acceptodds
Under review as a conference paper at ICLR 2027

SpikeEEG-3D: Spiking Neural Networks for 3D Visual Decoding from EEG

Abstract

EEG-based 3D visual decoding is an important yet highly challenging task that aims to recover three-dimensional visual content from neural activity. Spiking neural networks (SNNs), with their brain-inspired and energy-efficient computation, have been increasingly studied for EEG decoding, yet their potential for EEG-to-3D visual decoding remains unexplored. In this work, we present SpikeEEG-3D, to the best of our knowledge, the first SNN framework for EEG-based 3D visual decoding, spanning the complete pipeline from EEG representation learning to colored point-cloud reconstruction. For EEG representation learning, SpikeEEG-3D employs Learnable Dynamics Leaky Integrate-and-Fire (LD-LIF) neurons, whose input gain and membrane time constant are jointly optimized to provide adaptive and diverse neuronal dynamics. We further propose Event-Enhanced Cross-Attention (EECA), which exploits spike-state transitions to facilitate information exchange between the static and dynamic EEG streams. For 3D reconstruction, we propose SpikeBit, which leverages bit-weighted virtual timesteps to provide finer-grained numerical representations while preserving spike-driven computation. Experiments on EEG-3D demonstrate excellent EEG classification performance and effective 3D reconstruction, with 20.8 and 3.1 lower energy consumption than Neuro-3D, respectively. These results demonstrate the feasibility of extending SNNs to EEG-based 3D visual decoding.

Primary area:
applications to neuroscience & cognitive science
Keywords:
3D Visual Decoding, Spiking Neural Networks, Brain-inspired Intelligence

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CCTdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SpikeEEG-3D, an SNN framework for reconstructing colored 3D point clouds from EEG signals. The model uses LD-LIF neurons to learn adaptive neuronal dynamics, Event-Enhanced Cross-Attention (EECA) to fuse static and dynamic EEG representations, and SpikeBit neurons to reduce quantization errors during 3D reconstruction.

Experiments on EEG-3D show improvements in classification accuracy, several reconstruction metrics, and estimated energy efficiency compared with existing methods.

Critical strengths

1.The paper explores EEG-based 3D reconstruction, a challenging and relatively underexplored application for spiking neural networks. The proposed framework covers both EEG representation learning and point-cloud generation, making it a meaningful attempt to extend SNNs to this task.

2.LD-LIF, EECA, and SpikeBit address different challenges in EEG representation learning and reconstruction. The ablation studies support their effectiveness, particularly the improvement brought by SpikeBit over conventional LIF neurons.

Critical weaknesses

1.The decoder also uses predicted object categories, and the encoder is trained with CLIP-based visual supervision. While these choices are reasonable, the paper does not clearly establish how much of the reconstruction quality comes from EEG-specific information rather than category priors or the generative decoder. A shuffled-EEG or category-only baseline would help clarify this important point.

  1. In Table 2, SpikeEEG-3D improves semantic accuracy and SSIM over Neuro-3D, but its Chamfer Distance and F1 scores are worse, indicating weaker geometric fidelity. Furthermore, the energy savings are based on analytical operation-level estimates rather than actual hardware measurements. These limitations weaken the claims of superior reconstruction quality and practical efficiency.
Questions for authors

1.Can the authors clarify how much of the 3D reconstruction performance comes from EEG-specific information rather than category priors? Are shuffled-EEG or category-only results available?

2.How do the authors explain the worse Chamfer Distance and F1 scores compared with Neuro-3D? How well do the estimated energy savings reflect practical hardware performance?

4 comments
AuthorsOpenReview ↗

Response to Reviewer CCTd Part (1/2)

Dear Reviewer CCTd:

We sincerely thank you for your positive feedback and valuable suggestions on our work! We appreciate your recognition of the interesting application of SNNs to EEG-based 3D reconstruction and the effectiveness of our architectural designs. We are pleased to clarify and answer the questions and provide some additional experimental results.

W1: The contribution of EEG information to reconstruction is not sufficiently isolated. The decoder also uses predicted object categories, and the encoder is trained with CLIP-based visual supervision. While these choices are reasonable, the paper does not clearly establish how much of the reconstruction quality comes from EEG-specific information rather than category priors or the generative decoder. A shuffled-EEG or category-only baseline would help clarify this important point.

Q1: Can the authors clarify how much of the 3D reconstruction performance comes from EEG-specific information rather than category priors? Are shuffled-EEG or category-only results available?

Response to W1 & Q1: Thank you for your constructive comment. To assess whether 3D reconstruction genuinely benefits from EEG-specific information beyond category priors and the generative decoder, we conduct two additional control experiments: EEG Shuffling, where EEG features are randomly shuffled across test samples while keeping the predicted categories unchanged, and Category-only, where the decoder relies solely on predicted categories without EEG features. The results are presented below.

Method(2,1) ↑(10,3) ↑CD ↓F1 ↑SSIM ↑
EEG Shuffling56.7136.426.869268.1090.6643
Category-only56.9337.646.576068.7250.6858
SpikeEEG-3D58.5840.706.028070.9160.7189

Both EEG Shuffling and Category-only lead to performance degradation across all five reconstruction metrics. These results demonstrate that EEG-specific information makes a meaningful contribution to 3D reconstruction beyond category priors and learned generative priors.

We include these additional experiments in Appendix A.7 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer CCTd Part (2/2)

W2: The reconstruction and efficiency improvements are not fully convincing. In Table 2, SpikeEEG-3D improves semantic accuracy and SSIM over Neuro-3D, but its Chamfer Distance and F1 scores are worse, indicating weaker geometric fidelity. Furthermore, the energy savings are based on analytical operation-level estimates rather than actual hardware measurements. These limitations weaken the claims of superior reconstruction quality and practical efficiency.

Q2: How do the authors explain the worse Chamfer Distance and F1 scores compared with Neuro-3D? How well do the estimated energy savings reflect practical hardware performance?

Response to W2 & Q2: Thank you for this valuable question. We acknowledge that our CD and F1 are slightly worse than those of Neuro-3D. Nevertheless, SpikeEEG-3D achieves comparable geometric reconstruction quality, while obtaining higher semantic reconstruction accuracy and SSIM with substantially lower estimated energy consumption. Our primary objective is to achieve competitive reconstruction quality with higher energy efficiency.

Regarding energy efficiency, we provide a more comprehensive analysis that better accounts for practical implementation costs. Following SpikedAttention [1], we incorporate memory access, neuronal state updates, and output write-back costs to estimate System-Level Energy. The results are presented below.

Energy comparison for EEG classification.

MethodComputation-Level Energy (mJ)System-Level Energy (mJ)
DeepNet0.2410.768
EEGNetv40.1330.425
TSConv0.3060.982
EEG Conformer0.6432.052
ShallowFBCSPNet0.2360.752
Neuro-3D2.3047.327
BrainAlign1.4724.672
SpikeEEG-3D0.1110.490

Energy comparison for 3D reconstruction.

MethodComputation-Level Energy (J)System-Level Energy (J)
Neuro-3D327.7981040.640
SpikeEEG-3D w/ LIF-T4110.415497.490
SpikeEEG-3D w/ LIF-T8173.406705.170
SpikeEEG-3D w/ SpikeBit-B4102.098481.529
SpikeEEG-3D w/ SpikeBit-B8105.609512.720

Under the System-Level Energy evaluation, SpikeEEG-3D retains 14.95× and 2.03× higher estimated energy efficiency than Neuro-3D in EEG classification and 3D reconstruction, respectively. These findings further confirm that our method preserves considerable energy-efficiency advantages even under a more comprehensive energy evaluation. We include both energy estimates in Tables 1 and 2 of the revised manuscript and further clarify these energy comparisons.


We sincerely look forward to further discussion to clarify any other concerns. If you have additional suggestions, we are committed to addressing them and continuously improving our work based on your feedback.


[1] Hwang S, Lee S, Park D, et al. Spikedattention: Training-free and fully spike-driven transformer-to-snn conversion with winner-oriented spike shift for softmax operation[J]. Advances in Neural Information Processing Systems, 2024, 37: 67422-67445.

Dear reviewer, I'm sorry to bother you.We hope the above clarifications and the additional experiments in the revised draft sufficiently addressed your concerns. We remain committed to addressing any remaining points you may have during the discussion phase.

I appreciate the authors' thoughtful response and additional clarifications, which have resolved some of my concerns. As I acknowledge these improvements, I will keep my original rating unchanged.

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

This paper proposes SpikeEEG-3D, an SNN-based framework for reconstructing colored 3D objects from EEG signals. The framework introduces LD-LIF neurons for adaptive EEG representation learning, Event-Enhanced Cross-Attention (EECA) for static–dynamic EEG fusion, and SpikeBit neurons for improving numerical precision in iterative 3D reconstruction. Experiments on the EEG-3D benchmark demonstrate improvements in EEG classification and selected reconstruction metrics, along with lower estimated computational energy consumption compared with Neuro-3D.

Critical strengths
  1. The paper explores an interesting direction by extending spiking neural networks to EEG-based 3D visual decoding, integrating EEG representation learning and point-cloud reconstruction within a unified framework.

  2. The experiments show improvements in EEG classification and several 3D reconstruction metrics, while demonstrating potentially substantial energy savings through spike-driven computation.

Critical weaknesses

W1. Insufficient motivation and unclear methodological significance.

The motivation for the proposed framework is not sufficiently convincing. While the paper claims to be the first to apply SNNs to EEG-based 3D reconstruction, being the first to combine an existing architecture with a particular task does not necessarily constitute a meaningful methodological contribution. The paper does not clearly establish what additional scientific or practical benefits are enabled by EEG-to-3D reconstruction, or why SNNs are particularly suitable for this task compared with conventional ANN-based approaches. The main justification appears to be energy efficiency, yet the reported savings are theoretical estimates, and no architecture-matched ANN or quantized ANN baselines are provided. Therefore, the necessity and practical significance of the proposed framework remain unclear.

W2. Insufficient experimental comparisons and ablation studies.

The experimental evaluation is insufficient to support the claimed advantages. The 3D reconstruction experiments mainly compare against Neuro-3D, while omitting recent related methods such as BrainSSD and BrainGS. Moreover, although SpikeEEG-3D improves semantic reconstruction metrics, it performs worse than Neuro-3D on Chamfer Distance and F1, raising concerns about actual geometric reconstruction quality. Most ablations of LD-LIF and EECA focus on EEG classification rather than end-to-end 3D reconstruction, while SpikeBit is compared only with conventional LIF neurons without appropriate quantized alternatives. Furthermore, the paper lacks EEG-shuffling or category-only conditioning controls, making it unclear whether the reconstructed geometry genuinely reflects EEG-specific information or primarily relies on learned generative priors.

Questions for authors
  1. What is the key scientific or practical motivation for using SNNs in EEG-to-3D reconstruction? Beyond being the first SNN-based framework for this task, what concrete advantages does the proposed approach offer over conventional ANN-based methods?

  2. Why are recent EEG-to-3D methods such as BrainSSD and BrainGS not included in the main comparisons? Can the authors provide stronger baselines, particularly architecture-matched ANN and quantized ANN models, to justify the use of SNNs?

  3. Can the authors provide end-to-end 3D reconstruction ablations to isolate the contributions of LD-LIF, EECA, and SpikeBit? How does SpikeBit compare with existing multi-bit spiking or quantized neural representations under matched computational budgets?

  4. To what extent does the reconstructed 3D geometry depend on trial-specific EEG information rather than learned generative priors? Can the authors provide EEG-shuffling, mismatched EEG, or category-only conditioning experiments to validate the EEG-to-3D decoding capability?

  5. Since the reported energy savings are based on theoretical operation costs, can the authors provide hardware measurements or a more comprehensive efficiency analysis accounting for memory access, virtual timesteps, and implementation overhead?

4 comments
AuthorsOpenReview ↗

Response to Reviewer guXy Part (1/3)

Dear Reviewer guXy:

We sincerely thank you for your thoughtful feedback and valuable suggestions on our work! We also appreciate your recognition of SNN-based EEG-to-3D decoding as an interesting research direction, as well as the improvements in EEG classification and reconstruction performance and the potential energy savings.

We are pleased to clarify and answer these questions.

W1.1: Insufficient motivation and unclear methodological significance. The motivation for the proposed framework is not sufficiently convincing. While the paper claims to be the first to apply SNNs to EEG-based 3D reconstruction, being the first to combine an existing architecture with a particular task does not necessarily constitute a meaningful methodological contribution. The paper does not clearly establish what additional scientific or practical benefits are enabled by EEG-to-3D reconstruction, or why SNNs are particularly suitable for this task compared with conventional ANN-based approaches. The main justification appears to be energy efficiency, yet the reported savings are theoretical estimates, […] Therefore, the necessity and practical significance of the proposed framework remain unclear.

Q1: What is the key scientific or practical motivation for using SNNs in EEG-to-3D reconstruction? Beyond being the first SNN-based framework for this task, what concrete advantages does the proposed approach offer over conventional ANN-based methods?

Response to W1.1 & Q1: Sorry for the confusion. EEG-to-3D decoding aims to recover both 3D geometry and visual appearance from brain activity, providing a richer way to investigate how visual and spatial information is represented in EEG signals, with potential applications in brain–computer interfaces. We discuss the significance of EEG-to-3D decoding in the Introduction of the revised manuscript.

Our motivation goes beyond being the first to apply SNNs to EEG-to-3D reconstruction. We aim to achieve competitive reconstruction quality while substantially reducing energy consumption (Computation-Level and System-Level) compared with ANN-based methods (see Response to Q5). Moreover, given the brain-inspired nature of SNNs, applying them to the challenging task of EEG-to-3D reconstruction can further facilitate the investigation of how visual information is represented in and decoded from brain activity. This exploration opens a new research direction for both SNN development and understanding visual information represented in EEG signals, offering significant scientific value to both fields.

Furthermore, our framework is not a straightforward application of existing SNN architectures. To address the specific challenges of this task, we introduce LD-LIF for adaptive EEG representation learning, EECA for effective interaction between static and dynamic EEG streams, and SpikeBit to reduce numerical representation errors during iterative 3D reconstruction. These task-oriented designs constitute the methodological contributions of our work.

AuthorsOpenReview ↗

Response to Reviewer guXy Part (2/3)

W1.2: no architecture-matched ANN or quantized ANN baselines are provided

W2.1: Insufficient experimental comparisons and ablation studies. The experimental evaluation is insufficient to support the claimed advantages. The 3D reconstruction experiments mainly compare against Neuro-3D, while omitting recent related methods such as BrainSSD and BrainGS. Moreover, although SpikeEEG-3D improves semantic reconstruction metrics, it performs worse than Neuro-3D on Chamfer Distance and F1, raising concerns about actual geometric reconstruction quality. […] SpikeBit is compared only with conventional LIF neurons without appropriate quantized alternatives

Q2: Why are recent EEG-to-3D methods such as BrainSSD and BrainGS not included in the main comparisons? Can the authors provide stronger baselines, particularly architecture-matched ANN and quantized ANN models, to justify the use of SNNs?

Response to W1.2 & W2.1 & Q2: Thank you for your insightful question. Our primary objective is to investigate the effectiveness of SNNs for EEG-to-3D reconstruction. The official code for BrainSSD and BrainGS was not available. We encountered difficulties and deviations when reproducing these methods. Therefore, we adopt Neuro-3D as our primary baseline, as it provides publicly available code and a comparable reconstruction pipeline. Furthermore, since BrainSSD and BrainGS employ substantially different architectures from ours and our goal is to demonstrate the effectiveness of SNNs in this field, we do not include them in our comparisons.

We conduct additional comparisons with architecture-matched ANN and W32A8 quantized ANN decoders using post-training quantization (PTQ) and quantization-aware training (QAT). The results are presented below.

Method(2,1) ↑(10,3) ↑CD ↓F1 ↑SSIM ↑
ANN (FP32)56.7337.866.026570.5730.7059
W32A8 (PTQ)56.5537.036.141870.0920.7092
W32A8 (QAT)56.1237.586.035470.3080.7067
SpikeBit (B=8)58.5840.706.028070.9160.7189

SpikeBit achieves higher semantic reconstruction accuracy, F1, and SSIM than the FP32 ANN baseline, while maintaining comparable CD. Furthermore, SpikeBit outperforms both quantized ANN baselines across all five reconstruction metrics.

Finally, although our CD and F1 are slightly worse than those of Neuro-3D, we maintain comparable geometric reconstruction quality while achieving the best SSIM and higher semantic reconstruction accuracy with substantially lower estimated energy consumption. Overall, our objective is to achieve competitive reconstruction quality with higher energy efficiency, demonstrating the feasibility and effectiveness of SNNs for EEG-based 3D visual decoding.


W2.2: Most ablations of LD-LIF and EECA focus on EEG classification rather than end-to-end 3D reconstruction

Q3: Can the authors provide end-to-end 3D reconstruction ablations to isolate the contributions of LD-LIF, EECA, and SpikeBit? How does SpikeBit compare with existing multi-bit spiking or quantized neural representations under matched computational budgets?

Response to W2.2 & Q3: Thank you for your constructive suggestion. To further evaluate the contributions of LD-LIF and EECA to end-to-end 3D reconstruction, we conduct additional ablation experiments. The results are presented below.

Method(2,1) ↑(10,3) ↑CD ↓F1 ↑SSIM ↑
w/o LD-LIF57.4239.536.09169.9670.7128
w/o EECA56.8138.946.17569.8840.7029
SpikeEEG-3D58.5840.706.02870.9160.7189

Removing either LD-LIF or EECA degrades all five reconstruction metrics, demonstrating their contributions to downstream 3D reconstruction.

Furthermore, we compare SpikeBit with integer-valued I-LIF neurons [1] and burst-style neurons [2] under comparable synaptic operation (SOP) budgets. The results are presented below.

Method(2,1) ↑(10,3) ↑CD ↓F1 ↑SSIM ↑SOP (×1012\times 10^{12}) ↓
Burst Neuron (T=4)55.1434.836.286769.3930.673243.31
I-LIF (T=1, D=8)55.6537.836.416069.4110.696149.38
SpikeBit (B=8)58.5840.706.028070.9160.718947.42

Under comparable SOP budgets, SpikeBit consistently outperforms both I-LIF and burst-style neurons across all five reconstruction metrics.

We include the additional ablation studies in Appendix A.7 and the neuron comparisons in Section 4.4 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer guXy Part (3/3)

W2.3: Furthermore, the paper lacks EEG-shuffling or category-only conditioning controls, making it unclear whether the reconstructed geometry genuinely reflects EEG-specific information or primarily relies on learned generative priors.

Q4: To what extent does the reconstructed 3D geometry depend on trial-specific EEG information rather than learned generative priors? Can the authors provide EEG-shuffling, mismatched EEG, or category-only conditioning experiments to validate the EEG-to-3D decoding capability?

Response to W2.3 & Q4: Thank you for this exceptionally insightful question! To investigate the contribution of EEG-specific information to 3D reconstruction, we conduct two additional control experiments: EEG Shuffling, where EEG features are randomly shuffled across test samples while keeping the predicted categories unchanged, and Category-only, where the decoder relies solely on predicted categories without EEG features. The results are presented below.

Method(2,1) ↑(10,3) ↑CD ↓F1 ↑SSIM ↑
EEG Shuffling56.7136.426.869268.1090.6643
Category-only56.9337.646.576068.7250.6858
SpikeEEG-3D58.5840.706.028070.9160.7189

SpikeEEG-3D outperforms both control variants across all five reconstruction metrics. In particular, removing or shuffling the continuous EEG features leads to noticeable degradation in CD, F1, and SSIM, alongside reduced semantic reconstruction accuracy. These results demonstrate that EEG features provide additional information that contributes to the reconstruction of 3D geometry.

We include these additional experiments in Appendix A.7 of the revised manuscript.


Q5: Since the reported energy savings are based on theoretical operation costs, can the authors provide hardware measurements or a more comprehensive efficiency analysis accounting for memory access, virtual timesteps, and implementation overhead?

Response to Q5: Thank you for this valuable question. To provide a more comprehensive efficiency evaluation, we extend our original MAC/AC-based Computation-Level Energy model following SpikedAttention [3]. Specifically, the extended System-Level Energy analysis accounts for memory access, neuronal state updates, output write-back, and additional computational overhead associated with SpikeBit's virtual timesteps, including timestep-dependent coefficient scaling.

Energy comparison for EEG classification.

MethodComputation-Level Energy (mJ)System-Level Energy (mJ)
DeepNet0.2410.768
EEGNetv40.1330.425
TSConv0.3060.982
EEG Conformer0.6432.052
ShallowFBCSPNet0.2360.752
Neuro-3D2.3047.327
BrainAlign1.4724.672
SpikeEEG-3D0.1110.490

Energy comparison for 3D reconstruction.

MethodComputation-Level Energy (J)System-Level Energy (J)
Neuro-3D327.7981040.640
SpikeEEG-3D w/ LIF-T4110.415497.490
SpikeEEG-3D w/ LIF-T8173.406705.170
SpikeEEG-3D w/ SpikeBit-B4102.098481.529
SpikeEEG-3D w/ SpikeBit-B8105.609512.720

In the System-Level Energy evaluation, SpikeEEG-3D still achieves 14.95× and 2.03× higher estimated energy efficiency than Neuro-3D for EEG classification and 3D reconstruction, respectively. These results demonstrate that our method maintains substantial energy-efficiency advantages under a more comprehensive evaluation. We include both energy estimates in Tables 1 and 2 of the revised manuscript and further clarify these energy comparisons.


We sincerely look forward to further discussion to clarify any other concerns. If you have additional suggestions, we are committed to addressing them and continuously improving our work based on your feedback.


[1] Luo X, Yao M, Chou Y, et al. Integer-valued training and spike-driven inference spiking neural network for high-performance and energy-efficient object detection[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 253-272.

[2] Shen G, Zhao D, Zeng Y. Exploiting high performance spiking neural networks with efficient spiking patterns[J]. IEEE Transactions on Emerging Topics in Computational Intelligence, 2025, 9(2): 1480-1489.

[3] Hwang S, Lee S, Park D, et al. Spikedattention: Training-free and fully spike-driven transformer-to-snn conversion with winner-oriented spike shift for softmax operation[J]. Advances in Neural Information Processing Systems, 2024, 37: 67422-67445.

Dear reviewer, I'm sorry to bother you.We hope the above clarifications and the additional experiments in the revised draft sufficiently addressed your concerns. If you are satisfied, we kindly request you to consider updating the score to reflect the newly added results and discussion. We remain committed to addressing any remaining points you may have during the discussion phase.

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

SpikeEEG-3D is a spiking pipeline that reconstructs colored point clouds from EEG. Its encoder uses LD-LIF neurons, which learn an input gain g and a membrane time constant, and Event-Enhanced Cross-Attention (EECA), which uses XOR of consecutive query spikes to modulate the value tensor when fusing static and dynamic EEG. The decoder is a point-cloud diffusion model whose neurons are replaced by SpikeBit, a bit-weighted multi-level spike unfolded over B virtual timesteps. On the EEG-3D dataset (12 subjects, 72 categories, 8 training and 2 test objects per category), it beats Neuro-3D on EEG classification and on N-way Top-K, with modeled energy 20.8x lower for the encoder and 3.1x lower for the decoder.

Critical strengths

•It is the first SNN pipeline for EEG to 3D, covering classification and generation. The appendices give the dataset split, metric definitions, and the energy model with formulas.

•Gains point the same way across metrics: color Top-1 of 43.00 against 39.93 for Neuro-3D, reconstruction (2,1) of 58.58 against 55.72, (10,3) of 40.70 against 36.58, and SSIM of 0.7189 against 0.7085.

•SpikeBit's powers-of-two weighting means B virtual timesteps cover 2^B levels, so cost grows with the logarithm of the number of levels. This is a sensible design, and Table 2 shows it beating LIF at equal step counts.

•The region-wise analysis (Fig. 5) agrees with posterior-dominant responses, although it is only descriptive.

Critical weaknesses

•The classification gains are small next to the visible noise. Object Top-1 is 6.42% against 5.91% for Neuro-3D and 6.11% for BrainAlign on a 72-way task (chance 1.39%). The three-seed averages come without standard deviations. The ablations show variation of similar size: in Table 5, adding a sustained-activation cue to the transition cue gives 5.32% Top-1, below using no cue at all (5.50%).

•The reconstruction scores are close to chance and geometry gets worse. With uniformly random ranks, chance is K/N, which is 50% for (2,1) and 30% for (10,3); the model reaches 58.58% and 40.70%. Chamfer distance (6.028 against 5.954) and F1 (70.916 against 72.725) are worse than Neuro-3D (Table 2). Fig. 4 shows only successes from the new model, with no Neuro-3D outputs and no failures. The diffusion model is conditioned on the predicted category, so semantic scores are tied to a classifier with 6.4% Top-1 accuracy. The conclusion's "effective 3D reconstruction" should be toned down.

•LD-LIF adds little. The input gain g multiplies the synaptic input, and Appendix A.3 itself says g can be folded into the layer weights at inference. Functionally, learning g is then redundant with learning the weights, so the +0.5 points from g alone in Table 4 would be an optimization effect, and the claim of a broader learnable space overstates it. PLIF already learns the time constant.

•The energy figures are modeled, with generous assumptions. Appendix A.3 uses 4.6 pJ per MAC and 0.9 pJ per AC at 45 nm, weights operations by firing rate, and excludes memory access and control. The 20.8x in Table 1 is relative to Neuro-3D's encoder, while a small ANN in the same table (EEGNetv4, 0.133 mJ) is within 1.2x of the 0.111 mJ result. Table 2 gives energy in joules without saying whether it is per sample. At B=8, SpikeBit is an 8-bit activation spread over time, so an INT8 ANN decoder is the fair comparison, and the per-timestep 2^r γ± scaling is not something standard spike-driven hardware provides.

•SpikeBit has close relatives that are not compared. The paper cites integer-valued spiking neurons with spike-driven inference (Luo et al., 2024) only as a source of quantization error. The binary-weighted unfolding is a real difference, but there is no head-to-head test against integer-valued or burst-style neurons at equal synaptic events.

•Cross-subject behavior is not reported. The split is by object, and the paper does not say whether encoders are pooled across subjects or per subject, nor give per-subject results. The brain-region analysis uses representative subjects, and in Fig. 5b several regions (frontal, for example) have error bars that span zero. The EECA effect on the 3D stage is not isolated, and the number of diffusion steps is not given in the main text.

Questions for authors

1.Can you give mean and standard deviation over the three seeds for Tables 1 to 5, and test whether the 0.3 to 0.5 point classification gains are significant?

2.Why does adding a sustained cue to the transition cue fall below the no-cue baseline in Table 5?

3.Is g redundant with the weights? A control that fixes g=1 and rescales the weight learning rate, or learns a per-channel weight scale instead, would answer this.

4.How does SpikeBit compare with an integer-valued neuron and a burst-style neuron at equal synaptic events, and with an INT8 ANN decoder?

5.Is the energy in Table 2 per sample, and how many denoising steps do Neuro-3D and SpikeEEG-3D use?

6.Can you show Neuro-3D and SpikeEEG-3D reconstructions on the same test samples, including failures, and report chance-level references next to Table 2?

7.How would the per-virtual-timestep scaling be realized on spike-driven hardware?

6 comments
AuthorsOpenReview ↗

Response to Reviewer WXms Part (1/5)

Dear Reviewer WXms:

We sincerely thank you for your thoughtful review and for recognizing our work as the first SNN pipeline for EEG-to-3D decoding, the sensible design of SpikeBit, and the clarity of our presentation.

We are pleased to clarify and answer the questions and provide some additional experimental results.

W1.1: The classification gains are small next to the visible noise. Object Top-1 is 6.42% against 5.91% for Neuro-3D and 6.11% for BrainAlign on a 72-way task (chance 1.39%). The three-seed averages come without standard deviations.

Q1: Can you give mean and standard deviation over the three seeds for Tables 1 to 5, and test whether the 0.3 to 0.5 point classification gains are significant?

Response to W1.1 & Q1: Thank you for your valid point. In our original submission, the results of comparison methods were adopted from previously published evaluations on the EEG-3D benchmark, consistent with the practice in prior work [1]. These evaluations report average accuracies without providing standard deviations across random seeds. To assess the statistical significance and reliability of our improvements, we reproduced Neuro-3D (BrainAlign's code is not publicly available) with three independent random seeds under the same evaluation protocol. The Object Top-1 accuracy (%) results are reported below:

MethodSeed 1Seed 2Seed 3Mean ± Std
Neuro-3D (Reproduced)5.505.905.995.80 ± 0.26
SpikeEEG-3D6.316.486.486.42 ± 0.10

SpikeEEG-3D achieves a statistically significant improvement over the reproduced Neuro-3D (+0.63%, p = 0.040, two-sided Welch's t-test), with a lower standard deviation (0.10 vs. 0.26), demonstrating the effectiveness and stability of our method. Furthermore, we have updated Tables 1–5 in the revised manuscript to report the mean and standard deviation for our experiments.


W1.2: The ablations show variation of similar size: in Table 5, adding a sustained-activation cue to the transition cue gives 5.32% Top-1, below using no cue at all (5.50%).

Q2: Why does adding a sustained cue to the transition cue fall below the no-cue baseline in Table 5?

Response to W1.2 & Q2: Thank you for your insightful question. Combining transition and sustained-activation cues may lead to overly widespread event enhancement, thereby reducing the distinguishability of different query spike patterns during cross-stream interaction. This is inconsistent with the design motivation of EECA. Consequently, combining these cues does not necessarily improve performance and may even lead to degradation.


W2.1: The reconstruction scores are close to chance and geometry gets worse. With uniformly random ranks, chance is K/N, which is 50% for (2,1) and 30% for (10,3); the model reaches 58.58% and 40.70%. Chamfer distance (6.028 against 5.954) and F1 (70.916 against 72.725) are worse than Neuro-3D (Table 2).

Response to W2.1: Thank you for your constructive comment. Our reconstruction results exceed the chance levels by 8.58% and 10.70% on (2,1) and (10,3), respectively. This is a highly challenging task, and SpikeEEG-3D further outperforms Neuro-3D by 2.86% and 4.12% on these two metrics. Although our CD and F1 are slightly worse than those of Neuro-3D, we maintain comparable geometric reconstruction quality while achieving higher semantic reconstruction accuracy and SSIM. Our objective is to achieve competitive reconstruction quality with substantially lower estimated energy consumption, demonstrating the feasibility and effectiveness of SNNs for EEG-based 3D visual decoding.

AuthorsOpenReview ↗

Response to Reviewer WXms Part (2/5)

W2.2: Fig. 4 shows only successes from the new model, with no Neuro-3D outputs and no failures. The diffusion model is conditioned on the predicted category, so semantic scores are tied to a classifier with 6.4% Top-1 accuracy. The conclusion's "effective 3D reconstruction" should be toned down.

Q6: Can you show Neuro-3D and SpikeEEG-3D reconstructions on the same test samples, including failures, and report chance-level references next to Table 2?

Response to W2.2 & Q6: Thank you for pointing this out. We update Figure 4 in the revised manuscript to include reconstruction comparisons between Neuro-3D and SpikeEEG-3D on the same test samples, including failure cases, and further clarify the corresponding discussion in the text. SpikeEEG-3D better preserves distinctive structural details, as illustrated by the television antennae, while reproducing the dominant object colors more faithfully. In comparison, Neuro-3D sometimes fails to recover recognizable structures, as illustrated by the motorcycle and helicopter. Even when Neuro-3D successfully recovers the overall object shapes, it may still lose distinctive details or exhibit noticeable color deviations, as illustrated by the guitar and toilet. When neither method fully reconstructs the target geometry, SpikeEEG-3D can still retain meaningful structural cues, such as the elongated handle of the hammer. Nevertheless, our method still has limitations in accurately recovering the overall geometry in certain cases, as illustrated by the sneaker.

Furthermore, sorry for the confusion. Our decoder is conditioned on both continuous EEG-derived features and predicted categories. This allows the model to leverage coarse semantic guidance from category predictions while using richer EEG-derived information to complement and refine this guidance during 3D reconstruction. Therefore, the classification accuracy cannot be directly equated with reconstruction performance.

Finally, we revise the conclusion to emphasize the ability of SNN-based EEG-to-3D decoding to achieve competitive reconstruction quality with substantially lower estimated computational energy. We also add the chance levels to Table 2.


W3: LD-LIF adds little. The input gain g multiplies the synaptic input, and Appendix A.3 itself says g can be folded into the layer weights at inference. Functionally, learning g is then redundant with learning the weights, so the +0.5 points from g alone in Table 4 would be an optimization effect, and the claim of a broader learnable space overstates it. PLIF already learns the time constant.

Q3: Is g redundant with the weights? A control that fixes g=1 and rescales the weight learning rate, or learns a per-channel weight scale instead, would answer this.

Response to W3 & Q3: Thank you for this valuable question. The fact that the input gain gg can be absorbed into the preceding weights during inference does not imply that explicitly learning gg is redundant during training. Batch Normalization [2] and RepVGG [3] have similarly demonstrated that components that can be fused during inference can still provide meaningful benefits during training. Our statement that LD-LIF provides a broader learnable space for neuronal dynamics refers specifically to the neuronal level, where we introduce an additional learnable degree of freedom for neuronal dynamics.

Furthermore, we conducted the suggested learning-rate control experiments with g=1g=1 fixed, while varying the learning rate of the preceding synaptic weights. These experiments follow the gain-only ablation setting in Table 4. The results are presented below:

SettingObject Top-1 (%)
Fixed gg, LR = 0.5×4.63 ± 0.15
Fixed gg, LR = 1×4.75 ± 0.12
Fixed gg, LR = 2×4.82 ± 0.07
Learnable gg5.27 ± 0.10

The learnable-gg configuration significantly outperforms all learning-rate controls (all p<0.01p<0.01, two-sided Welch's t-test). These results demonstrate that the benefits of explicitly learning gg cannot be replicated by simply adjusting the learning rate within the evaluated settings.

AuthorsOpenReview ↗

Response to Reviewer WXms Part (3/5)

W4.1: The energy figures are modeled, with generous assumptions. Appendix A.3 uses 4.6 pJ per MAC and 0.9 pJ per AC at 45 nm, weights operations by firing rate, and excludes memory access and control. The 20.8x in Table 1 is relative to Neuro-3D's encoder, while a small ANN in the same table (EEGNetv4, 0.133 mJ) is within 1.2x of the 0.111 mJ result.

Response to W4.1: Thank you for your insightful comment. Our original energy evaluation adopts the widely used MAC/AC-based Computation-Level Energy model. To provide a more comprehensive evaluation, we further extend the analysis following SpikedAttention [4], incorporating memory access, neuronal state updates, and output write-back costs to estimate System-Level Energy. The energy comparisons are presented below.

Energy comparison for EEG classification.

MethodComputation-Level Energy (mJ)System-Level Energy (mJ)
DeepNet0.2410.768
EEGNetv40.1330.425
TSConv0.3060.982
EEG Conformer0.6432.052
ShallowFBCSPNet0.2360.752
Neuro-3D2.3047.327
BrainAlign1.4724.672
SpikeEEG-3D0.1110.490

Energy comparison for 3D reconstruction.

MethodComputation-Level Energy (J)System-Level Energy (J)
Neuro-3D327.7981040.640
SpikeEEG-3D w/ LIF-T4110.415497.490
SpikeEEG-3D w/ LIF-T8173.406705.170
SpikeEEG-3D w/ SpikeBit-B4102.098481.529
SpikeEEG-3D w/ SpikeBit-B8105.609512.720

Under the System-Level Energy model, SpikeEEG-3D still achieves 14.95× and 2.03× higher estimated energy efficiency than Neuro-3D for EEG classification and 3D reconstruction, respectively. Compared with EEGNetv4, SpikeEEG-3D achieves 1.20× higher computation-level energy efficiency but 13.3% lower system-level energy efficiency. Nevertheless, our method achieves a 62.9% relative improvement in Object Top-1 accuracy (6.42% vs. 3.94%), demonstrating a favorable trade-off between decoding performance and energy efficiency.

We include both energy estimates in Tables 1 and 2 of the revised manuscript and further clarify these energy comparisons.


W4.2: Table 2 gives energy in joules without saying whether it is per sample.

Q5: Is the energy in Table 2 per sample, and how many denoising steps do Neuro-3D and SpikeEEG-3D use?

Response to W4.2 & Q5: Thank you for the question. The energy values in Table 2 are estimated per sample. Following the official implementation of Neuro-3D, both Neuro-3D and SpikeEEG-3D use 1,000 diffusion denoising steps for shape generation. We clarify these details in Table 2 and the implementation details of the revised manuscript.


W4.3: At B=8, SpikeBit is an 8-bit activation spread over time, so an INT8 ANN decoder is the fair comparison.

W5: SpikeBit has close relatives that are not compared. The paper cites integer-valued spiking neurons with spike-driven inference (Luo et al., 2024) only as a source of quantization error. The binary-weighted unfolding is a real difference, but there is no head-to-head test against integer-valued or burst-style neurons at equal synaptic events.

Q4: How does SpikeBit compare with an integer-valued neuron and a burst-style neuron at equal synaptic events, and with an INT8 ANN decoder?

Response to W4.3, W5 & Q4: Thank you for your constructive suggestion. We conduct additional comparisons with integer-valued I-LIF neurons [5] and burst-style neurons [6] under comparable synaptic operation (SOP) budgets. The results are presented below.

Method(2,1) ↑(10,3) ↑CD ↓F1 ↑SSIM ↑SOP (×1012\times 10^{12}) ↓
Burst Neuron (T=4)55.1434.836.286769.3930.673243.31
I-LIF (T=1, D=8)55.6537.836.416069.4110.696149.38
SpikeBit (B=8)58.5840.706.028070.9160.718947.42

Under comparable SOP budgets, SpikeBit consistently outperforms both I-LIF and burst-style neurons across all five reconstruction metrics.

Furthermore, we compare SpikeBit with architecture-matched W32A8 quantized ANN decoders using post-training quantization (PTQ) and quantization-aware training (QAT), where W32A8 denotes FP32 weights and 8-bit activations. The results are presented below.

Method(2,1) ↑(10,3) ↑CD ↓F1 ↑SSIM ↑
W32A8 (PTQ)56.5537.036.141870.0920.7092
W32A8 (QAT)56.1237.586.035470.3080.7067
SpikeBit (B=8)58.5840.706.028070.9160.7189

SpikeBit also outperforms both quantized ANN baselines across all five reconstruction metrics.

We include these additional comparisons in Section 4.4 of the revised manuscript.

Comments on the paper

2 comments

Author Follow-up Summary for Reviewers and the Area Chair

Dear Reviewers and Area Chair,

We sincerely thank all reviewers for their careful reviews and constructive suggestions. We appreciate the recognition of our unified SNN-based EEG-to-3D framework, its classification and reconstruction improvements, and SpikeBit's design. To facilitate further discussion, we summarize the main strengths and concerns, followed by the additional evidence and corresponding revisions.

Detailed responses are provided in Parts 1–5 for Reviewer WXms, Parts 1–3 for Reviewer guXy, and Parts 1–2 for Reviewer CCTd.

Summary of Reviews

R-WXmsR-guXyR-CCTdStatus and Supporting Evidence
Strengths
Unified SNN framework covering EEG representation learning and 3D point-cloud generation✓✓✓
Improvements in EEG classification and selected reconstruction metrics✓✓✓
SpikeBit's bit-weighted design and improvements over conventional LIF neurons✓✓
Clear and easy-to-follow presentation✓✓
Concerns
Research motivation, methodological contributions, and baseline selection✓Clarified: EEG-to-3D research significance, task-specific contributions, and the rationale for selecting Neuro-3D explained.
Statistical reliability and encoder-design effectiveness✓Addressed: three-seed results and significance tests provided; learnable gain validated against learning-rate controls; EECA cue-combination behavior discussed.
Architecture-matched ANN and quantized ANN comparisons✓✓Addressed: FP32 and W32A8 PTQ/QAT comparisons demonstrate SpikeBit's reconstruction advantages.
Comparisons with alternative spiking neurons✓✓Addressed: SpikeBit outperforms I-LIF and Burst Neuron across all five metrics under comparable SOP budgets.
EEG information and downstream component contributions✓✓✓Addressed: decoder conditioning clarified; EEG Shuffling, Category-only, and LD-LIF/EECA ablations demonstrate their contributions to reconstruction.
Reconstruction quality and qualitative evidence✓✓✓Addressed: the quality–energy trade-off explained; same-sample comparisons, failure cases, and chance references provided; reconstruction claims refined.
Energy evaluation and neuromorphic implementation✓✓✓Addressed: System-Level Energy analysis preserves substantial advantages; the EEGNetv4 trade-off and a shift-and-add hardware mapping discussed.
Subject-wise analyses and evaluation settings✓Addressed: results and topographies for all 12 subjects provided; training protocol, regional-response interpretation, energy units, and denoising steps clarified.

Main Responses and Revisions

1. Architecture-matched ANN and alternative spiking comparisons (R-WXms Q4; R-guXy Q2–Q3). We compare SpikeBit with architecture-matched FP32 ANN and W32A8 quantized ANN decoders using PTQ and QAT. SpikeBit outperforms both quantized baselines across all five reconstruction metrics. Against FP32 ANN, it improves semantic accuracies from 56.73% / 37.86% to 58.58% / 40.70%, achieves higher F1 and SSIM, and maintains comparable CD (6.0280 vs. 6.0265). We also compare with I-LIF and Burst Neuron under comparable SOP budgets: 47.42 × 10¹², 49.38 × 10¹², and 43.31 × 10¹², respectively, for SpikeBit, I-LIF, and Burst Neuron. SpikeBit achieves the best results across all five reconstruction metrics, supporting its effectiveness relative to existing spiking alternatives (Section 4.4, Tables 6–7).

2. EEG-specific information and downstream reconstruction (R-WXms W2.2, W6.3; R-guXy Q3–Q4; R-CCTd Q1). We clarify that the decoder uses both continuous EEG features and predicted categories. EEG Shuffling retains the original predicted categories while shuffling EEG features; Category-only omits continuous EEG features. Both controls degrade all five reconstruction metrics. CD increases from 6.0280 to 6.8692 / 6.5760, and SSIM decreases from 0.7189 to 0.6643 / 0.6858, respectively, demonstrating the contribution of EEG features beyond category guidance. Removing either LD-LIF or EECA also degrades all five metrics, confirming that their benefits extend to downstream 3D reconstruction (Appendix A.7, Tables 9–10, referenced in Section 4.3).

3. Statistical reliability and encoder-design controls (R-WXms Q1–Q3). Across three independent seeds, SpikeEEG-3D achieves 6.42 ± 0.10% Object Top-1 accuracy versus 5.80 ± 0.26% for reproduced Neuro-3D, with p = 0.040 in a two-sided Welch's t-test and lower variability. We report mean and standard deviation for our experiments in Tables 1–5. In the gain-only setting, learnable g achieves 5.27 ± 0.10%, outperforming all tested fixed-gain learning-rate controls, whose best result is 4.82 ± 0.07% (all p < 0.01). This supports a training benefit beyond the tested learning-rate adjustments. Our response also explains that combining transition and sustained-activation cues may cause overly widespread enhancement, reducing query-pattern distinguishability (responses to R-WXms Q1–Q3).

4. More comprehensive energy evaluation and hardware mapping (R-WXms W4.1, Q7; R-guXy Q5; R-CCTd Q2). Following SpikedAttention, we extend the energy model to account for memory access, neuronal state updates, and output write-back, while retaining SpikeBit's virtual-timestep and coefficient-scaling costs. Under the System-Level Energy model, SpikeEEG-3D retains 14.95× and 2.03× higher energy efficiency than Neuro-3D for classification and reconstruction. Compared with EEGNetv4, it achieves 62.9% higher Object Top-1 accuracy with slightly higher SLE (0.490 vs. 0.425 mJ). These results support a favorable performance–energy trade-off (Tables 1–2; Section 4.2; Appendix A.3). We also describe a possible MSB-to-LSB shift-and-add mapping onto programmable neuromorphic hardware such as Loihi 2, preserving binary spike communication without per-timestep synaptic-weight modification (Appendix A.3, Eq. 26).

5. Reconstruction quality and qualitative evidence (R-WXms W2.1, Q6; R-guXy Q2; R-CCTd Q2). Our responses clarify the reconstruction quality–energy trade-off: SpikeEEG-3D achieves higher semantic reconstruction accuracy and SSIM with competitive geometric quality and substantially lower computational energy. Figure 4 now compares SpikeEEG-3D and Neuro-3D on the same test samples, including failure cases. Section 4.2 discusses distinctive structural details, such as television antennae, alongside remaining geometric limitations. Table 2 includes chance references of 50% and 30%, compared with our 58.58% and 40.70%. The Conclusion is revised to emphasize competitive reconstruction quality and energy efficiency.

6. All-subject evaluation and protocol clarification (R-WXms W6.1–W6.2, Q5). We provide classification results for all 12 subjects in Appendix A.6, Table 8, and their static, dynamic, and combined CAT topographies in Appendix A.5, Figure 7. The topographies broadly support the spatial-response observations in the main text. Our response clarifies subject-specific encoder training and distinguishes relative regional-response patterns from claims of statistically significant activation in every region. We also specify per-sample reconstruction energy in Table 2 and 1,000 diffusion denoising steps, matching Neuro-3D, in Appendix A.1.

7. Research motivation, methodological contributions, and baseline selection (R-guXy Q1–Q2). We discuss the significance of EEG-to-3D decoding in the Introduction and explain how LD-LIF, EECA, and SpikeBit address adaptive EEG representation learning, static–dynamic interaction, and numerical precision during iterative reconstruction. Our response also explains our choice of Neuro-3D as the primary baseline, given its publicly available code and comparable reconstruction pipeline. The additional controlled comparisons and ablations provide direct evidence for our task-oriented designs, beyond simply applying an existing SNN architecture to a new task.

Revised Manuscript

The revisions appear in Tables 1–5 (standard deviations); Tables 1–2, Section 4.2, and Appendix A.3 (energy and hardware); Figure 4, Table 2, and the Conclusion (reconstruction comparisons); Section 4.4, Tables 6–7 (ANN, quantized ANN, and spiking-neuron comparisons); Appendices A.5–A.6, Figure 7, and Table 8 (all-subject analyses); and Appendix A.7, Tables 9–10 (reconstruction ablations and EEG conditioning controls). The Introduction and Appendix A.1 clarify the motivation and diffusion steps, respectively. The reproduction-based statistical test and gain learning-rate controls are reported in our detailed responses to R-WXms Q1 and Q3.

We would greatly appreciate the reviewers' follow-up feedback on whether our responses and revisions have addressed their main concerns. We also respectfully invite the Area Chair to consider the additional evidence and help facilitate further discussion. If any concerns remain, we would be happy to provide further clarification during the discussion period. Thank you again to all reviewers and the Area Chair for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.