Pith. sign in

REVIEW 2 major objections 39 references

XOResNet: Exclusive-OR Meta-Residuals Facilitate Deep Spiking Neural Networks Learning

T0 review · 2 major / 0 minor · reviewed 2026-06-30 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read XOResNet uses OR-ADD shortcuts and XOR meta-residuals to fix spike redundancy and information loss in deep spiking neural networks.

desk verdict XOResNet adds an OR-ADD shortcut and XOR meta-residual selection to residual blocks in SNNs and claims better results on four image datasets than prior gradient-trained models. read the letter →

arxiv 2605.30362 v1 pith:WEBHHZJW submitted 2026-05-15 cs.NE cs.AIcs.CV

Jianfang Wu , Junsong Wang This is my paper · ORCID
classification cs.NEcs.AIcs.CV
keywords spikingneuralnetworksresiduallearningXORmeta-residualsOR-ADDshortcutdeepSNNsimageclassificationneuromorphiccomputing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper seeks to make residual learning work reliably in deep spiking neural networks by fixing three problems: relative spike redundancy during identity mappings, information loss during non-identity mappings, and redundant learning inside the backbone branch. It introduces an OR-ADD shortcut that merges spikes or currents from the two branches and XOR meta-residuals that select pre-learning residuals via the exclusive-or operation for the backbone. These two elements are combined into the XOR residual block, from which networks of varying depth called XOResNet are built. Experiments across Fashion-MNIST, CIFAR-10, CIFAR-100 and miniImageNet show the resulting models exceed prior state-of-the-art deep spiking networks trained by gradient descent. The claim matters because residual blocks are the standard route to depth in ordinary networks; a version that works for spiking models would open the same scaling path for neuromorphic hardware.

What carries the argument

The XOR residual block, formed by an OR-ADD shortcut that merges spikes or currents and XOR meta-residuals that select pre-learning residuals for the backbone branch via the exclusive-or operation.

What would settle it

Reproducing the four-dataset experiments and finding that XOResNet fails to exceed the accuracy of the compared prior deep SNNs would falsify the performance claim.

Watch

Extended reading notes

Core claim

Integrating the OR-ADD shortcut connection to merge branch outputs with XOR meta-residuals that select residuals for the backbone produces an XOR residual block; stacking these blocks yields XOResNet, which outperforms existing gradient-descent-optimized deep SNNs on Fashion-MNIST, CIFAR-10, CIFAR-100 and miniImageNet.

Load-bearing premise

The OR-ADD shortcut and XOR meta-residuals fix spike redundancy, information loss and redundant backbone learning without creating new offsetting problems.

Editorial extensions

If this is right

  • Deeper SNN architectures become feasible without the prior limits on identity and non-identity mappings.
  • Higher classification accuracy is obtained on the tested image datasets relative to earlier residual SNN designs.
  • The same block can be used to build networks of different depths while preserving the reported gains.
  • The components supply concrete architectural guidance for neuromorphic hardware that relies on spiking representations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same OR-ADD and XOR selection pattern could be tested in non-spiking residual networks that also suffer from feature redundancy.
  • Hardware energy measurements on neuromorphic chips would show whether the reduced spike redundancy translates into lower power draw.
  • Scaling the same blocks to larger image or video datasets would test whether the gains hold beyond the four reported benchmarks.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes XOResNet for deep spiking neural networks. It introduces an OR-ADD (OA) shortcut to merge spikes/currents from residual branches, addressing relative spike redundancy in identity mappings and information loss in non-identity mappings. It further defines XOR meta-residuals via the Exclusive-OR operation to reduce redundant learning in the backbone branch. These are combined into XOR residual blocks used to construct networks of varying depths. The central claim is that XOResNet outperforms existing state-of-the-art deep SNNs trained via gradient descent on Fashion-MNIST, CIFAR-10, CIFAR-100, and miniImageNet.

Significance. If the reported gains are shown to arise specifically from the OA shortcut and XOR meta-residuals under controlled conditions, the work would supply a concrete architectural pattern for scaling residual connections in SNNs. This could be useful for neuromorphic hardware design, as the components directly target named limitations of prior residual SNN structures while remaining compatible with gradient-based training.

major comments (2)
  1. [Abstract / Experimental evaluation] The central experimental claim (outperformance on four datasets) is load-bearing, yet the abstract provides no quantitative metrics, number of runs, error bars, or baseline implementation details. Without these, it is impossible to assess whether gains survive hyperparameter controls or data handling variations.
  2. [Experimental evaluation] No ablation results are referenced that isolate the contribution of the OA shortcut versus the XOR meta-residual selection. This is required to substantiate the claim that these components specifically resolve spike redundancy, information loss, and redundant learning without offsetting drawbacks.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major comment below and will revise the manuscript to strengthen the presentation of results.

read point-by-point responses

  1. Referee: [Abstract / Experimental evaluation] The central experimental claim (outperformance on four datasets) is load-bearing, yet the abstract provides no quantitative metrics, number of runs, error bars, or baseline implementation details. Without these, it is impossible to assess whether gains survive hyperparameter controls or data handling variations.

    Authors: We agree that the abstract should report key quantitative results. In the revised manuscript we will add the main accuracy figures for XOResNet versus the strongest baselines on each of the four datasets, together with the number of independent runs and standard deviations. Full hyperparameter settings and baseline re-implementation details already appear in the experimental section; we will ensure the abstract points to these controls. revision: yes

  2. Referee: [Experimental evaluation] No ablation results are referenced that isolate the contribution of the OA shortcut versus the XOR meta-residual selection. This is required to substantiate the claim that these components specifically resolve spike redundancy, information loss, and redundant learning without offsetting drawbacks.

    Authors: The current experiments demonstrate end-to-end gains of the complete XOResNet. To isolate the two proposed components we will add ablation tables in the revision that compare (i) standard residual blocks, (ii) blocks with only the OA shortcut, (iii) blocks with only XOR meta-residuals, and (iv) the full XOR residual block, all trained under identical conditions on the same datasets. These results will directly quantify the individual and combined effects on the issues of spike redundancy and redundant learning. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity

full rationale

The paper proposes an architectural modification (OA shortcut + XOR meta-residual block) to address named problems in residual SNNs and validates it via standard gradient-based training and accuracy comparisons on four external benchmark datasets. No equations, fitted parameters, or self-citations are presented that reduce the central claim to a tautology or to the input data by construction. The derivation chain consists of design motivation followed by empirical measurement; it remains self-contained against external benchmarks and does not invoke any of the enumerated circularity patterns.

Assumptions & free parameters 0 free parameters · 0 assumptions · 0 invented entities

No information available from abstract to populate free parameters, axioms, or invented entities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of XOResNet: Exclusive-OR Meta-Residuals Facilitate Deep Spiking Neural Networks Learning." pith.science (2026). https://pith.science/paper/WEBHHZJW

@misc{pith2026260530362,
  author       = {Pith},
  title        = {Pith review of: XOResNet: Exclusive-OR Meta-Residuals Facilitate Deep Spiking Neural Networks Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WEBHHZJW}},
  note         = {Machine review of arXiv:2605.30362}
}
read the original abstract

Spiking neural networks (SNNs) hold promise for demonstrating superior learning and representation capabilities in deep models. Given the tremendous success of ResNet in deep learning, it would naturally follow to train deep SNNs with residual learning. However, existing residual structures for constructing deep SNNs still present challenges of spike redundancy or information loss, as well as redundant learning. In the present study, we first aim to address issues of relative spike redundancy in identity mapping and information loss in non-identity mapping. To this end, we propose an OR-ADD (OA) shortcut connection to merge output spikes/currents from two branches in the residual structure. Furthermore, to mitigate redundant learning in the backbone branch of the residual structure, we introduce the concept of XOR meta-residuals, i.e., selecting pre-learning residuals using the Exclusive-OR (XOR) operation for the backbone branch. Finally, by integrating the OA shortcut and XOR meta-residuals, we devise the XOR residual block and further construct XOResNet with varying depths based on this block. Extensive experiments on four datasets, Fashion-MNIST, CIFAR-10, CIFAR-100, and miniImageNet, show that the proposed XOResNet outperforms existing state-of-the-art deep SNNs optimized via gradient descent. These results validate the effectiveness of our OA shortcut and XOR meta-residual components in overcoming fundamental limitations of residual learning in SNNs, providing new architectural insights for building high-performance neuromorphic systems.

Figures

Figures reproduced from arXiv: 2605.30362 by the authors.

Figure 1
Figure 1. The basic building blocks for deep SNNs. (a) The basic building block in Plain Network. (b) The basic building block in Spiking ResNet. S l [t]/O l [t] denotes the input/output spikes of layer l at time t. 1 × 1 and 3 × 3 denote the convolution kernel size. BN is a batch normalization operation. S N denotes the spiking neuron. In this work, we consider the specificity of spike communication in SNNs. For the shortcut… view at source ↗
Figure 2
Figure 2. Training accuracy and test accuracy of models with different depths on CIFAR-10. rents from the two branches are summed as input to the spiking neuron (SN), avoiding information loss [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Shortcut connection. (a) OR shortcut connection. (b) (c) OR-ADD (OA) shortcut connection [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Exclusive-OR meta-residuals structure. (a) OR shortcut connection. (b) ADD shortcut connection. primarily used for channel transformation of feature maps with approximate input and output features, the output feature scale of f l 2 (S l [t]) is the same as that of f l …
Figure 5
Figure 5. Figure 5: The network structure of XOResNet and its unfolded formulation. ×M denotes the number of iterations of the identity mapping, and ×N denotes the number of downsample steps. Note that the network’s parameters are shared at all time-steps [PITH_FULL_IMAGE:figures/full_fi…
Figure 6
Figure 6. Figure 6: Evaluation of models with varying depths on the CIFAR-10 dataset. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: The training and test accuracy of different models with different depths on different datasets. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: The Jaccard similarity coefficient between the input and output binary spike features of the backbone branch in the residual block with identity mapping. 4.2. Ablation study The proposed XOResNet shows superior performance compared to other models and methods on four d…
Figure 9
Figure 9. Figure 9: Confusion matrices of 50-layer networks on the Fashion-MNIST dataset. The left column represents the training dataset, and the right column represents the test dataset. CIFAR-10 We construct seven networks with varying depths, as shown in Ta￾ble 5, using different comp…
Figure 10
Figure 10. Figure 10: Confusion matrices of 110-layer networks on the CIFAR-10 dataset. The left column represents the training dataset, and the right column represents the test dataset. residual component markedly enhances learning capabilities [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]
Figure 11
Figure 11. Figure 11: Evaluation of models with varying depths on the CIFAR-100 dataset. the conclusions drawn from [PITH_FULL_IMAGE:figures/full_fig_p024_11.png]
Figure 12
Figure 12. Figure 12: Evaluation of models with varying depths on the miniImageNet dataset. Through extensive experiments, we demonstrate that the proposed XOR meta￾residuals structure can promote the learning of deep SNNs. Meanwhile, systematic ablation studies reveal the effectiveness of…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages

  1. [1]

    K. Roy, A. Jaiswal, P. Panda, Towards spike-based machine intelligence with neuromorphic computing, Nature 575 (7784) (2019) 607–617

    work page 2019

  2. [2]

    Maass, Networks of spiking neurons: the third generation of neural network models, Neural networks 10 (9) (1997) 1659–1671

    W. Maass, Networks of spiking neurons: the third generation of neural network models, Neural networks 10 (9) (1997) 1659–1671. 28

    work page 1997

  3. [3]

    Y . Guo, W. Peng, Y . Chen, L. Zhang, X. Liu, X. Huang, Z. Ma, Joint a-snn: Joint training of artificial and spiking neural networks via self-distillation and weight factorization, Pattern Recognition 142 (2023) 109639

    work page 2023

  4. [4]

    W. Fang, Z. Yu, Y . Chen, T. Huang, T. Masquelier, Y . Tian, Deep residual learning in spiking neural networks, Advances in Neural Information Processing Systems 34 (2021) 21056–21069

    work page 2021

  5. [5]

    R. Xian, X. Xiong, H. Peng, J. Wang, A. R. de Arellano Marrero, Q. Yang, Feature fusion method based on spiking neural convolutional network for edge detection, Pattern Recognition 147 (2024) 110112

    work page 2024

  6. [6]

    C. Tan, M. Šarlija, N. Kasabov, Neurosense: Short-term emotion recognition and understanding based on spiking neural network modelling of spatio-temporal eeg patterns, Neurocomputing 434 (2021) 137–148

    work page 2021

  7. [7]

    Vicente-Sola, D

    A. Vicente-Sola, D. L. Manna, P. Kirkland, G. Di Caterina, T. J. Bihl, Spiking neural networks for event-based action recognition: A new task to understand their advantage, Neurocomputing 611 (2025) 128657

    work page 2025

  8. [8]

    Bengio, Y

    Y . Bengio, Y . LeCun, et al., Scaling learning algorithms towards ai, Large-scale kernel machines 34 (5) (2007) 1–41

    work page 2007

Show all 39 references
  1. [9]

    G. F. Montufar, R. Pascanu, K. Cho, Y . Bengio, On the number of linear regions of deep neural networks, Advances in neural information processing systems 27 (2014)

    2014

  2. [10]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  3. [11]

    Y . Hu, H. Tang, G. Pan, Spiking deep residual networks, IEEE Transactions on Neural Networks and Learning Systems 34 (8) (2021) 5200–5205. 29

    2021

  4. [12]

    Sengupta, Y

    A. Sengupta, Y . Ye, R. Wang, C. Liu, K. Roy, Going deeper in spiking neural networks: Vgg and residual architectures, Frontiers in neuroscience 13 (2019) 425055

    2019

  5. [13]

    B. Han, G. Srinivasan, K. Roy, Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, 2020, pp. 13558–13567

    2020

  6. [14]

    Y . Wu, L. Deng, G. Li, L. Shi, Spatio-temporal backpropagation for training high-performance spiking neural networks, Frontiers in neuroscience 12 (2018) 323875

    2018

  7. [15]

    H. Shen, H. Wang, Y . Ma, L. Li, S. Duan, S. Wen, Multi-lra: Multi logical resid- ual architecture for spiking neural networks, Information Sciences 660 (2024) 120136

    2024

  8. [16]

    Y . Shan, X. Qiu, R.-j. Zhu, R. Li, M. Wang, H. Qu, Or residual connection achiev- ing comparable accuracy to add residual connection in deep residual spiking neu- ral networks, arXiv preprint arXiv:2311.06570 (2023)

    2023

  9. [17]

    Stöckl, W

    C. Stöckl, W. Maass, Optimized spiking neurons can classify images with high accuracy through temporal coding with two spikes, Nature Machine Intelligence 3 (3) (2021) 230–238

    2021

  10. [18]

    Tang, J.-H

    J. Tang, J.-H. Lai, X. Xie, L. Yang, W.-S. Zheng, Ac2as: Activation consistency coupled ann-snn framework for fast and memory-efficient snn training, Pattern Recognition 144 (2023) 109826

    2023

  11. [19]

    Z. Shao, X. Fang, Y . Li, C. Feng, J. Shen, Q. Xu, Eicil: Joint excitatory inhibitory cycle iteration learning for deep spiking neural networks, Advances in Neural Information Processing Systems 36 (2023) 32117–32128. 30

    2023

  12. [20]

    C. Duan, J. Ding, S. Chen, Z. Yu, T. Huang, Temporal effective batch normal- ization in spiking neural networks, Advances in Neural Information Processing Systems 35 (2022) 34377–34390

    2022

  13. [21]

    LeCun, D

    Y . LeCun, D. Touresky, G. Hinton, T. Sejnowski, A theoretical framework for back-propagation, in: Proceedings of the 1988 connectionist models summer school, V ol. 1, 1988, pp. 21–28

    1988

  14. [22]

    T. P. Lillicrap, A. Santoro, L. Marris, C. J. Akerman, G. Hinton, Backpropagation and the brain, Nature Reviews Neuroscience 21 (6) (2020) 335–346

    2020

  15. [23]

    Y . Hu, L. Deng, Y . Wu, M. Yao, G. Li, Advancing spiking neural networks to- ward deep residual learning, IEEE transactions on neural networks and learning systems 36 (2) (2024) 2353–2367

    2024

  16. [24]

    M. Yao, J. Hu, Z. Zhou, L. Yuan, Y . Tian, B. Xu, G. Li, Spike-driven transformer, Advances in neural information processing systems 36 (2023) 64043–64058

    2023

  17. [25]

    H. Xiao, K. Rasul, R. V ollgraf, Fashion-mnist: a novel image dataset for bench- marking machine learning algorithms, arXiv preprint arXiv:1708.07747 (2017)

    2017 arXiv

  18. [26]

    Krizhevsky, G

    A. Krizhevsky, G. Hinton, et al., Learning multiple layers of features from tiny images (2009)

    2009

  19. [27]

    Vinyals, C

    O. Vinyals, C. Blundell, T. Lillicrap, D. Wierstra, et al., Matching networks for one shot learning, Advances in neural information processing systems 29 (2016)

    2016

  20. [28]

    Zheng, Y

    H. Zheng, Y . Wu, L. Deng, Y . Hu, G. Li, Going deeper with directly-trained larger spiking neural networks, in: Proceedings of the AAAI conference on artificial intelligence, V ol. 35, 2021, pp. 11062–11070

    2021

  21. [29]

    Cheng, Y

    X. Cheng, Y . Hao, J. Xu, B. Xu, Lisnn: Improving spiking neural networks with lateral interactions for robust object recognition., in: IJCAI, Yokohama, 2020, pp. 1519–1525. 31

    2020

  22. [30]

    Zhang, P

    W. Zhang, P. Li, Spike-train level backpropagation for training deep recurrent spiking neural networks, Advances in neural information processing systems 32 (2019)

    2019

  23. [31]

    Y . Zhu, Z. Yu, W. Fang, X. Xie, T. Huang, T. Masquelier, Training spiking neural networks with event-driven backpropagation, Advances in Neural Information Processing Systems 35 (2022) 30528–30541

    2022

  24. [32]

    D. Zhao, Y . Zeng, Y . Li, Backeisnn: A deep spiking neural network with adaptive self-feedback and balanced excitatory–inhibitory neurons, Neural Networks 154 (2022) 68–77

    2022

  25. [33]

    Zhang, P

    W. Zhang, P. Li, Temporal spike sequence learning via backpropagation for deep spiking neural networks, Advances in neural information processing systems 33 (2020) 12022–12033

    2020

  26. [34]

    Hunsberger, C

    E. Hunsberger, C. Eliasmith, Spiking deep networks with lif neurons, arXiv preprint arXiv:1510.08829 (2015)

    2015 arXiv

  27. [35]

    Y . Cao, Y . Chen, D. Khosla, Spiking deep convolutional neural networks for energy-efficient object recognition, International Journal of Computer Vision 113 (2015) 54–66

    2015

  28. [36]

    Y . Wu, L. Deng, G. Li, J. Zhu, Y . Xie, L. Shi, Direct training for spiking neu- ral networks: Faster, larger, better, in: Proceedings of the AAAI conference on artificial intelligence, V ol. 33, 2019, pp. 1311–1318

    2019

  29. [37]

    Q. Zhan, B. Wang, A. Jiang, X. Xie, M. Zhang, G. Liu, A two-stage spiking meta-learning method for few-shot classification, Knowledge-Based Systems 284 (2024) 111220

    2024

  30. [38]

    Y . Qi, Y . Ban, T. Wei, J. Zou, H. Yao, J. He, Meta-learning with neural bandit scheduler, Advances in Neural Information Processing Systems 36 (2024). 32

    2024

  31. [39]

    S. Bag, S. K. Kumar, M. K. Tiwari, An efficient recommendation generation using relevant jaccard similarity, Information Sciences 483 (2019) 53–64. 33

    2019

Pith tools

Reviewed June 30, 2026 · model on record in the stance chip above.