Self-cross Feature based Spiking Neural Networks for Efficient Few-shot Learning

We sincerely appreciate your insightful suggestions.

Q1:The experimental part lacks the robustness analysis of the model under different datasets such as event-based neuromorphic datasets.

R1:We are very grateful for this valuable suggestion. We supplemented the performance of the model on different datasets and different noise levels to further verify the generalization performance and robustness of the model. The experimental results are shown as follows:

Q2:Whether the network structure of SNNs could be improved to further enhance the performance?

R2:We sincerely appreciate the reviewer's constructive suggestion. Indeed, we agree that the network architecture of spiking neural networks (SNNs) holds significant potential for further performance enhancement. While our current model has achieved state-of-the-art performance on the neuromorphic N-Omniglot dataset, we recognize several promising directions for future architectural improvements. These include exploring spiking residual connections with learnable skip weights to address the vanishing gradient problem in deep SNNs while maintaining event-driven sparsity, as well as developing sparse event-driven attention modules and spike-optimized Transformer structures that operate exclusively on active spikes. Such innovations could not only improve model performance but also further reduce energy consumption, thereby advancing the development of efficient and biologically plausible neuromorphic systems. We plan to thoroughly investigate these directions in our future work.

Q3.The effect of the time step T on the results on the static dataset is not shown.

R3: We are very grateful for your valuable suggestions. We have added experiments on static datasets CUB and miniImageNet with time steps of 2, 4, and 8 to show the effect of time step T on static datasets.

We sincerely appreciate your insightful suggestions.

Q1:The choice of the optimal λ coefficient lacks a detailed experimental basis or theoretical explanation.

R1:We appreciate this insightful feedback. We supplement the detailed experimental basis from three aspects:

1.Empirical Evidence (Existing Results) Table 4 already shows λ=0.7 yields peak accuracy on CUB/miniImageNet.

2.Extended ablation studies will demonstrate: λ>0.5 consistently outperforms pure contrastive learning (λ=0) by +6.2% accuracy λ<0.9 avoids overfitting to temporal features (test loss ↓15%)

3.Biological Plausibility:λ≈0.7 aligns with hippocampal learning studies :70% local plasticity (TET) + 30% global modulation (InfoNCE)

Q2: There are not enough visualizations in the article to better intuitively feel the effectiveness of the model.

R2: We sincerely appreciate the reviewer's valuable suggestion regarding the need for more visualizations to better demonstrate our model's effectiveness. We acknowledge that the lack of visual elements in our original submission may have reduced the paper's intuitive appeal. In response to this constructive feedback, we have enhanced our revision by incorporating spike-based visualizations that specifically illustrate the activity patterns within our model's encoder component. These visualizations clearly demonstrate how increasing time steps lead to progressive activation of features through spike emissions, effectively capturing richer representations. While we regret that we cannot share these visualization results during the rebuttal phase, we are confident they will significantly improve readers' understanding of our model's dynamic feature extraction capabilities in the final version. This addition will make the model's effectiveness more tangible and visually apparent to the research community.

Q3:.Experiments should be set up at different time steps on static datasets to further verify the effectiveness of the model.

R3:We are very grateful for your valuable suggestions. We have added experiments on static datasets CUB and miniImageNet with time steps of 2, 4, and 8 to show the effect of time step T on static datasets

Q4.:Essential References Not Discussed.

R4: In our revision, we will incorporate a dedicated discussion comparing two key approaches: (1) Zhan et al.'s two-stage spiking meta-learning method [1] that introduces a novel bio-inspired framework combining spike-timing-dependent plasticity with meta-learning for few-shot classification, and (2) Yu et al.'s graph-based few-shot learning approach [2] that proposes a unified framework bridging meta-learning, pre-training and prompting techniques for graph-structured data. While Zhan's work focuses on biologically plausible SNN architectures and Yu's emphasizes graph neural networks, our method distinguishes itself by developing an event-driven, temporally-aware architecture that achieves superior computational efficiency while maintaining competitive accuracy. These comparisons will help better position our contributions within the broader landscape of few-shot learning research.

Q5:Could the proposed model be applied to larger datasets?

R5:We sincerely appreciate this forward-looking question, which aligns perfectly with our research vision. Indeed, our architecture is designed to be extensible to larger datasets, as demonstrated by its robust performance on medium-scale benchmarks like N-Omniglot (1,623 classes) and miniImageNet (100 classes). The model's strong backbone and efficient spike-based processing make it particularly suitable for scaling up. In our ongoing and future work, we plan to systematically evaluate the model's performance on larger-scale datasets such as tieredImageNet, which will further validate its scalability and generalization capabilities while maintaining computational efficiency. This extension represents a natural and important direction for our research.

I have carefully revised and proofread the spelling errors and citation errors. Thank you again for your valuable suggestions.

Reference

[1] A two-stage spiking meta-learning method for few-shot classification[J]. Knowledge-Based Systems, 2024.
[2] Few-shot learning on graphs: from meta-learning to pre-training and prompting, 2024.

We sincerely appreciate the insightful feedback, which has helped us clarify our contributions and identify areas for improvement. Below, we address each concern point-by-point:

Q1:Design motivation and theoretical analysis.

R1:Our paper introduces a novel framework, SSCF (Spiking Self-Cross Feature Network), which advances SNN-based few-shot learning through two key innovations: (1) a Self-Feature Extractor that captures intra-class temporal correlations, overcoming SNNs' spatiotemporal feature extraction limitations, and (2) a Cross-Feature Contrastive module employing 4D convolutions and joint attention to optimize support-query alignment. Unlike existing SNN-FSL works focusing on spike-time encoding, our framework uniquely integrates self-correlation analysis with cross-attention mechanisms, offering more comprehensive feature learning while maintaining SNNs' energy efficiency advantages.

Q2:About the last Energy consumption model comparison.

R2:Thank you for highlighting this important aspect. Because of the real condition limitation in the lab, we are sorry that we do not have the conditions to use TPU yet. The experiments in our paper are conducted on NVIDIA 4060ti, with a performance of 22 TFLOPS. That is why we employ the energy consumption estimation based on the value 1.732mJ, which is consistent with [1]Efficient Event-based Semantic Segmentation with Spike-driven Lightweight Transformer-based Networks,2024. Here, we further supplement the new energy consumption estimation by referring the recent energy as [2]Spiking-physformer: camera-based remote photoplethysmography with parallel spike-driven transformer,2025.

SOPs= fr × T × FLOPs

Power(ANN)=4.6pJ × FLOPs

Power(SNN)=0.9pJ × SOPs

Q3:Insufficient evidence for the energy efficiency experiment, CUB dataset not achieved SOTA.

R3:Thanks for your suggestion. The significant energy reduction stems from SNNs’ event-driven, sparse processing, aligning with neuromorphic principles. While performance is slightly lower than SOTA (-3.22% in 5-way-1-shot), our method bridges SNNs and ANNs, offering a low-power alternative with minimal accuracy loss—a key contribution for resource-constrained applications.Since SNNs have great spatiotemporal dynamics thanks to the spike-based neuron computation model, thus have advantages on processing event-based neuromorphic dataset, our experiments verified that the proposed SNNs’ few shot learning method remains and even improves these advantages. On event datasets (e.g., N-Omniglot), our model achieves 84.4% (20-way-1-shot), improving +20.3% over [3], leveraging SNNs’ spatiotemporal dynamics for event-data advantages. Moreover,on miniImageNet, it also outperforms [4] by +7.3%, demonstrating competitive performance.Therefore, we emphasize that our model achieves competitive performance among these datasets.

Q4:Comparison with methods from recent years on the N-Omniglot dataset.

R4:First of all, thank you very much for the two references you provided. [5] Knowledge Transfer for SNNs (AAAI 2024) explores static-to-event domain adaptation, which is orthogonal to our work. Their experiments are conducted on event-based datasets, and our performance on N-Omniglot surpasses it. [6] Brain-Inspired Meta-Learning (TNNLS 2024) focuses on bearing fault diagnosis, due to the data employed in this paper focus on bearing part failure problem and is not public, which makes it difficult for me to compare it under the same experimental setting. Instead, we add discussion about this study in the related work section.

I have carefully revised and proofread the spelling errors and citation errors. Thank you again for your valuable suggestions.

Reference

[1] Efficient Event-based Semantic Segmentation with Spike-driven Lightweight Transformer-based Networks, 2024.
[2] Spiking-physformer: camera-based remote photoplethysmography with parallel spike-driven transformer. Neural Networks, 2025
[3] An Efficient Knowledge Transfer Strategy for Spiking Neural Networks from Static to Event Domain, AAAI 2024.
[4] A two-stage spiking meta-learning method for few-shot classification[J]. Knowledge-Based Systems, 2024.
[5] An Efficient Knowledge Transfer Strategy for Spiking Neural Networks from Static to Event Domain, AAAI 2024.
[6] Brain-Inspired Meta-Learning for Few-Shot Bearing Fault Diagnosis, TNNLS 2024

We sincerely appreciate the insightful feedback, which has helped us clarify our contributions and identify areas for improvement.

Q1:How to tackles the challenges of SNN’s few shot learning? Simply adds two simple attention-like modules ?

R1:We appreciate this critique and clarify that our key innovations go beyond simple "attention-like" modules: Our innovations transcend conventional attention mechanisms: The Self-feature extractor enhances intra-class representation through relational feature integration ("observe yourself for what" ); The Cross Feature Contrastive module dynamically models support-query relationships via joint attention maps. Together, they synergistically improve SNN few-shot learning through dual optimization of self-representation and cross-class comparison, not mere attention replication.

Q2: Cross-class comparison?

R2:Our method enables cross-class comparison via two specialized modules: A bottleneck-structured self-feature extractor capturing intra-class patterns while boosting inter-class discrimination; A 4D-convolution based cross-feature comparator generating attention maps for explicit query-support relationship modeling. This SNN-compatible design combines event-driven LIF dynamics with advanced feature extraction, achieving 98.9% accuracy (5-way 5-shot on N-Omniglot) with clear feature separation (t-SNE verified).

Q3:Data scarcity

R3:Data scarcity drives few-shot learning research, motivating our development of SNN-based solutions that leverage meta-learning and metric learning to extract meaningful patterns from limited samples while maintaining generalization capability. Our systematic validation demonstrates significant improvements through: comprehensive benchmarking against state-of-the-art SNN few-shot methods (see R4 results), and synergistic operation of our novel self-feature extraction and cross-feature comparison modules, which collectively enhance few-shot learning performance in SNNs while preserving their energy-efficient characteristics.

Q4:Compared to other recent advanced SNNs’ few shot learning methods.

R4:Thanks for your suggestion.We have added some of the latest SNN few-shot learning methods for comparison, The performance comparison is show as follows:

Q5:The Generalization ability of the proposed model.

Thanks for your suggestion. Here we validate the performance of our model on N-Omniglot, CUB, and miniImageNet datasets, which are common-used validation settings as [5-7] . These datasets are recognized benchmark datasets in the FSL field, covering neuromorphic data, fine-grained classification, and general object classification scenarios, respectively.Meanwhile, in order to further verify the generalization performance and robustness of the model, we supplemented the performance of the model under different datasets and different noise levels as Table 1 in Reviewer 4's comments.

Thanks for your feedback. We have re-uploaded high-resolution images in the revision, and have carefully revised and proofread spelling errors and citation errors. We are sorry that we can not upload the figure during the rebuttal phase, but we really add more method details in the method d diagram to make it clear. Thank you again for your valuable suggestions.

Reference

[1] N-omniglot, a large-scale neuromorphic dataset for spatio-temporal sparse few-shot learning, 2022
[2] An Efficient Knowledge Transfer Strategy for Spiking Neural Networks from Static to Event Domain, AAAI 2024
[3] Fast on-device adaptation for spiking neural networks via online-within-online meta-learning, IEEE 2021
[4] A two-stage spiking meta-learning method for few-shot classification[J]. Knowledge-Based Systems, 2024
[5] Spatial-aware Metric Network via Patch-wise Feature Alignment for Few-shot Learning[J]. IEEE 2025.
[6] Hard-Positive Prototypical Networks for Few-shot Classification, IEEE 2025.
[7] EMNet: A Novel Few-Shot Image Classification Model with Enhanced Self-Correlation Attention and Multi-Branch Joint Module,2025