TS-SNN: Temporal Shift Module for Spiking Neural Networks

Questions for Authors

Performance Discrepancy for ResNet-19 on CIFAR-100 (Timestep=1)

Thank you for pointing out this discrepancy. As you correctly observed, the performance of our TS-SNN with ResNet-19 on CIFAR-100 under a single timestep lags slightly behind MPBN, whereas it outperforms MPBN on CIFAR-10. This behavior can be attributed to the nature of the Temporal Shift (TS) operation when T=1. In this case, there are no past or future information to reference, and the shift operation defaults to zero-padding. The resulting update becomes a residual connection:

$Z' = \alpha Z + X$.

As a result, the TS operation brings limited benefit when T=1, and the model's performance largely relies on the backbone architecture and input statistics. Notably, CIFAR-100 is a more challenging dataset with fewer samples per class, which amplifies the limitations of reduced temporal modeling under T=1. This observation demonstrates that the primary advantage of our method arises from the effective feature fusion brought by the temporal shift operations across multiple timesteps. We will include this clarification in the camera-ready version of the paper.

---

Weakness

1. Lack of Discussion on Limitations

We appreciate your suggestion. We will include two primary limitations of our method:

• Restricted Effectiveness at Low Timesteps
  As mentioned above, when T=1, the TS module cannot utilize temporal context, limiting its effectiveness. The model reverts to near-baseline behavior, and performance gains are minimal or inconsistent.

• Memory Requirements
  The TS operation assumes access to spike outputs from both past and future timesteps within a local window. On neuromorphic hardware, this may require additional memory and access latency to buffer or store spike states across timesteps. While this overhead is generally modest compared to the savings in computation and energy (since spike buffers are lightweight), it could become a bottleneck for highly constrained or streaming systems. We plan to explicitly discuss these limitations and potential mitigations in the camera-ready version.

---

2. Performance Discrepancy in Table 1

Please see our response above under ##Questions for Authors## for a detailed explanation.

---

Other Comments and Suggestions

Formatting Issues

We appreciate your feedback on formatting. We have revised the LaTeX alignment of all equations to ensure consistency—either center alignment or equation numbering, depending on context. This significantly improves the visual quality and readability of the manuscript.

---

We sincerely thank the reviewer for their thoughtful and helpful suggestions. We will incorporate all the suggested changes into the camera-ready version of the paper.

Thank you for your detailed review and constructive feedback. We address your points as follows:

---

Questions for Authors

How TS operation when timestep = 1

As you correctly observed, the performance of our TS-SNN with ResNet-19 on CIFAR-100 under a single timestep lags slightly behind MPBN, whereas it outperforms MPBN on CIFAR-10. This behavior can be attributed to the nature of the TS operation when T=1. In this case, there are no past or future information to reference, and the shift operation defaults to zero-padding, leaving only the residual connection $Z' = \alpha Z + X$ active. Although this residual fusion can subtly affect feature extraction, it does not consistently enhance model stability. This observation demonstrates that the primary advantage of our method arises from the effective feature fusion brought by the temporal shift operations across multiple timesteps. We will include this clarification in the camera-ready version of the paper.

---

Weakness

1. Theoretical Analysis and Experimental Verification

We appreciate your observation. In the camera-ready version, we will incorporate theoretical insights that will substantially strengthen the paper, including:

• Expansion of the Temporal Receptive Field
  By shifting features forward and backward, the TS module enables neurons to access information from adjacent timesteps, which is effectively allowing each neuron to "see" more temporal context. This operation expands the temporal receptive field from $d$ to $2d + 1$ without incurring extra parameters or computational cost.

• Increase in Mutual Information and Entropy
  The TS module fuses inputs ${X_{t-1}, X_t, X_{t+1}}$ to compute $S_t$, thereby increasing its temporal entropy $H(S_t)$. Under stationary input conditions, the mutual information $I(S_t; X_{t-1:t+1})$ exceeds $I(S_t; X_t)$, demonstrating enhanced temporal context modeling.

---

2. Implementation on Neuromorphic Hardware

Your concern that the TS module may incur additional memory costs on neuromorphic chips is reasonable. However, relative to the significant energy efficiency gains demonstrated by our approach, the extra storage requirement is minimal. The hardware optimization strategies will be promising in future work to further mitigate this cost in practical deployments.

---

3. Setting of the $\alpha$ Parameter

We appreciate your insightful observation regarding the $\alpha$ parameter. In our implementation details, we mistakenly stated that the initial value of $\alpha$ is 0.5 for all datasets. The correct settings, as detailed in the supplementary material, are as follows: for CIFAR-10 and CIFAR-100, the initial $\alpha$ is 0.5; for ImageNet and CIFAR10-DVS, the initial $\alpha$ is 0.2. We will correct this error in the camera-ready version.

Furthermore, we specified in the manuscript that the experimental range for$\alpha$ is between 0.2 and 0.5. Our experiments validated the effectiveness of this range across different datasets. We found that the initial value of $\alpha$ is critical—if it is set above 0.5, training tends to collapse. Therefore, we recommend tuning $\alpha$ as a hyperparameter within the 0.2–0.5 range based on the specific dataset.

---

4. Training Epochs and Convergence

Although our experiments on CIFAR-10 and CIFAR-100 used 500 epochs, we observed that competitive performance can be achieved with as few as 200 epochs—the performance difference is within 0.4% as shown in the table below. We chose 500 epochs to align with some current SOTA settings and to ensure robust convergence. In the camera-ready version, we will include this experiments to clarify this.

---

Other Comments and Suggestions

Thank you for your suggestion regarding Figures 5 and 6. In the camera-ready revision, we will update these figures to include baseline data (i.e., models without the TS module), clearly demonstrating the performance improvements attributable solely to the TS module.

---

We greatly appreciate your valuable advice, which will help enhance the quality and clarity of our work.

Thank you for this insightful question. Indeed, one of the advantages of the Temporal Shift module is its lightweight and general design, making it applicable beyond classification.

While our current work focuses on image classification due to space constraints and the need for standardized comparisons with prior SNN models, we agree that the TS module has strong potential in other tasks such as object detection, semantic segmentation, or event-based video understanding.

In future work, we plan to extend TS-SNN to such domains. Additionally, the TS module is designed to be plug-and-play and architecture-agnostic, so it can be seamlessly integrated into models for these tasks with minimal modifications.

We appreciate the suggestion and will include a discussion of this in the camera-ready version.