Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

ECC-SNN: Cost-Effective Edge-Cloud Collaboration for Spiking Neural Networks

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read ECC-SNN claims that an entropy-gated edge SNN backed by a cloud ANN teacher achieves 4.15% higher accuracy with 79.4% less energy and 39.1% lower latency than standalone baselines.

desk verdict A coherent integration of distillation, entropy-based offloading, and on-device incremental learning for SNN edge inference—but the headline energy/latency savings are backed by no usable cost model. read the letter →

arxiv 2505.20835 v1 pith:SHFEJQX2 submitted 2025-05-27 cs.DC

Di Yu , Changze Lv , Xin Du , Linshan Jiang , Wentao Tong , Zhenyu Liao , Xiaoqing Zheng , Shuiguang Deng This is my paper
classification cs.DC
keywords spikingneuralnetworksedge-cloudcollaborationknowledgedistillationentropy-basedoffloadingincrementallearningclass-incrementalenergy-efficientinferenceneuromorphiccomputing
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 proposes ECC-SNN, an edge-cloud inference framework that replaces the usual edge model with an energy-efficient spiking neural network (SNN) while keeping a powerful artificial neural network (ANN) in the cloud. Its central claim is that three mechanisms—distilling the cloud ANN's knowledge into the edge SNN, using normalized entropy to send only low-confidence inputs to the cloud, and retraining the SNN on those cloud-answered cases with a self-distillation objective—let a small spiking model match or exceed a cloud-only system at a fraction of the cost. Across CIFAR-10, CIFAR-100, Caltech, and Tiny-ImageNet under class-incremental settings, the framework reports an average accuracy improvement of 4.15%, an average energy reduction of 79.4%, and an average latency reduction of 39.1% over standalone edge SNN and cloud ANN baselines. If correct, this makes edge-cloud AI less dependent on continuous cloud involvement and more practical for resource-constrained IoT devices.

What carries the argument

The load-bearing machinery is a three-stage loop: Setup, Execution, and Update. In Setup, the loss $L(f_1(x),y)=L_{ce}+\lambda_1 L_{logit}+\lambda_2 L_{align}$ distills the cloud ANN's logits and, when architectures overlap, its features after a linear alignment layer into the edge SNN. In Execution, the normalized-entropy filter $s(1|x)=-\sum_{k=1}^{K}\sigma(f_{1,k}(x))\log\sigma(f_{1,k}(x))/\log K$ sends the input to the cloud exactly when the SNN's softmax entropy exceeds a threshold $\delta$, and the cloud's argmax is stored as that sample's label in the local buffer. In Update, the edge SNN trains on the buffer with the objective $L_{new}+\lambda_3 L_{old}$, where $L_{old}$ is the KL divergence between the old and new SNN output distributions, preventing catastrophic forgetting without an exemplar buffer. Equation (7), which idealizes the cloud ANN as a zero-loss oracle, is what reduces collaborative optimization to a cost-constrained minimization of edge-side loss alone.

What would settle it

Run the same framework under the same class-incremental protocol on a benchmark subset where the cloud teacher's accuracy is lower than the edge SNN's on the offloaded samples, such as a corrupted-noise split where the ViT or ResNet teacher systematically fails. If ECC-SNN's self-distillation update degrades accuracy relative to a no-update control, the oracle assumption in Equation (7) is violated and the claimed benefit depends on near-perfect teachers.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that an edge SNN can be made the primary inference engine of an edge-cloud system, with the cloud demoted to a teacher and a fallback for ambiguous inputs rather than the dominant compute source. The authors show that a pre-trained cloud ANN can transfer its knowledge to an arbitrary-architecture SNN through logit distillation plus optional feature alignment, that normalized entropy of the SNN's softmax output reliably flags the long-tail inputs the SNN cannot confidently handle, and that incrementally fine-tuning the SNN on those cloud-labeled hard samples with a learning-without-forgetting term steadily reduces how often the cloud must be consulted. The reported average gains of 4.15% accuracy, 79.4% lower energy, and 39.1% lower latency are the concrete form of this claim.

Load-bearing premise

The whole accuracy gain rests on trusting the cloud ANN's answer on every offloaded input as a correct training label; if the cloud is wrong on those hard examples, the incremental update can learn the wrong answer and the reported gains shrink.

Editorial extensions

If this is right

  • Edge SNNs can absorb knowledge from cloud ANNs without requiring matching ANN and SNN architectures, because the transfer runs through logits and aligned features rather than shared structure.
  • During inference, only low-confidence inputs pay the communication cost of cloud offloading, so the average per-input cost falls as the edge model improves.
  • On-device incremental updates with a self-distillation regularizer make continual learning feasible on resource-constrained devices while mitigating catastrophic forgetting.
  • If the reported numbers hold, ECC-SNN offers a concrete accuracy-energy-latency trade-off that is better than either running the full ANN in the cloud or running the SNN alone at the edge.

Reading between the lines

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

  • The entropy-gated loop is a form of self-training, so part of the reported gain may come from cloud-provided pseudo-labels; an extension would ablate this by retraining on ground-truth labels for the same buffered hard samples and measuring the difference.
  • The framework's upload rate decreases as the edge model learns, which implies a deployment-time cost curve: the longer a device operates, the cheaper each inference becomes, and quantifying that amortized cost is a natural next step.
  • If the mechanism generalizes, the same pattern could apply to other SNN-friendly modalities such as audio or time-series sensing, provided a strong ANN teacher exists for that domain.
  • The oracle idealization in Equation (7) suggests a testable robustness bound: the system's expected loss is approximately the cloud's error on offloaded inputs plus the SNN's error on confident inputs, and measuring those two error terms separately would predict when the framework helps or hurts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes ECC-SNN, an edge-cloud collaboration framework for image classification in which a cloud-based ANN teacher is used to distill knowledge into an edge-deployed SNN, inputs whose normalized entropy exceeds a threshold are offloaded to the cloud for re-inference, and the cloud's outputs on those ambiguous inputs are stored in a local buffer for on-device self-distillation incremental learning. The framework is evaluated on CIFAR-10, CIFAR-100, Caltech, and Tiny-ImageNet under a class-incremental learning protocol. The Abstract and Introduction claim that ECC-SNN improves average accuracy by 4.15%, reduces average energy consumption by 79.4%, and reduces average latency by 39.1% compared with standalone edge SNNs and cloud ANNs.

Significance. The central idea of coupling an ANN teacher in the cloud with an SNN at the edge, together with entropy-based offloading and on-device incremental learning, is relevant and timely for cost-efficient edge inference. The accuracy experiments are clearly presented: Table 1 and Table 2 make the joint-training gains traceable, Figure 4 shows the cloud upload rate falling as the edge model improves, and Figure 5 gives a concrete comparison with other incremental-learning methods. The code is said to be publicly available, which is a strength for reproducibility. However, the energy and latency reductions that carry the paper's cost-effectiveness thesis are not supported by any checkable model in the manuscript, and the cloud-oracle assumption in Eq. (7) interacts with noisy pseudo-labels in Algorithm 1. The paper is therefore not yet publishable in its current form.

major comments (4)
  1. [§5.2, Figs. 6-7] The manuscript never defines how Cu, Cf, Tu, and Tf are computed. No equations are given for energy per spike, cloud GPU energy per inference, communication energy per byte, bandwidth, or round-trip time, and no hardware measurements are reported; Figure 6's own caption calls the energy values 'theoretical.' The Abstract's 79.4% energy reduction and 39.1% latency reduction are therefore not derivable from anything a reader can check, and favorable unstated parameter choices could make the reductions arbitrarily large. Because cost-effectiveness is the paper's central claim, this is load-bearing. Please provide an explicit cost model with equations and parameter values, add a sensitivity analysis over those parameters, or replace the theoretical numbers with measured values.
  2. [§4.2, Eq. (7); Algorithm 1, lines 7-11] Equation (7) idealizes the cloud model f0 as an oracle whose loss L(f0(x),y) is zero, while Algorithm 1 stores the cloud's argmax as the pseudo-label in the local buffer. Table 1 shows that the actual teachers are far from perfect: for example, VGG-16 on CIFAR-100 reaches only 89.10% accuracy. Wrong pseudo-labels from the cloud can therefore poison the incremental updates in Eq. (12) and inflate the reported accuracy gains. The authors should quantify this effect, for example by comparing against a variant that uses ground-truth labels in the buffer, or by measuring pseudo-label accuracy and applying a label-noise-robust loss.
  3. [§5.2, Tables 1-2] The headline '4.15% average accuracy improvement' is never tied to a table, equation, or explicit aggregation rule. Similarly, the sentence 'the joint training design in ECC-SNN can achieve an average accuracy improvement of 2.87%' does not state whether the average is over all dataset-architecture pairs or over tasks. The authors should state exactly how each aggregate number is computed and which columns of which table support it.
  4. [§5.2, Figs. 4-7] The evaluation compares ECC-SNN with standalone edge SNNs and cloud ANNs, but not with an equally equipped edge-cloud system using an ANN at the edge or a simpler non-learning offloading rule. Since the paper's contribution is the combination of SNN, entropy-based offloading, and incremental learning, an ablation that isolates the benefit of the SNN edge model over an ANN edge model under the same offloading and incremental-learning machinery would help establish that the accuracy and cost benefits come from the proposed design rather than from the offloading strategy alone.
minor comments (4)
  1. [§4.3] The sentence 'our proposed framework will provide a bio-plausible explanation for this method)' is incomplete and contains a stray closing parenthesis; it should be rewritten or removed.
  2. [§5.3] The heading 'Sensitive Analysis' should be 'Sensitivity Analysis'.
  3. [References] The Zhou et al. 2022 entry appears with a corrupted title: 'A model or 603 exemplars' should be checked and corrected.
  4. [§5.3, Fig. 8] The sensitivity study varies CUR while keeping the entropy threshold δ fixed at 0.3; a direct sensitivity analysis over δ would make the behavior of the filter clearer.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ECC-SNN's claims rest on measured experiments and standard training objectives, not on self-referential definitions.

full rationale

The derivation chain in ECC-SNN is largely self-contained and does not reduce its claims to its inputs. The optimization objective in Eqs. 5-7 is a standard expected-loss/cost formulation; the simplifying assumption that the cloud ANN f0 is an oracle is stated explicitly and is not used to fabricate the empirical accuracy numbers, which are measured against true labels in Tables 1-2 and Figures 4-5. The joint training loss (Eqs. 8-10) combines cross-entropy, logit distillation, and feature alignment with explicit weights; these are training objectives, not fitted predictions. Incremental learning loss Eq. 12 is a standard LwF-style combination of a new-task loss and a KL regularization term, with the local buffer filled by cloud predictions (Algorithm 1, line 11); this is a self-training strategy, not a circularly defined metric or a parameter fitted to the reported outcome. The energy and latency claims in Section 5.2 and Figures 6-7 are not derived from any disclosed equation or measured hardware parameters, so they are a serious verifiability gap, but a missing derivation is not circularity. Self-citations such as EC-SNN [Yu et al., 2024] and [Lv et al., 2024] are background references and do not carry the central argument. No load-bearing circular step was found.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central result rests on standard SNN machinery, on the empirical reliability of entropy-based offloading, and on the trustworthiness of cloud labels for self-training. The oracle idealization in Eq. 7 is the most significant unstated assumption because it is contradicted by the imperfect teachers used in experiments.

free parameters (4)
  • entropy threshold delta = 0.3
    Fixed threshold for cloud offloading in the execution stage; it controls the cloud upload rate and therefore the reported energy and latency reductions (Figures 4, 6, 7).
  • distillation weights lambda1, lambda2, lambda3 = not reported
    Weights for logit distillation, feature alignment, and old-model KL in Eqs. 8, 10, and 12; no values or sensitivity analysis are provided.
  • SNN simulation settings and LIF constants = not reported
    The LIF neuron model in Eqs. 1-3 requires the decay factor, threshold, reset potential, and number of time steps; these are not specified, though they affect SNN accuracy and cost.
  • energy and latency cost model constants = not reported
    Figures 6-7 report 'theoretical' energy and latency values, but the cost model, hardware parameters, and communication costs are left unspecified, making the headline 79.4% and 39.1% reductions non-recomputable.
assumptions (5)
  • domain assumption The LIF neuron model and surrogate-gradient BPTT are valid and are used as in prior work (Eqs. 1-3, Section 3.1).
    The SNN training and inference rely on standard LIF dynamics and surrogate gradients, which the paper takes from the cited literature.
  • domain assumption The cloud ANN f0 can be treated as a perfect oracle with zero loss (Eq. 7, Section 4.2).
    This simplification reduces the objective to only the edge model loss, but the experiments use imperfect ViT, ResNet, and VGG teachers, so the assumption is violated.
  • domain assumption Normalized entropy of the edge SNN softmax is a reliable measure of input difficulty (Eq. 11).
    The entropy filter is borrowed from Huang et al. 2021; the paper assumes it cleanly separates learned from unlearned classes, as shown in the Figure 3 case study.
  • domain assumption Cloud-predicted labels stored in the local buffer are accurate enough to supervise incremental learning (Algorithm 1, lines 7-11).
    The framework has no mechanism to detect or correct cloud errors, which is load-bearing once the oracle assumption is relaxed.
  • domain assumption Class-incremental learning is a representative testbed for prior probability distribution drift (Section 5.1).
    The paper evaluates only class-incremental splits and does not test other drift types despite discussing general prior drift.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ECC-SNN: Cost-Effective Edge-Cloud Collaboration for Spiking Neural Networks." pith.science (2026). https://pith.science/paper/SHFEJQX2

@misc{pith2026250520835,
  author       = {Pith},
  title        = {Pith review of: ECC-SNN: Cost-Effective Edge-Cloud Collaboration for Spiking Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SHFEJQX2}},
  note         = {Machine review of arXiv:2505.20835}
}
read the original abstract

Most edge-cloud collaboration frameworks rely on the substantial computational and storage capabilities of cloud-based artificial neural networks (ANNs). However, this reliance results in significant communication overhead between edge devices and the cloud and high computational energy consumption, especially when applied to resource-constrained edge devices. To address these challenges, we propose ECC-SNN, a novel edge-cloud collaboration framework incorporating energy-efficient spiking neural networks (SNNs) to offload more computational workload from the cloud to the edge, thereby improving cost-effectiveness and reducing reliance on the cloud. ECC-SNN employs a joint training approach that integrates ANN and SNN models, enabling edge devices to leverage knowledge from cloud models for enhanced performance while reducing energy consumption and processing latency. Furthermore, ECC-SNN features an on-device incremental learning algorithm that enables edge models to continuously adapt to dynamic environments, reducing the communication overhead and resource consumption associated with frequent cloud update requests. Extensive experimental results on four datasets demonstrate that ECC-SNN improves accuracy by 4.15%, reduces average energy consumption by 79.4%, and lowers average processing latency by 39.1%.

Figures

Figures reproduced from arXiv: 2505.20835 by the authors.

Figure 1
Figure 1. Overview of the proposed ECC-SNN. In all three stages, the cloud-based ANN model directly or indirectly supports the edge SNN [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Optional feature distillation in the joint training approach. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Case Study: a spiking VGG-9 model learned with CIFAR￾10 limited to samples labeled 1 and 6. The entropy distributions and corresponding cloud upload rates for each label are derived from each test sample’s model output predictive distribution. Algorithm 1 Collaborative Inference in ECC-SNN Input: collected input feature x; filter threshold δ. Parameter: cloud ANN model f0; edge SNN model f1. Output: the inference re… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Changing patterns of accuracy performance and CUR as [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Average accuracy of edge SNN w.r.t. different IL methods. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Average theoretical energy cost (mJ) per input for infer [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 8
Figure 8. Figure 8: Accuracy improvement (AccI) w.r.t CUR (%) with differ￾ent learned classes at current task. prising observation is that ECC-SNN can improve accuracy in simple classification tasks. For example, the overall accu￾racy of ECC-SNN can exceed that of the standalone cloud￾bas…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Edge Intelligence with Spiking Neural Networks

    cs.DC 2025-07 conditional novelty 4.0 of 10

    A comprehensive review of spiking neural networks for edge computing, covering neuron models, learning algorithms, hardware, deployment, security, and evaluation, with a claim to be the first survey on this specific i...

Reference graph

Works this paper leans on

38 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [1]

    Adaptive neural networks for efficient inference

    [Bolukbasi et al., 2017] Tolga Bolukbasi, Joseph Wang, Ofer Dekel, and Venkatesh Saligrama. Adaptive neural networks for efficient inference. In International Confer- ence on Machine Learning, pages 527–536. PMLR,

    work page 2017

  2. [5]

    Oebench: Investigating open environment challenges in real-world relational data streams

    [Diao et al., 2024] Yiqun Diao, Yutong Yang, Qinbin Li, Bingsheng He, and Mian Lu. Oebench: Investigating open environment challenges in real-world relational data streams. Proc. VLDB Endow., 17(6):1283–1296,

    work page 2024

  3. [7]

    An integrated cloud-edge-device adaptive deep learning service for cross-platform web.IEEE Transactions on Mo- bile Computing, 22(4):1950–1967,

    [Huang et al., 2021] Yakun Huang, Xiuquan Qiao, Jian Tang, Pei Ren, Ling Liu, Calton Pu, and Junliang Chen. An integrated cloud-edge-device adaptive deep learning service for cross-platform web.IEEE Transactions on Mo- bile Computing, 22(4):1950–1967,

    work page 2021

  4. [9]

    Training on the edge: The why and the how

    [Kukreja et al., 2019] Navjot Kukreja, Alena Shilova, Olivier Beaumont, Jan Huckelheim, Nicola Ferrier, Paul Hovland, and Gerard Gorman. Training on the edge: The why and the how. In2019 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), pages 899–903. IEEE,

    work page 2019

  5. [11]

    Carm: Hierarchical episodic memory for continual learning

    [Lee et al., 2022] Soobee Lee, Minindu Weerakoon, Jonghyun Choi, Minjia Zhang, Di Wang, and Myeongjae Jeon. Carm: Hierarchical episodic memory for continual learning. In Proceedings of the 59th ACM/IEEE Design Automation Conference, pages 1147–1152,

    work page 2022

  6. [13]

    Appealnet: An efficient and highly-accurate edge/cloud collaborative architecture for dnn inference

    [Li et al., 2021] Min Li, Yu Li, Ye Tian, Li Jiang, and Qiang Xu. Appealnet: An efficient and highly-accurate edge/cloud collaborative architecture for dnn inference. In 2021 58th ACM/IEEE Design Automation Conference (DAC), pages 409–414. IEEE,

    work page 2021

  7. [14]

    Distributed dnn infer- ence with fine-grained model partitioning in mobile edge computing networks

    [Li et al., 2024] Hui Li, Xiuhua Li, Qilin Fan, Qiang He, Xi- aofei Wang, and Victor CM Leung. Distributed dnn infer- ence with fine-grained model partitioning in mobile edge computing networks. IEEE Transactions on Mobile Com- puting,

    work page 2024

  8. [15]

    Complexity-aware adaptive training and inference for edge-cloud distributed ai systems

    [Long et al., 2021] Yinghan Long, Indranil Chakraborty, Gopalakrishnan Srinivasan, and Kaushik Roy. Complexity-aware adaptive training and inference for edge-cloud distributed ai systems. In 2021 IEEE 41st International Conference on Distributed Computing Systems (ICDCS), pages 573–583. IEEE,

    work page 2021

Show all 38 references
  1. [16]

    Ef- ficient and effective time-series forecasting with spiking neural networks

    [Lv et al., 2024] Changze Lv, Yansen Wang, Dongqi Han, Xiaoqing Zheng, Xuanjing Huang, and Dongsheng Li. Ef- ficient and effective time-series forecasting with spiking neural networks. In Forty-first International Conference on Machine Learning,

    2024

  2. [17]

    Cost- effective on-device continual learning over memory hierar- chy with miro

    [Ma et al., 2023] Xinyue Ma, Suyeon Jeong, Minjia Zhang, Di Wang, Jonghyun Choi, and Myeongjae Jeon. Cost- effective on-device continual learning over memory hierar- chy with miro. In Proceedings of the 29th Annual Interna- tional Conference on Mobile Computing and Networking,...

    2023

  3. [18]

    Darwin3: a large-scale neu- romorphic chip with a novel isa and on-chip learning

    [Ma et al., 2024] De Ma, Xiaofei Jin, Shichun Sun, Yitao Li, Xundong Wu, Youneng Hu, Fangchao Yang, Huajin Tang, Xiaolei Zhu, Peng Lin, et al. Darwin3: a large-scale neu- romorphic chip with a novel isa and on-chip learning. Na- tional Science Review, 11(5):nwae102,

    2024

  4. [19]

    Networks of spiking neu- rons: the third generation of neural network models

    [Maass, 1997] Wolfgang Maass. Networks of spiking neu- rons: the third generation of neural network models. Neu- ral networks, 10(9):1659–1671,

    1997

  5. [21]

    Structured sparse back- propagation for lightweight on-device continual learning on microcontroller units

    [Paissan et al., 2024] Francesco Paissan, Davide Nadalini, Manuele Rusci, Alberto Ancilotto, Francesco Conti, Luca Benini, and Elisabetta Farella. Structured sparse back- propagation for lightweight on-device continual learning on microcontroller units. In Proceedings of the I...

    2024

  6. [22]

    Supervised algorith- mic fairness in distribution shifts: A survey

    [Shao et al., 2024] Minglai Shao, Dong Li, Chen Zhao, Xin- tao Wu, Yujie Lin, and Qin Tian. Supervised algorith- mic fairness in distribution shifts: A survey. In Kate Larson, editor, Proceedings of the Thirty-Third Interna- tional Joint Conference on Artificial Intelligence, ...

    2024

  7. [23]

    Spikingresformer: Bridging resnet and vision trans- former in spiking neural networks

    [Shi et al., 2024] Xinyu Shi, Zecheng Hao, and Zhaofei Yu. Spikingresformer: Bridging resnet and vision trans- former in spiking neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5610–5619,

    2024

  8. [24]

    Chal- lenges in benchmarking stream learning algorithms with real-world data

    [Souza et al., 2020] Vinicius MA Souza, Denis M dos Reis, Andre G Maletzke, and Gustavo EAPA Batista. Chal- lenges in benchmarking stream learning algorithms with real-world data. Data Mining and Knowledge Discovery , 34(6):1805–1858,

    2020

  9. [25]

    Towards robust graph incremental learning on evolving graphs

    [Su et al., 2023] Junwei Su, Difan Zou, Zijun Zhang, and Chuan Wu. Towards robust graph incremental learning on evolving graphs. In International Conference on Machine Learning, pages 32728–32748. PMLR,

    2023

  10. [26]

    Transtab: Learning transferable tabular transformers across tables

    [Wang and Sun, 2022] Zifeng Wang and Jimeng Sun. Transtab: Learning transferable tabular transformers across tables. Advances in Neural Information Processing Systems, 35:2902–2915,

    2022

  11. [27]

    A new ann-snn conversion method with high accuracy, low latency and good robust- ness

    [Wang et al., 2023] Bingsen Wang, Jian Cao, Jue Chen, Shuo Feng, and Yuan Wang. A new ann-snn conversion method with high accuracy, low latency and good robust- ness. In IJCAI, pages 3067–3075,

    2023

  12. [28]

    End-edge-cloud col- laborative computing for deep learning: A comprehensive survey

    [Wang et al., 2024] Yingchao Wang, Chen Yang, Shulin Lan, Liehuang Zhu, and Yan Zhang. End-edge-cloud col- laborative computing for deep learning: A comprehensive survey. IEEE Communications Surveys & Tutorials,

    2024

  13. [29]

    Direct training for spiking neural networks: Faster, larger, better

    [Wu et al., 2019] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 1311–1318,

    2019

  14. [30]

    Hebbian learning based orthogonal projection for continual learning of spik- ing neural networks

    [Xiao et al., 2024] Mingqing Xiao, Qingyan Meng, Zong- peng Zhang, Di He, and Zhouchen Lin. Hebbian learning based orthogonal projection for continual learning of spik- ing neural networks. In The Twelfth International Confer- ence on Learning Representations,

    2024

  15. [31]

    Constructing deep spik- ing neural networks from artificial neural networks with knowledge distillation

    [Xu et al., 2023] Qi Xu, Yaxin Li, Jiangrong Shen, Jian K Liu, Huajin Tang, and Gang Pan. Constructing deep spik- ing neural networks from artificial neural networks with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogniti...

    2023

  16. [32]

    Gkt: A novel guidance-based knowledge transfer framework for efficient cloud-edge collaboration llm deployment

    [Yao et al., 2024b] Yao Yao, Zuchao Li, and Hai Zhao. Gkt: A novel guidance-based knowledge transfer framework for efficient cloud-edge collaboration llm deployment. arXiv preprint arXiv:2405.19635,

    arXiv

  17. [33]

    Ec-snn: Splitting deep spik- ing neural networks for edge devices

    [Yu et al., 2024] Di Yu, Xin Du, Linshan Jiang, Wentao Tong, and Shuiguang Deng. Ec-snn: Splitting deep spik- ing neural networks for edge devices. In Proceedings of the Thirty-Third International Joint Conference on Artifi- cial Intelligence, IJCAI-24, pages 5389–5397,

    2024

  18. [34]

    Trainable spiking-yolo for low-latency and high-performance object detection

    [Yuan et al., 2024] Mengwen Yuan, Chengjun Zhang, Zim- ing Wang, Huixiang Liu, Gang Pan, and Huajin Tang. Trainable spiking-yolo for low-latency and high-performance object detection. Neural Networks , 172:106092,

    2024

  19. [35]

    Dvfo: Learning-based dvfs for energy-efficient edge-cloud collaborative inference

    [Zhang et al., 2024] Ziyang Zhang, Yang Zhao, Huan Li, Changyao Lin, and Jie Liu. Dvfo: Learning-based dvfs for energy-efficient edge-cloud collaborative inference. IEEE Transactions on Mobile Computing,

    2024

  20. [36]

    A model or 603 exem- plars: Towards memory-efficient class-incremental learn- ing

    [Zhou et al., 2022] Da-Wei Zhou, Qi-Wei Wang, Han-Jia Ye, and De-Chuan Zhan. A model or 603 exem- plars: Towards memory-efficient class-incremental learn- ing. CoRR, abs/2205.13218,

    2022 arXiv

  21. [37]

    Ecseg: Edge-cloud switched image segmentation for au- tonomous vehicles

    [Zhou et al., 2024b] Siyuan Zhou, Duc Van Le, and Rui Tan. Ecseg: Edge-cloud switched image segmentation for au- tonomous vehicles. In 2024 21th Annual IEEE Interna- tional Conference on Sensing, Communication, and Net- working (SECON). IEEE,

    2024

  22. [38]

    Gilpin, and Jason Eshraghian

    [Zhu et al., 2024] Rui-Jie Zhu, Ziqing Wang, Leilani H. Gilpin, and Jason Eshraghian. Autonomous driving with spiking neural networks. InThe Thirty-eighth Annual Con- ference on Neural Information Processing Systems, 2024

    2024

  23. [1997]

    Class-incremental learning: sur- vey and performance evaluation on image classification

    [Masana et al., 2022] Marc Masana, Xialei Liu, Bartłomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost Van De Weijer. Class-incremental learning: sur- vey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine In- tellige...

    2022

  24. [2017]

    Backpropagation-based learning techniques for deep spik- ing neural networks: A survey.IEEE Transactions on Neu- ral Networks and Learning Systems,

    [Dampfhoffer et al., 2023] Manon Dampfhoffer, Thomas Mesquida, Alexandre Valentian, and Lorena Anghel. Backpropagation-based learning techniques for deep spik- ing neural networks: A survey.IEEE Transactions on Neu- ral Networks and Learning Systems,

    2023

  25. [2019]

    Exploring system performance of continual learning for mobile and embedded sensing applications

    [Kwon et al., 2021] Young D Kwon, Jagmohan Chauhan, Abhishek Kumar, Pan Hui Hkust, and Cecilia Mascolo. Exploring system performance of continual learning for mobile and embedded sensing applications. In 2021 IEEE/ACM Symposium on Edge Computing (SEC), pages 319–332. IEEE,

    2021

  26. [2020]

    Surrogate module learning: Reduce the gradi- ent error accumulation in training spiking neural networks

    [Deng et al., 2023] Shikuang Deng, Hao Lin, Yuhang Li, and Shi Gu. Surrogate module learning: Reduce the gradi- ent error accumulation in training spiking neural networks. In International Conference on Machine Learning , pages 7645–7657. PMLR,

    2023

  27. [2021]

    A unified optimiza- tion framework of ann-snn conversion: towards optimal mapping from activation values to firing rates

    [Jiang et al., 2023] Haiyan Jiang, Srinivas Anumasa, Giulia De Masi, Huan Xiong, and Bin Gu. A unified optimiza- tion framework of ann-snn conversion: towards optimal mapping from activation values to firing rates. In Inter- national Conference on Machine Learning, pages 14945...

    2023

  28. [2022]

    Learning without forgetting

    [Li and Hoiem, 2017] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence , 40(12):2935–2947,

    2017

  29. [2023]

    Rethinking the performance comparison between snns and anns

    [Deng et al., 2020] Lei Deng, Yujie Wu, Xing Hu, Ling Liang, Yufei Ding, Guoqi Li, Guangshe Zhao, Peng Li, and Yuan Xie. Rethinking the performance comparison between snns and anns. Neural networks, 121:294–307,

    2020

  30. [2024]

    A lightweight collaborative deep neural network for the mobile web in edge cloud

    [Huang et al., 2020] Yakun Huang, Xiuquan Qiao, Pei Ren, Ling Liu, Calton Pu, Schahram Dustdar, and Junliang Chen. A lightweight collaborative deep neural network for the mobile web in edge cloud. IEEE Transactions on Mobile Computing, 21(7):2289–2305,

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.