acceptodds
Under review as a conference paper at ICLR 2027

UNISPIKE: Boosting the Performance of Spiking Neural Networks with Hybrid Training

Abstract

Spiking neural networks (SNNs) offer an efficient inference paradigm with two desirable properties: addition-only computation and step-by-step inference. Addition-only computation removes expensive multipliers and reduces arithmetic energy, while step-by-step inference enables temporally progressive outputs for faster response and early exiting. However, achieving high accuracy while preserving these properties remains challenging. Several accuracy-oriented SNN operators, such as SEW-addition, temporal attention, and temporal addition, improve representation learning by introducing multi-valued activations or temporal synchronization, thereby violating addition-only computation or step-by-step inference. To improve accuracy under these property-preserving constraints, we propose UniSpike, a hybrid-training framework that initializes an SNN from a quantized ANN and further fine-tunes it with back-propagation-through-time (BPTT). The key design is Unified Clip, a unified ANN-side activation that replaces operators incompatible with spike-based inference, including softmax, layer normalization, and GeLU. Its quantized form is equivalent to the ST-BIF spiking neuron, providing effective ANN-to-SNN initialization for SNN fine-tuning. Built upon Unified Clip, we design UniFormer, a spike-compatible vision transformer tailored for accurate, addition-only, and step-by-step SNN inference. On ImageNet-1K, UniSpike achieves 80.8% accuracy at 4 time-steps, outperforming the SOTA addition-only and step-by-step direct-training method Spike-Driven Transformer V2.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Spiking Nerual Network, Hybrid Training, Energy Efficiency

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

This work proposes UNISPIKE, a hybrid SNN training framework aiming to improving accuracy while preserving two important properties: addition-only computation and step-by-step inference. The framework first introduces Unified Clip as an ANN activation function to replace ANN operators that are not compatible with SNNs (e.g., softmax, LayerNorm, and GeLU). A SNN-friendly Transformer architecture called Uniformer is then built. The full training pipeline involves ANN pretraining, ANN quantization-aware training, ANN2SNN conversion, and BPTT-based SNN finetuning. Experiments on image classification tasks demonstrate competitive accuracy among SNNs with addition-only and step-by-step properties and low theoretical energy consumption during inference.

Critical strengths
  1. The motivation is sound. The importance of addition-only computation and step-by-step inference in SNNs are well justified.
  2. UNISPIKE presents a comprehensive training pipeline integrating pretraining, quantization, conversion and finetuning. It achieves good performance on image classification benchmarks.
  3. Training cost and early-exit strategies are carefully evaluated. The energy analysis is relatively comprehensive as it considers memory.
Critical weaknesses
  1. The equivalence between Unified Clip and ST-BIF in Equations (5) - (8) is not sufficiently justified, as the temporal unevenness error [1] is not considered here. Consider input x=[0.7,0.7,0.7,0.7]x=[0.7, 0.7, 0.7, 0.7] and x′=[0,0,0,2.8]x'=[0,0,0,2.8]. Here, Quan(∑txt)=Quan(∑txt′)=3\text{Quan}(\sum_t x_t)=\text{Quan}(\sum_t x'_t)=3, but ∑tSTBIF(xt)=3, ∑tSTBIF(xt′)=1\sum_t \text{STBIF}(x_t)=3,\ \sum_t \text{STBIF}(x'_t)=1 (assuming firing threshold and quantization step are both 1 ).
  2. Insufficient ablation. Figure 6 provides a partial comparison of neuron models, and Appendix A3 analyzes training costs. However, the individual contributions of the hybrid training pipeline and the UniFormer architecture to classification accuracies and energy consumption are not clearly disentangled.

[1] Bu et. al., Optimal ANN-SNN Conversion for High-accuracy and Ultra-low-latency Spiking Neural Networks, ICLR 2022.

Questions for authors
  1. (Weakness 1) Please clarify the required conditions under which the equivalence between Unified Clip and ST-BIF holds.
  2. (Weakness 2) What's the individual contribution of the UniFormer architecture and the hybrid training pipeline on classification accuracy, training cost and theoretical energy consumption? It's better to compare UniFormer trained directly with Uniformer trained through the hybrid training pipeline.
  3. The memory cost of QKFormer seems to be exaggerated. Appendix A4 assumes FP32 numerical activations for non-binary values and no operand-cache reuse. Under these assumptions, QKFormer's 48.09 GMACs alone imply approximately 240 mJ of activation reads, accounting for nearly all of the reported 247.7 mJ activation memory energy. However, practical implementations may exploit activation reuse and lower-precision representations for the integers introduced by SEW connections. Please clarify these assumptions and evaluate their impact on the energy comparison.
  4. Minor: qthrq_{\text{thr}} in Equation (2) is not defined and not used elsewhere. Its relationship with VthrV_{\text{thr}} and the quantization scale ss should be clarified.
  5. Minor: Page 4, Line 200, "for similarity" should be "for simplicity". xtx_t should be "input current". A careful language revision would improve clarity.
6 comments
AuthorsOpenReview ↗

Response to Reviewer qtXE (Part 1/4)

[Question 1] Please clarify the required conditions under which the equivalence between Unified Clip and ST-BIF holds.

We thank the reviewer for the careful analysis. The output of quantized Unified Clip is equivalent to the accumulated output of its converted ST-BIF neuron once the SNN reaches equilibrium, i.e., when no further input spikes are received and no additional output spikes are generated. As shown in Table R1, the L1 error falls below 1e-5 when the inference time-step exceeds 42. At this point, the SNN produces virtually no further output, with the L1 norm of its per-step output being zero.

Table R1: Conversion results of UniSpike before BPTT training. L1 Error denotes the average L1 distance between the SNN output at each time-step and the corresponding QANN output for the same input, averaged over 1,000 randomly selected ImageNet-1K test images.

AccuracyL1 Norm for SNN OutputL1 Error
QANN81.6%--
SNN T=12.1%139.416430427.347134
SNN T=235.1%214.116407494.385983
SNN T=466.5%132.484945516.086995
SNN T=880.9%90.546196241.131246
SNN T=1282.0%38.644254127.026343
SNN T=1682.0%29.199896108.070606
SNN T=3281.7%2.4126632.057155
SNN T=4181.6%1.23828e-32.88151e-4
SNN T=4281.6%2.88151e-43.74996e-7
SNN T=4381.6%0.03.74996e-7
AuthorsOpenReview ↗

Response to Reviewer qtXE (Part 2/4)

[Question 2] What's the individual contribution of UniFormer architecture and the hybrid training pipeline on classification accuracy, training cost, and theoretical energy consumption?

We provide controlled ablation as reviewer acquired.

1) Ablation for UniFormer:

As shown in Table R2, progressively replacing the original ANN operators results in an overall accuracy drop of only 0.5% (82.8% to 82.3%). Specifically, replacing GeLU with ReLU reduces accuracy by 0.4% (82.8% to 82.4%), while further replacing LayerNorm with ReLU causes an additional 0.5% drop (82.4% to 81.9%). Interestingly, replacing Softmax with Unified Clip recovers 0.4% accuracy (81.9% to 82.3%), suggesting that Unified Clip is better suited to the unified activation design in the attention block.

2) Ablation of Hybrid Training:

Accuracy Analysis: As shown in Table R3, hybrid training achieves the highest accuracy of 80.8%, outperforming direct training (54.3%) and conversion alone (66.5%) by 26.5% and 14.3%, respectively. Direct training suffers from poor random initialization and inaccurate surrogate-gradient estimation during BPTT, hindering effective optimization. ANN-to-SNN conversion provides a better initialization, while subsequent BPTT fine-tuning further improves accuracy, demonstrating the effectiveness of the proposed hybrid training pipeline.

Energy and Training Cost: Higher accuracy requires more neuronal firing, resulting in the highest SOPs and inference energy in hybrid training. Direct training takes 327.8 hours, as 4 time-step SNN training incurs approximately 4× the computation per epoch of ANN training. In contrast, hybrid training requires only 10 epochs of SNN fine-tuning, reducing the total training time to 65.00 hours (80.2% reduction).

Table R2: The Ablation Study for operator replacement in UniFormer architecture on ANN pre-training stage. All the models are trained from scratch for 100 epochs under the same training configuration, which differs from the 300-epoch setting used in the main paper.

Accuracy
Naive82.8%
+Replace GeLU82.4%
+Repalce Layernorm81.9%
+Replace Softmax (UniFormer)82.3%

Table R3: The Ablation Study for the hybrid training pipeline. Training Time is estimated on 8 RTX 4090 GPUs. We use torch.compile to further accelerate the training in the direct training method.

ModelTime-StepSOP(G)Energy(mJ)Epoch of Stage1/Stage2/Stage3Training Time (h)Accuracy
Direct TrainingUniFormer416.011121.9300/0/0327.854.3%
ConversionUniFormer418.377136.0300/40/051.3466.5%
Hybrid TrainingUniFormer419.009138.2300/40/1065.0080.8%
AuthorsOpenReview ↗

Response to Reviewer qtXE (Part 3/4)

[Question 3] The memory cost of QKFormer seems to be exaggerated.

We thank the reviewer for the careful examination. We profile QKFormer's SEW-shortcut activations and find that their values range from 0 to 16, requiring only 5-bit unsigned integers (UInt5). With this reduced precision, the activation memory energy decreases from 247.7 mJ to 50.36 mJ, reducing total energy from 726.3 mJ to 528.9 mJ (Table R4). Nevertheless, QKFormer still consumes substantially more energy than UniFormer (138.2 mJ), with weight memory and computation accounting for 90.3% of its total energy.

Table R4: Energy breakdown of SDT-V2, QKFormer, QKFormer with 5-bit SEW-shortcut activations, and UniFormer.

Energy(mJ)SDT-V2QKFormerQKFormer with INT5 SEW-ShortCut ActivationUniFormer
Weight Memory282.2 (75.9%)254.1 (35.0%)254.1 (48.0%)90.90 (65.8%)
Activation Memory37.99 (10.2%)247.7 (34.1%)50.36 (9.52%)26.99 (19.5%)
Neuron State Memory1.030 (0.28%)0.971 (0.13%)0.971 (0.18%)3.194 (2.31%)
Computation50.68 (13.6%)223.5 (30.8%)223.5 (42.3%)17.11 (12.4%)
Total371.9 (100%)726.3 (100%)528.9 (100%)138.2 (100%)

Regarding activation reuse, the same optimization can be applied to all SNN architectures; thus, it would benefit all compared architectures similarly. Moreover, after INT5 optimization, activation memory accounts for only 9.52% of QKFormer's total energy, while weight memory and computation remain the dominant costs. Therefore, excluding activation reuse has limited impact on our energy comparison.

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

This paper aims to achieve high accuracy in spiking neural networks (SNNs) while preserving two desirable properties: addition-only computation and step-by-step inference. To this end, the authors adopt a hybrid training scheme in which an SNN is first initialized through ANN-to-SNN conversion using an architecturally equivalent ANN, followed by training with BPTT. To preserve the two desired properties, the authors propose Unified Clip to replace operations in ANNs that violate these properties, along with a corresponding Transformer-based architecture called UniFormer. Experiments on standard benchmarks demonstrate that the proposed methods achieve improved accuracy and energy efficiency.

Critical strengths
  1. The proposed Unified Clip is simple and effective, providing a unified approach to replacing operations in ANNs that are incompatible with addition-only computation and step-by-step inference in SNNs.

  2. The experimental results are promising, and the additional analyses provide useful insights for the proposed methods.

Critical weaknesses
  1. Hybrid training that combines ANN-to-SNN conversion with subsequent BPTT training has already been explored in previous studies. If hybrid training is intended to be the main contribution of this paper, as suggested by the title, its novelty appears rather limited.

  2. In my view, addition-only computation and step-by-step inference are desirable properties of SNNs in general, rather than properties specifically associated with hybrid training. Moreover, the proposed Unified Clip and UniFormer do not appear to be closely tied to the hybrid training scheme. It is therefore unclear why the proposed methods are specifically motivated and presented within the context of hybrid training.

  3. Several previous studies on ANN-to-SNN conversion have explored similar ideas as Unified Clip, so the novelty of Unified Clip appears to be limited.

  4. Replacing standard ANN operations with Unified Clip introduces accuracy drop before conversion, which is undesirable.

  5. The paper contains hallucinated references, such as the citation of MS-shortcut (Han et al., 2022) in Line 235.

Questions for authors
  1. Why do the softmax and layer normalization operations require access to spikes from all time steps?

  2. How much accuracy degradation is introduced in the original ANNs when standard operations are replaced with Unified Clip?

  3. Can the proposed method be generalized to arbitrary Transformer-based architectures?

6 comments
AuthorsOpenReview ↗

Response to Reviewer bKMW (Part 1/6)

[Question 1] Why do the softmax and layer normalization operations require access to spikes from all time steps?

In SNNs, the inputs to Softmax and LayerNorm in the QANN cannot be fully obtained within a single time-step, because each QANN activation is represented by the accumulated spike outputs across all SNN time steps. Therefore, correctly computing these operations requires waiting until the corresponding activations are fully accumulated.

Specifically, Softmax requires attention scores across all tokens to compute the normalization denominator, while LayerNorm requires all feature values to calculate the mean and variance. Therefore, both operations must wait until the accumulation across all time-steps is complete, preventing step-by-step computation.

[Question 2] How much accuracy degradation is introduced in the original ANNs when standard operations are replaced with Unified Clip?

We provide the ANN accuracy with/without operator replacement in Table R1. After replacing all three operators with Unified Clip, the accuracy degradation is only 0.5% (82.8% → 82.3%), showing that UniFormer maintains most of the model's accuracy on ImageNet-1K.

Table R1: Accuracy Comparison UniFormer with/without operator replacement on ImageNet-1K. All the models are trained from scratch for 100 epochs under the same training configuration, which differs from the 300-epoch setting used in the main paper.

OperatorAcc
OriginalSoftmax; LayerNorm; GeLU82.8%
UniFormerUnified Clip82.3%
AuthorsOpenReview ↗

Response to Reviewer bKMW (Part 2/6)

[Question 3] Can the proposed method be generalized to arbitrary Transformer-based architectures?

To evaluate the generalizability of UniSpike, we apply it to a representative Transformer architecture, vision transformer (ViT) [1]. As shown in Table R2, UniSpike achieves 79.3% accuracy on ViT, with only 4 time-steps. These results demonstrate that UniSpike is not limited to UniFormer and can generalize to other Transformer architectures while preserving accuracy.

Table R2: Evaluation of the UniSpike Framework on ViT and UniFormer Architectures.

Model Arch.Parameter(M)Time-StepANNQANNSNN
ViT86.0481.6%80.2%79.3%
UniFormer66.5483.3%81.6%80.8%

[1] Dosovitskiy A. An image is worth 16x16 words: Transformers for image recognition at scale[J]. arXiv preprint arXiv:2010.11929, 2020.

AuthorsOpenReview ↗

Response to Reviewer bKMW (Part 3/6)

[Weakness 1] Hybrid training that combines ANN-to-SNN conversion with subsequent BPTT training has already been explored in previous studies. If hybrid training is intended to be the main contribution of this paper, as suggested by the title, its novelty appears rather limited.

We do not claim to propose a new hybrid training paradigm. Instead, our contribution is making hybrid training work for Transformers while preserving high accuracy, addition-only computation, and step-by-step inference.

Unlike previous HT methods [1], [2], which focus on CNNs and convert ReLU into IF/LIF neurons, UniSpike addresses Transformer-specific operators, including Softmax, LayerNorm, and GELU, which are incompatible with the desired SNN inference properties. We replace these operators with Unified Clip, whose quantized form is mathematically equivalent to ST-BIF neurons, enabling accurate ANN-to-SNN conversion and effective SNN initialization.

Table R3 further validates the importance of this equivalence. Before BPTT training, ST-BIF-based conversion achieves substantially higher accuracy than IF-based conversion across all evaluated time-steps, recovering the QANN accuracy (81.6%) at T=42. This demonstrates that our contribution lies not in the HT paradigm itself, but in making it effective for Transformer-based SNNs.

Table R3: Conversion Results of UniSpike with Different Spiking Neurons Before BPTT Training.

QANNT=1T=2T=4T=8T=12T=16T=32T=42
ST-BIF81.6%2.1%35.1%66.5%80.9%82.0%82.0%81.7%81.6%
IF81.6%0.1180%0.1180%0.1340%0.1060%0.1120%0.1000%0.1000%0.1000%
LIF81.6%0.1160%0.1160%0.1040%0.1000%0.1000%0.1000%0.1000%0.1000%

[1] Rathi N, Srinivasan G, Panda P, et al. Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation[J]. arXiv preprint arXiv:2005.01807, 2020.

[2] Baltes M, Abuhajar N, Yue Y, et al. Joint ann-SNN co-training for object localization and image segmentation[C]//ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023: 1-5.

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

This paper proposes UniSpike, a hybrid-training framework for Transformer-based SNNs. It introduces Unified Clip to replace Softmax, LayerNorm, and GeLU, and converts the quantized ANN into an SNN using ST-BIF neurons, followed by fine-tuning. The method achieves 80.81% ImageNet-1K accuracy at four time steps and reports reduced inference energy.

Critical strengths
  1. The paper improves SNN accuracy while preserving addition-only computation and step-by-step inference.
  2. The results are competitive, particularly the 80.81% ImageNet accuracy at four time steps. The evaluation also covers smaller datasets and early-exit inference.
Critical weaknesses
  1. Incremental novelty. Several key components build on existing work. Hybrid ANN-to-SNN training was explored by Rathi et al. (ICLR 2020), the ST-BIF/quantization equivalence is closely related to SpikeZIP-TF (ICML 2024), and MS-Shortcut and spike-driven attention are also established techniques. Unified Clip is essentially an affine transformation followed by clipping. While combining these techniques is useful, the paper does not clearly establish what is fundamentally new compared with prior work.
  2. Insufficient evidence for the claimed improvements. The accuracy gain over SDFormer V2 is only 0.81, with a larger model (66.5M vs. 55.4M parameters). There is no controlled comparison of direct training, conversion-only, and hybrid training using the same backbone. Also, the claimed equivalence under four time steps is unclear given the 10-level quantization.
Questions for authors
  1. What is the main technical novelty beyond SpikeZIP-TF and existing hybrid-training methods? Can the authors provide a controlled ablation showing the individual contributions of Unified Clip and hybrid training?
  2. How does the claimed exact equivalence hold with 10 quantization levels but only four inference time steps?
3 comments
AuthorsOpenReview ↗

Response to Reviewer iMSL (Part 1/3)

[Question 1-1] What is the main technical novelty beyond SpikeZIP-TF and existing hybrid-training methods

Our contribution is making hybrid training work for Transformers while achieving high accuracy and maintaining high energy efficiency, which these components alone do not achieve.

vs. prior hybrid training (HT). Existing HT methods [1], [2] only convert ReLU activations into IF/LIF neurons in CNNs, but do not support transformer-specific operations such as Softmax, LayerNorm, and GELU, resulting in inaccurate Artifical Neural Network(ANN)-to-Spiking Neural Network(SNN) conversion. We replace these operators with functions (Quantized Unified Clip) that equivalent to ST-BIF neuron. As shown in Table R1, UniFormer converted by ST-BIF neuron has the highest accuracy across all time-steps, demonstrating better initialization in UniSpike HT framework.

vs. SpikeZIP-TF. SpikeZIP-TF is conversion-only and keeps Softmax, LayerNorm, and GELU. We (i) replace all three operators with one element-wise activation (Unified Clip) whose quantized form is provably equivalent to ST-BIF given enough time-steps, costing only 0.5% ANN accuracy (82.8% → 82.3%) in Table R2, and (ii) derive BPTT for ST-BIF with its spike tracer, which reduces time-steps from 64 to 4.

To our knowledge, this is the first Transformer-based hybrid-training SNN validated on ImageNet-1K.

Table R1: Conversion Results of UniSpike with Different Spiking Neurons Before BPTT Training.

QANNT=1T=2T=4T=8T=12T=16T=42
ST-BIF81.6%2.1%35.1%66.5%80.9%82.0%82.0%81.6%
IF81.6%0.1180%0.1180%0.1340%0.1060%0.1120%0.1000%0.1000%
LIF81.6%0.1160%0.1160%0.1040%0.1000%0.1000%0.1000%0.1000%

Table R2: Accuracy Comparison UniFormer with/without operator replacement on ImageNet-1K. All the models are trained from scratch for 100 epoch under the same training configuration, which differs from the 300-epoch setting used in the main paper.

OperatorAccuracy
OriginalSoftmax; LayerNorm; GeLU82.8%
UniFormerUnified Clip82.3%

[1] Rathi N, Srinivasan G, Panda P, et al. Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation[J]. arXiv preprint arXiv:2005.01807, 2020.

[2] Baltes M, Abuhajar N, Yue Y, et al. Joint ann-SNN co-training for object localization and image segmentation[C]//ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023: 1-5.

AuthorsOpenReview ↗

Response to Reviewer iMSL (Part 2/3)

[Question 1-2] Can the authors provide a controlled ablation showing the individual contributions of Unified Clip and hybrid training?

We provide controlled ablation as reviewer acquired.

1) Ablation for Unified Clip: As shown in Table R3, progressively replacing the original ANN operators results in an overall accuracy drop of only 0.5% (82.8% to 82.3%). Specifically, replacing GeLU with ReLU reduces accuracy by 0.4% (82.8% to 82.4%), while further replacing LayerNorm with ReLU causes an additional 0.5% drop (82.4% to 81.9%). Interestingly, replacing Softmax with Unified Clip recovers 0.4% accuracy (81.9% to 82.3%), suggesting that Unified Clip is better suited to the unified activation design in the attention block.

2) Ablation of Hybrid Training: As shown in Table R4, hybrid training achieves the highest accuracy of 80.8%, outperforming direct training (54.3%) and conversion alone (66.5%) by 26.5% and 14.3%, respectively. Direct training suffers from poor random initialization and inaccurate surrogate-gradient estimation during BPTT, hindering effective optimization. ANN-to-SNN conversion provides a better initialization, while subsequent BPTT fine-tuning further improves accuracy, demonstrating the effectiveness of the proposed hybrid training pipeline.

Table R3: Ablation for Unified Clip replacement in ANN pre-training. All the models are trained from scratch for 100 epoch under the same training configuration, which differs from the 300-epoch setting used in the main paper.

Accuracy
Naive82.8%
+Replace GeLU82.4%
+Repalce LayerNorm81.9%
+Replace Softmax (UniFormer)82.3%

Table R4: Ablation for hybrid training:

Accuracy
Direct Training54.3%
Conversion66.5%
Hybrid Training80.8%
AuthorsOpenReview ↗

Response to Reviewer iMSL (Part 3/3)

[Question 2] How does the claimed exact equivalence hold with 10 quantization levels but only four inference time steps?

We apologize for the misleading wording. As proven in [1], a quantized ANN (QANN) and its converted SNN produce equivalent outputs once the SNN reaches equilibrium, i.e., when no further input spikes are provided, and no additional output spikes are generated. As shown in Table R5, the QANN and SNN have the same output (L1 Error < 1e-5), when the SNN output is zero. Nevertheless, as shown in Table R1, UniSpike achieves higher initial accuracy immediately after conversion, before hybrid training, than SNNs converted using the IF/LIF neurons adopted in previous HT methods [2], [3], demonstrating better weight initialization.

Table R5: Conversion results of UniSpike before BPTT training. L1 Error denotes the average L1 distance between the SNN output at each time-step and the corresponding QANN output for the same input, averaged over 1,000 randomly selected ImageNet-1K test images.

AccuracyL1 Norm for SNN OutputL1 Error
QANN81.6%--
SNN T=12.1%139.416430427.347134
SNN T=235.1%214.116407494.385983
SNN T=466.5%132.484945516.086995
SNN T=880.9%90.546196241.131246
SNN T=1282.0%38.644254127.026343
SNN T=1682.0%29.199896108.070606
SNN T=3281.7%2.4126632.057155
SNN T=4181.6%1.23828e-32.88151e-4
SNN T=4281.6%2.88151e-43.74996e-7
SNN T=4381.6%0.03.74996e-7

[1] You K, Xu Z, Nie C, et al. Spikezip-tf: Conversion is all you need for transformer-based snn[J]. arXiv preprint arXiv:2406.03470, 2024.

[2] Rathi N, Srinivasan G, Panda P, et al. Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation[J]. arXiv preprint arXiv:2005.01807, 2020.

[3] Baltes M, Abuhajar N, Yue Y, et al. Joint ann-SNN co-training for object localization and image segmentation[C]//ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023: 1-5.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.