real-world applications, which also inspire our design principles.

Table 3: Breakdown of theoretical compute-energy estimates across RFspkSSM architectures. Dense MACs and spike-driven ACs are reported per sample (per clip for EgoHands). Energy is estimated using 4.6 pJ/MAC and 0.9 pJ/AC.

ModelDense MACsSpike ACsEnergy(mJ)
mm-OscSpikeSSM (AOPHand)1.42×1081.42 \times 10^{8}5.7×1025.7 \times 10^{2}0.65
mm-OscSpikeSSM (mmFiT)5.7×1075.7 \times 10^{7}4.3×1034.3 \times 10^{3}0.26
mm-OscSpikeSSM (MMActivity)1.2×1081.2 \times 10^{8}1.4×1021.4 \times 10^{2}0.58
Soli-OscSpikeSSM1.0×1071.0 \times 10^{7}1.9×1051.9 \times 10^{5}0.05
MMVR-OscSpikeSSM4.7×1094.7 \times 10^{9}4.9×1084.9 \times 10^{8}22.29
EgoHands-OscSpikeSSM1.1×10101.1 \times 10^{10}5.6×1065.6 \times 10^{6}51.06
RF-SpecSpkSSM1.9×1081.9 \times 10^{8}1.9×1071.9 \times 10^{7}0.93

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.