SEENN: Towards Temporal Spiking Early-Exit Neural Networks

Spiking Neural Networks (SNNs) have recently become more popular as a biologically plausible substitute for traditional Artificial Neural Networks (ANNs). SNNs are cost-efficient and deployment-friendly because they process input in both spatial and temporal manner using binary spikes. However, we observe that the information capacity in SNNs is affected by the number of timesteps, leading to an accuracy-efficiency tradeoff. In this work, we study a fine-grained adjustment of the number of timesteps in SNNs. Specifically, we treat the number of timesteps as a variable conditioned on different input samples to reduce redundant timesteps for certain data. We call our method Spiking Early-Exit Neural Networks (SEENNs). To determine the appropriate number of timesteps, we propose SEENN-I which uses a confidence score thresholding to filter out the uncertain predictions, and SEENN-II which determines the number of timesteps by reinforcement learning. Moreover, we demonstrate that SEENN is compatible with both the directly trained SNN and the ANN-SNN conversion. By dynamically adjusting the number of timesteps, our SEENN achieves a remarkable reduction in the average number of timesteps during inference. For example, our SEENN-II ResNet-19 can achieve 96.1% accuracy with an average of 1.08 timesteps on the CIFAR-10 test dataset. Code is shared at https://github.com/Intelligent-Computing-Lab-Yale/SEENN.

Paper

References (64)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Sp8z5/10 · confidence 4/52023-07-04

Summary

This paper considers to adaptively determine inference time steps of spiking neural networks to improve the tradeoff between accuracy and time. Two methods are proposed. The first one uses confidence score thresholding. The second one introduces an additional policy network to predict the number of timesteps by reinforcement learning. Experiments show the effectiveness of the proposed methods for direct SNN training and ANN-SNN conversion methods.

Strengths

1. This paper considers early-exit of spiking neural networks, which can effectively improve the efficiency of SNNs. The idea of dynamically determining inference time steps based on the difficulty of inputs is interesting. 2. Extensive experiments on static and neuromorphic datasets as well as qualitative assessment are conducted.

Weaknesses

1. Dynamic inference time steps of SNNs, especially based on confidence scores, were also explored by (possibly concurrent) recent works [1,2], which can be discussed. The idea of confidence score is very simple and straight-forward. 2. For SEENN-II, an additional policy network is required for inference, and it is unclear how this can be deployed. From the descriptions (and codes), the policy network is an ANN rather than SNN, so it is not compatible with the main SNN for hardware deployment. This poses challenges and also arouses questions of why considering such hybrid architectures. Additionally, SEENN-II seems not to be flexible, i.e. after training it cannot make tradeoff between accuracy and time/energy. 3. From the descriptions, it is not clear enough whether the energy consumption estimation of SEENN-II consider policy networks. This should be included and more detailedly discussed. For example, on ImageNet, is the policy network an ANN ResNet-34 (as shown in code)? Then it may consume more than the SNN part. The energy result on ImageNet is missing. [1] Wu, D., Jin, G., Yu, H., Yi, X., & Huang, X. (2023). Optimising Event-Driven Spiking Neural Network with Regularisation and Cutoff. arXiv preprint arXiv:2301.09522. [2] Li, C., Jones, E., & Furber, S. (2023). Unleashing the Potential of Spiking Neural Networks by Dynamic Confidence. arXiv preprint arXiv:2303.10276.

Questions

Can the authors discuss more about conceptually how to determine the inference of (deep) SNNs (in small time steps) with confidence signals from the top layer? Note that the common time steps of SNNs simply refer to time steps for each layer and do not consider propagation across layers. However, if we consider asynchronous neurons, e.g. on neuromorphic hardware, it also takes time to propagate signals across layers, and the actual delay may be layer number plus time steps per layer. For deep SNNs (e.g. >18 layers in this paper), it takes much time to obtain the confidence score from the top layer, which may be longer than the small time steps per layer. Then how can it effectively control such small time steps?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The authors do not discuss limitations and societal impact.

Reviewer DV4t7/10 · confidence 5/52023-07-05

Summary

This paper proposes an accuracy-efficiency tradeoff method by adjusting the number of timesteps in SNNs, which is new and interesting. The authors accomplish this idea with two methods that uses a confidence score thresholding and reinforcement learning. These methods can be applied to directly training SNN and the ANN-SNN both, and results show these methods can save energy greatly but with negligible accuracy decreasing. I like the work that has potential and would provide a new direction for the following SNN work.

Strengths

1. well-written, easy to read. 2.The idea to adjust the number of timesteps to balance accuracy-efficiency is new and interesting. 3. The method is simple and easy to follow. 4. Experimental results are really good. The method can keep similar performance while reduce cost.

Weaknesses

To show the effectiveness of the method. A similar timesteps for other methods should be provided, for example, 2 or 3 timesteps for TET on ImageNet.

Questions

1.Can this method be used to other backbones, like transformer? 2.What is the performance of the method using vanilla CE loss?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

1.I find no potential negative societal impact.

Reviewer J7UY7/10 · confidence 5/52023-07-05

Summary

This paper introduces a new inference scheme for spiking neural networks (SNNs), the early exit on the time dimension. To make sure that relatively easy images can be predicted with less number of timesteps, this work build two frameworks to identify the appropriate timestep to minimize the latency while maintaining decent performance. The first method is an on-the-fly approach and the second one is more complicated that requires finetuning. The authors test their SEENN on various recognition benchmarks and obtain quite good accuracy with even lower number of timesteps compared to existing papers.

Strengths

1. Compared to other SNNs papers that focus on either conversion and training, what is work studied is a universal approach (i.e. time) that can be applied to any types SNNs. 2. The methodology is technically sound and covers different user resources (whether finetune or just plug-and-play) 3. Experiments results is thorough and solid and verify the effectiveness of this approach.

Weaknesses

1. There is no empirical evidence showing that the confidence score will increase along with the number of timesteps. The authors are suggested to add a figure to show the confidence score evolution. The above limitations has been explained in the rebuttal.

Questions

Can SEENN-II be applied to the conversion approach? if not, why is that, please elaborate it. I suggest the following response about this problem is discussed into the revision.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

see above.

Reviewer ZThW6/10 · confidence 3/52023-07-24

Summary

This paper proposes a novel manner to improve the efficiency of the spiking neural network. Specifically, the SEENN is proposed to determine the appropriate number of timesteps, which therefore reduce the inference time-cost. The proposed method is evaluated on CIFAR-10 and Imagenet, achieving good performance.

Strengths

+ The main contribution of this paper is to treat the number of timesteps as a variable in SNN model. Accordingly, several variations of early-exit manner is deigned for better accuracy-efficiency tradeoff. + The paper writing and organization is good.

Weaknesses

- The proposed method for determining the best timesteps is not novel. The first manner is to simply set the right number with the confidence score, while the second one introduces a policy network for better prediction. Although above two manners could work, these would inevitably introduce extra computation cost or human-based prior, which is not expected. In addition, the proposed SNN model still rely heavily on the ANN backbones, e.g., Resnet or VGGnet. - As for the hardware efficiency, the adopted nvidia V 100 maybe not proper, which is not designed for SNN application. - The ablation study is not clear, it seems the SEEN adopts different backbones with SNN.

Questions

See weakness.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

None.

Reviewer J7UY2023-08-12

Thanks for the responses, my concerns have been resovled. I appreciate the additional experiment and the discussion in the above response, the corresponding content should be added into the revision if accepted.

Reviewer Sp8z2023-08-15

I would like to thank the authors for their detailed responses. Most of my questions are addressed and I will raise my score. Some additional comments: 1. Directly trained SNNs can be tested under different timesteps for trade-off and some previous works reported such accuracy [1,2]; 2. The design to send the next input data to the pipeline requires frequent memory exchange and additional memory (e.g., for membrane potentials), which can cause much additional energy consumption. [1] Temporal efficient training of spiking neural network via gradient re-weighting. ICLR 2022. [2] Online training through time for spiking neural networks. NeurIPS 2022.

Authorsrebuttal2023-08-18

Authors reply

We'd like to thank reviewer Sp8z for his appreciation and genuine discussion of our work. We'd like to discuss the two comments raised by reviewer Sp8z. (1) *Directly trained SNNs can be tested under different timesteps for trade-off and some previous works reported such accuracy [1,2]* Reply: Yes, technically directly trained SNNs can be tested with other numbers of timesteps. Nonetheless, we emphasize that these two works report did not report those accuracies when they are compared against other paper. For example, TET[1] has shown its ability to adjust $T$ in Figure 4, but the model trained with $T=4$ only gets 70.5% accuracy evaluated at $T=2$. However, they report 72.87% accuracy in their main table (Table 3). OTTT[2] also has a huge accuracy gap (7%) between $T=2$ and $T=6$, and they did not compare $T=2$ results with their current state-of-the-art methods like tdBN, TET. Having said that, our emphasis is that the results we presented in Table 1 are evaluated under a fair setting with existing methods. (2) *The design to send the next input data to the pipeline requires frequent memory exchange and additional memory (e.g., for membrane potentials), which can cause much additional energy consumption.* Reply: We agree with the reviewer on this matter. Let's formulate this entire question from the beginning. In such a pipeline inference architecture, we can describe the overall latency as $$Latency(T) = C + \Delta (T-1),$$ where $C$ is a constant representing the latency to finish all layer's computation and $\Delta$ is the latency interval between any two timesteps. As the reviewer suggested, if we have a very deep network (i.e. $C$ is large) under a low number of timesteps (i.e., max $T$ is small, for example 4), SEENN-I cannot save much energy. This is true. However, the problem is under such cases, any timesteps reduction is not meaningful since $Latency(4)-Latency(1)$ has only $3\Delta$ magnitude, which is a lot lower than $C$. One may just use the full timesteps to utilize the full performance of SNNs. If the max $T$ is a lot higher, for example, $>500$, which is often the case for this type of hardware, timesteps reduction can bring a significant energy reduction. The ratio of energy saving is $$Ratio = \frac{C + \Delta (\hat{T}-1)}{C+\Delta(T-1)}$$ where $\hat{T}$ is the average number of timesteps for all test samples in SEENN-I. To summarize, the relationship between $C$, $\Delta$, and max $T$ will impact the overall saving ratio. The choice of $C$ is a hardware design problem and decides the upper limit of the acceleration ratio in SEENN-I or other timestep reduction work. In a few extreme cases ($C \gg \Delta \times T$), timestep reduction brings a limited advantage. This hardware design problem is somewhat beyond the scope of our work as it impacts the common algorithmic aspects in SNNs like network architecture and the choice of $T$.

Authorsrebuttal2023-08-18

Authors reply

Thank you again for your time in reviewing our article. Given that the deadline for the author-reviewers discussion is approaching, we want to kindly remind you to let us know if you have further questions or concerns about our rebuttal. We welcome any further feedback, comments, or questions and we are open to discussion.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC