Pith. sign in

REVIEW 4 major objections 6 minor 70 references

Spiking Neural Network Feature Discrimination Boosts Modality Fusion

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read L2-normalizing the final spiking features of each modality pushes audio-visual fusion to 98.6% on CIFAR10-AV and 97.2% on UrbanSound8K-AV, beating a transformer-based spiking baseline.

desk verdict A plausible but unproven claim: the paper shows L2-normalized SNN features can be fused competitively, yet never isolates the L2 layer, so the central 'boosts' claim rests on a missing ablation. read the letter →

arxiv 2502.10423 v1 pith:WDIOVUDV submitted 2025-02-05 cs.NE cs.CVcs.LGeess.IV

Katerina Maria Oikonomou , Ioannis Kansizoglou , Antonios Gasteratos This is my paper · ORCID
classification cs.NEcs.CVcs.LGeess.IV
keywords spikingneuralnetworksfeaturediscriminationL2normalizationaudio-visualfusionmulti-modalclassificationResNetMLPhypersphericalembeddings
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

This paper tries to establish that feature discrimination, making same-class features more alike and different-class features more distinct, can be transplanted into spiking neural networks by L2-normalizing the time-aggregated feature vectors of each modality before fusion. The authors report that this normalization, applied after the last hidden layer of a spiking visual network and a spiking audio network, makes the learned features more compact within a class and more separated between classes, and that it pushes audio-visual classification to 98.6% on CIFAR10-AV and 97.2% on UrbanSound8K-AV. Those numbers beat the transformer-based multi-modal spiking baseline they compare against. The wider point, if true, is that spiking networks do not need to give up representation quality for energy efficiency: a simple geometric constraint on the final embedding can sharpen the features they pass to a fusion module.

What carries the argument

The load-bearing object is the custom L2 normalization layer inserted after the last hidden layer of each unimodal spiking network. For each input it computes the time-averaged feature vector $\mathbf{z}_{av} = \frac{1}{T}\sum_{t=1}^{T}\mathbf{z}_t$, normalizes it as $\hat{\mathbf{z}} = \mathbf{z}_{av}/\|\mathbf{z}_{av}\|_2$, and likewise normalizes each class weight vector $\hat{\mathbf{w}}_j = \mathbf{w}_j/\|\mathbf{w}_j\|_2$, so the logit for class $j$ becomes the cosine similarity $\hat{\mathbf{z}}^T \hat{\mathbf{w}}_j$. This projection onto the unit hypersphere is what forces angular concentration of same-class embeddings, and it is paired with an MSE spike-count loss and a spiking MLP fusion network that concatenates the normalized visual and audio embeddings.

What would settle it

Train the same audio-visual spiking pipeline on CIFAR10-AV twice, with identical architecture, hyperparameters, and random seeds, changing only the presence of the L2 normalization layer after each modality's last hidden layer; if the normalized model does not show both a clear increase in cosine-distance diagonal structure and a higher fusion accuracy than the vanilla model, the paper's central claim is not supported.

Watch

Extended reading notes

Core claim

The central claim is that L2 normalization of the final hidden-layer features in spiking networks constitutes the first feature discrimination mechanism for SNNs, and that it improves both unimodal separability and multimodal fusion. Concretely, the paper time-averages each modality's feature vectors across simulation steps, normalizes both the averaged embeddings and the class weight vectors to unit norm, and uses their dot product as the logit; this makes classification depend on cosine similarity on a hypersphere. The paper then feeds the normalized visual embedding (512 dimensions) and normalized audio embedding (27136 dimensions) through a spiking multilayer perceptron for fusion. On CIFAR10-AV the fusion reaches 98.6% accuracy, with unimodal audio at 99.6% and visual at 92.74%; on UrbanSound8K-AV fusion reaches 97.2%, with audio at 87.49% and visual at 95.42%. Cosine-distance matrices of the normalized features show stronger diagonal structure than features from a vanilla fully connected layer, which the authors take as evidence of higher intra-class compactness and inter-class separability.

Load-bearing premise

The argument depends on the 'vanilla' model being exactly the same network, loss, and training procedure as the L2-normalized model except for the added normalization layer; if the two pipelines differ in any other way, the reported gain in feature separability cannot be pinned on feature discrimination.

Editorial extensions

If this is right

  • Spiking networks can inherit a well-studied DNN tool, hyperspherical feature normalization, without changing their event-driven computation or their spike-based loss.
  • Future multi-modal SNN designs can apply the same L2 layer on any modality's final embedding before fusion; the paper demonstrates it on vision and audio with very different feature dimensionalities.
  • Cosine-distance matrices give a direct training-time diagnostic for whether a spiking feature space is separating classes, not just a final accuracy number.
  • On two public audio-visual benchmarks, the proposed fusion outperforms the transformer-based multi-modal SNN baseline, suggesting the combination of normalization and spiking MLP fusion is competitive at low reported timesteps.

Reading between the lines

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

  • The normalization's scale-invariance likely also helps fusion by putting the 512-dimensional visual features and the 27136-dimensional audio features on the same geometric footing before concatenation; the paper does not isolate this effect from the separability gain.
  • Because the motivating lemmas assume LIF neurons and an MSE spike-count loss, the benefit of L2 normalization may not transfer unchanged to other neuron models or to cross-entropy-trained spiking networks; this is an untested extension.
  • A controlled comparison that varies only the normalization layer, holding architecture, loss, optimizer, and random seeds fixed, would turn the reported accuracy gap into a causal claim about feature discrimination.
  • The same normalization could be tested on unimodal spiking tasks alone to see whether the fusion gain comes from better unimodal embeddings or from the fusion network itself.
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 / 6 minor

Summary. The paper proposes an L2 normalization layer after the last hidden layer of uni-modal spiking visual and audio networks, followed by a spiking MLP (SMLP) fusion network, and evaluates this pipeline on two audio-visual datasets (CIFAR10-AV and UrbanSound8K-AV). The authors report accuracies of 98.6% on CIFAR10-AV and 97.2% on UrbanSound8K-AV, compare against a transformer-based multi-modal SNN baseline, and include an ablation study of several architectural components. The central claim is that L2 normalization-induced feature discrimination 'boosts modality fusion.'

Significance. If the results were rigorously supported, the contribution would be a simple, potentially useful extension of well-known L2-normalized cosine-similarity classifiers to spiking neural networks, and the first empirical study of feature discrimination in SNNs. The paper contains a clear architecture description, standard benchmark datasets, and a focused ablation of non-L2 architectural choices. However, the current evidence is insufficient to establish the central claim: no controlled with/without-L2 comparison is reported, the evaluation lacks error bars or multiple runs, and on CIFAR10-AV the fusion underperforms its own audio-only branch. With additional controlled experiments and statistical reporting, the contribution could be validated, but as written the significance is not demonstrated.

major comments (4)
  1. [Section V-C, V-D, Table III] The central claim that L2 normalization 'boosts modality fusion' is not tested, because no controlled ablation removes or adds the L2 layer. Tables I and II report accuracies only for the full L2 pipeline, and Table III varies LIF placement, dropout, blocks, and pooling but never the L2 layer. The 'vanilla' baseline used in the cosine-distance heatmaps of Figs. 6 and 9 is not specified in terms of architecture, loss function, training schedule, or hyperparameters, so the observed differences in separability cannot be attributed to the L2 normalization layer specifically. A same-pipeline with/without-L2 comparison is required to support the title claim and contribution (1).
  2. [Table I, Section V-F] On CIFAR10-AV, the reported fusion accuracy of 98.60% is lower than the audio-only branch's 99.60% (Table I), yet Section V-F states that the approach 'outperforms the uni-modal SNN models.' This internal contradiction undermines the claim that L2-normalized feature discrimination boosts fusion on this dataset. Even if the comparison to the transformer baseline [31] shows an improvement, the improvement cannot be attributed to L2 normalization without a controlled ablation of that layer in the fusion pipeline.
  3. [Section III, Lemmas 1-2] The theoretical lemmas do not provide a rigorous foundation for the proposed normalization. Lemma 1 assumes the optimal-case condition that the SNN classifies correctly at t=0 and then refers to the 'similar approach with the Feature Space Division proof in [63]' without reproducing that argument. Lemma 2 reformulates class dominance as a weighted sum of dot products, which is a restatement of the LIF dynamics, and does not imply that an explicit L2 normalization layer improves separability or accuracy. Moreover, because Eqs. (11)-(13) define classification by cosine similarity, the block-diagonal appearance of the L2 heatmaps in Figs. 6 and 9 is partly a consequence of this construction. We recommend that the theoretical section either prove a direct benefit of the normalization layer or be repositioned as motivational intuition.
  4. [Section V-F, Table IV] The comparative evaluation is not controlled. Table IV compares the proposed multi-modal method against uni-modal SNNs trained on different datasets (CIFAR-10 and UrbanSound8K) and against a single multi-modal transformer baseline [31] with different timesteps, architectures, and likely training procedures. No error bars, multiple seeds, or statistical significance tests are reported anywhere in the paper, so the claimed improvements over [31] (98.6 vs. 97.01 and 97.2 vs. 96.85) may be within run-to-run variation. The authors should report mean and standard deviation over at least three runs and, if possible, reproduce the baseline under matched conditions.
minor comments (6)
  1. [Section V-B1] The initial learning rate for the visual modality is given as '20−3', which appears to be a typo for '10−3'; please correct and ensure consistent formatting of numerical values.
  2. [Section V-C] The heatmap description states 'with yellow denoting high angular distance and darker ones denoting darker ones'; this sentence is incomplete. The intended meaning is likely 'yellow denotes high angular distance and darker colors denote lower angular distance.'
  3. [Table IV, Section V-F] In Table IV, the UrbanSound8K-AV row labels the proposed method as 'SMLP Multi-modal transformer,' but the method is a spiking MLP, not a transformer; this is inconsistent with the method name used elsewhere. Also, Section V-F contains the typo 'UltraSound8k-AV' instead of 'UrbanSound8k-AV.'
  4. [Section IV-D] The audio feature space dimension is given as A ∈ R^27136, and the concatenated dimension as Vconcat ∈ R^27648 (Section IV-E), but the audio network architecture in Fig. 4 does not clearly explain how this large 27136-dimensional representation is produced. Please specify the output shapes of each layer.
  5. [Section IV-E] It is not stated whether the uni-modal visual and audio networks are frozen during SMLP fusion training or fine-tuned jointly. This detail is essential for reproducing the fusion results and for interpreting the contribution of the L2 normalization layers.
  6. [Chapter III and Conclusion] The claim of being the 'first work investigating feature discrimination in SNNs' is strong and may be difficult to verify; consider softening it to 'to the best of our knowledge' and citing any prior L2-normalization techniques applied to spiking networks, if they exist.

Circularity Check

2 steps flagged · score 4.0 of 10

Feature-discrimination evidence reduces to the method's own cosine objective, and the theoretical lemmas import angular separability from a self-cited prior theorem; the fusion accuracies are empirical, but the causal 'boosts' claim is not isolated by any no-L2 ablation.

  1. self definitional [Section IV-B, Eqs. (11)-(13), and Section V-C, Eq. (23), Figs. 6/9]
    "Accordingly, the logits ... are computed as follows: Lij = ˆzT i ˆwj. (13) ... To achieve that, we calculated the cosine distances between two feature vectors, by applying the cosine distance formula: Cdist = 1 − fi · fj/∥fi∥∥fj∥ (23)."

    The proposed L2 layer defines the classifier as cosine similarity between L2-normalized embeddings and L2-normalized class weights, so Eq. 13 is the training objective. The evidence for feature discrimination is then the cosine distance matrix of those same normalized features, Eq. 23. Thus the claimed property, improved angular separability, is not an independent finding but exactly the quantity the construction is designed and trained to optimize. The comparison against an underspecified 'vanilla FC' model compares a cosine-optimized model with a dot-product model under a cosine metric, so the reported improvement is imposed by the definition of the method and metric rather than demonstrated by a controlled ablation.

  2. self citation load bearing [Section III, Lemma 1, Eqs. (4)-(6)]
    "In [63], we have proved that the logits of the i-th output neuron (zi) can be described as: zi = ¯a · ¯wi ... Following the similar approach with the Feature Space Division proof in [63], it is straightforward that Eq. 6 ensures angular separability at time-step 0."

    The paper's central theoretical premise, that SNN output features are angularly discriminated, is not derived from the LIF dynamics here. The proof substitutes I_in = z_i and then appeals to the authors' own prior theorem [63] for the claim that the dot-product inequality Eq. 6 ensures angular separability. The load-bearing geometric step is therefore a self-citation, with the actual derivation left as 'straightforward.' The lemma reduces to a restatement of the dot-product logits plus an imported prior result, rather than an independent proof from the spiking model.

full rationale

The reported accuracy numbers on CIFAR10-AV and UrbanSound8K-AV (Tables I, II, IV) are empirical and benchmarked against the transformer-based SNN [31], so the raw fusion performance is not circular. However, the paper's title claim that feature discrimination 'boosts' modality fusion is not isolated by any ablation that removes or adds the L2 layer while holding everything else fixed; the 'vanilla' comparison in Figs. 6 and 9 lacks architecture, loss, and training details, and Table III never varies the L2 layer. That is primarily a missing-control correctness issue, not circularity. The genuine circularity is narrower: the method's classifier is defined as cosine similarity on L2-normalized features and weights (Eq. 13), and the feature-discrimination evidence is a cosine-distance matrix of the same features (Eq. 23), so the measured 'improvement' is aligned with the construction by design. Additionally, the theoretical Lemmas 1-2 lean on the authors' own [63] theorem for the angular-separability claim, with Lemma 2 explicitly described as 'motivation' rather than proof. These elements make the feature-discrimination narrative partially circular, while leaving the benchmark accuracies as independent empirical content; hence a score of 4 rather than higher.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central empirical claim depends mainly on the chosen loss targets (rcorrect/rincorrect), timesteps, and surrogate gradient settings, all chosen by hand. The theoretical motivation leans on a self-cited prior theorem and a LIF model without reset. The empirical evaluation itself adds no invented entities.

free parameters (6)
  • rcorrect (visual) = 0.90
    Target firing ratio for the correct class in the MSE spike-count loss; chosen by hand in Section V-B1, no sensitivity analysis.
  • rincorrect (visual) = 0.10
    Target firing ratio for incorrect classes; chosen by hand in Section V-B1.
  • rcorrect (audio) = 0.85
    Target firing ratio for the correct class in the audio network; chosen by hand in Section V-B1.
  • rincorrect (audio) = 0.15
    Target firing ratio for incorrect classes in the audio network; chosen by hand in Section V-B1.
  • number of timesteps T = visual=8, audio=4, fusion=8
    Temporal unrolling length per modality chosen without ablation; affects feature aggregation and loss targets.
  • surrogate gradient parameters = a (arctan sharpness) and k=5 (fast sigmoid slope)
    Gradient approximation shapes chosen per modality (Section V-B1) and affect training dynamics.
assumptions (5)
  • domain assumption LIF neuron dynamics without explicit reset
    Eq. 1-2 in Section III describe membrane potential update and threshold, but omit the reset mechanism; the theoretical lemmas in Section III ignore reset, which is standard in LIF implementations.
  • domain assumption Transfer of the angular-discrimination theorem from [63] to SNNs
    Section III, Lemma 1 uses the prior work's result that logits are dot products of features and weights and asserts this ensures angular separability in SNNs with MSE spike-count loss; the transfer is not proven.
  • ad hoc to paper Optimal-case assumption that the SNN classifies correctly at every timestep
    Lemma 2 (Section III, Eq. 9) assumes the network confidently classifies each sample at all timesteps, which is the desired outcome of training, not a premise.
  • domain assumption Surrogate gradients approximate true gradients well enough for training
    Section V-B1 defines arctan and fast-sigmoid surrogate gradients; standard in SNN training but not proven for this architecture.
  • domain assumption CIFAR10-AV and UrbanSound8K-AV datasets and preprocessing are correct
    Section V-B states data preparation follows [31]; no external validation or code is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spiking Neural Network Feature Discrimination Boosts Modality Fusion." pith.science (2026). https://pith.science/paper/WDIOVUDV

@misc{pith2026250210423,
  author       = {Pith},
  title        = {Pith review of: Spiking Neural Network Feature Discrimination Boosts Modality Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WDIOVUDV}},
  note         = {Machine review of arXiv:2502.10423}
}
read the original abstract

Feature discrimination is a crucial aspect of neural network design, as it directly impacts the network's ability to distinguish between classes and generalize across diverse datasets. The accomplishment of achieving high-quality feature representations ensures high intra-class separability and poses one of the most challenging research directions. While conventional deep neural networks (DNNs) rely on complex transformations and very deep networks to come up with meaningful feature representations, they usually require days of training and consume significant energy amounts. To this end, spiking neural networks (SNNs) offer a promising alternative. SNN's ability to capture temporal and spatial dependencies renders them particularly suitable for complex tasks, where multi-modal data are required. In this paper, we propose a feature discrimination approach for multi-modal learning with SNNs, focusing on audio-visual data. We employ deep spiking residual learning for visual modality processing and a simpler yet efficient spiking network for auditory modality processing. Lastly, we deploy a spiking multilayer perceptron for modality fusion. We present our findings and evaluate our approach against similar works in the field of classification challenges. To the best of our knowledge, this is the first work investigating feature discrimination in SNNs.

Figures

Figures reproduced from arXiv: 2502.10423 by the authors.

Figure 1
Figure 1. Overview of the proposed multi-modal spiking fusion architecture [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. To the best of our knowledge, this is the first work attempt￾ing to introduce a feature discrimination approach in SNNs, which is further evaluated in the practical task of audio-visual classification. II. RELATED WORKS In this section, we explore recent research trends in spik￾ing architectures, multi-modal classification techniques, and feature discrimination approaches. Building on the aforemen￾tioned research fi… view at source ↗
Figure 2
Figure 2. Architecture of the L2-ActSpikeNet visual feature extraction network, incorporating residual learning with spiking neurons and our proposed L2 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figures from the paper (6 more)
Figure 3
Figure 3. Figure 3: Comparison of Fig. A the standard residual block in conven [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]
Figure 4
Figure 4. Figure 4: The proposed spiking audio feature extraction network, with the L2 normalization layer for audio feature discrimination. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Architecture of the Spiking Multilayer Perceptron (SMLP) for multi [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Cosine distance matrices of CIFAR10-AV features comparing L2-normalized and vanilla models. Figs. 6a, 6c, 6e, 6g, show L2-normalized visual/audio [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: The normalized confusion matrix of the audio-visual model on the [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: Cosine distance matrices of UrbanSound8K-AV features comparing L2-normalized and vanilla models. [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 52 canonical work pages

  1. [31]

    Transformer-based spiking neural networks for multimodal audiovisual classification,

    L. Guo, Z. Gao, J. Qu, S. Zheng, R. Jiang, Y . Lu, and H. Qiao, “Transformer-based spiking neural networks for multimodal audiovisual classification,” IEEE Transactions on Cognitive and Developmental Systems, vol. 16, no. 3, pp. 1077–1086, 2024

    work page 2024

  2. [63]

    Deep feature space: A geometrical perspective,

    I. Kansizoglou, L. Bampis, and A. Gasteratos, “Deep feature space: A geometrical perspective,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 10, pp. 6823–6838, 2022

    work page 2022

  3. [1]

    Evaluating the energy efficiency of deep convolutional neural networks on cpus and gpus,

    D. Li, X. Chen, M. Becchi, and Z. Zong, “Evaluating the energy efficiency of deep convolutional neural networks on cpus and gpus,” in 2016 IEEE International Conferences on Big Data and Cloud Computing (BDCloud), Social Computing and Networking (SocialCom), Sustainable Computing and Communications (SustainCom) (BDCloud- SocialCom-SustainCom), pp. 477–484, 2016

    work page 2016

  4. [2]

    Efficient processing of deep neural networks: A tutorial and survey,

    V . Sze, Y .-H. Chen, T.-J. Yang, and J. S. Emer, “Efficient processing of deep neural networks: A tutorial and survey,” Proceedings of the IEEE , vol. 105, no. 12, pp. 2295–2329, 2017

    2017

  5. [3]

    Going deeper with convolutions,

    C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,” in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1–9, 2015

    2015

  6. [4]

    Go- ing deeper with image transformers,

    H. Touvron, M. Cord, A. Sablayrolles, G. Synnaeve, and H. J ´egou, “Go- ing deeper with image transformers,” in Proceedings of the IEEE/CVF international conference on computer vision , pp. 32–42, 2021

    work page 2021

  7. [5]

    Deep residual learning for image recognition,

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

    2016

  8. [6]

    Convolutional neural networks at constrained time cost,

    K. He and J. Sun, “Convolutional neural networks at constrained time cost,” in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 5353–5360, 2015

    work page 2015

Show all 70 references
  1. [7]

    Highway networks,

    R. K. Srivastava, K. Greff, and J. Schmidhuber, “Highway networks,” arXiv preprint arXiv:1505.00387 , 2015

    2015 arXiv

  2. [8]

    Deep residual learning for image recognition: A survey,

    M. Shafiq and Z. Gu, “Deep residual learning for image recognition: A survey,” Applied Sciences, vol. 12, no. 18, p. 8972, 2022

    2022

  3. [9]

    Residual networks of residual networks: Multilevel residual networks,

    K. Zhang, M. Sun, T. X. Han, X. Yuan, L. Guo, and T. Liu, “Residual networks of residual networks: Multilevel residual networks,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 28, no. 6, pp. 1303–1314, 2017

    2017

  4. [10]

    Towards spike-based machine intelligence with neuromorphic computing,

    K. Roy, A. Jaiswal, and P. Panda, “Towards spike-based machine intelligence with neuromorphic computing,” Nature, vol. 575, no. 7784, pp. 607–617, 2019

    2019

  5. [11]

    Hussaini, Spiking Neural Networks for Scalable Visual Place Recog- nition

    S. Hussaini, Spiking Neural Networks for Scalable Visual Place Recog- nition. PhD thesis, Queensland University of Technology, 2024. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 11

    2024

  6. [12]

    The spike response model,

    W. Gerstner, “The spike response model,” The Handbook of Biological Physics, vol. 4, pp. 469–516, 1999

    1999

  7. [13]

    A hybrid spiking neural network reinforcement learning agent for energy-efficient object manipulation,

    K. M. Oikonomou, I. Kansizoglou, and A. Gasteratos, “A hybrid spiking neural network reinforcement learning agent for energy-efficient object manipulation,” Machines, vol. 11, no. 2, p. 162, 2023

    2023

  8. [14]

    A hybrid rein- forcement learning approach with a spiking actor network for efficient robotic arm target reaching,

    K. M. Oikonomou, I. Kansizoglou, and A. Gasteratos, “A hybrid rein- forcement learning approach with a spiking actor network for efficient robotic arm target reaching,” IEEE Robotics and Automation Letters , vol. 8, no. 5, pp. 3007–3014, 2023

    2023

  9. [15]

    Reinforcement co-learning of deep and spiking neural networks for energy-efficient mapless nav- igation with neuromorphic hardware,

    G. Tang, N. Kumar, and K. P. Michmizos, “Reinforcement co-learning of deep and spiking neural networks for energy-efficient mapless nav- igation with neuromorphic hardware,” in 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pp. 6090–6097, 2020

    2020

  10. [16]

    Sfod: Spiking fusion object detector,

    Y . Fan, W. Zhang, C. Liu, M. Li, and W. Lu, “Sfod: Spiking fusion object detector,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 17191–17200, 2024

    2024

  11. [17]

    Swat: A spiking neural network training algorithm for classification problems,

    J. J. Wade, L. J. McDaid, J. A. Santos, and H. M. Sayers, “Swat: A spiking neural network training algorithm for classification problems,” IEEE Transactions on Neural Networks, vol. 21, no. 11, pp. 1817–1830, 2010

    2010

  12. [18]

    Supervised learning in spiking neural networks with resume: sequence learning, classification, and spike shifting,

    F. Ponulak and A. Kasi ´nski, “Supervised learning in spiking neural networks with resume: sequence learning, classification, and spike shifting,” Neural computation, vol. 22, no. 2, pp. 467–510, 2010

    2010

  13. [19]

    Online deterministic annealing for classification and clustering,

    C. N. Mavridis and J. S. Baras, “Online deterministic annealing for classification and clustering,” IEEE Transactions on Neural Networks and Learning Systems , vol. 34, no. 10, pp. 7125–7134, 2023

    2023

  14. [20]

    Optimal ann-snn conversion for fast and accurate inference in deep spiking neural networks,

    J. Ding, Z. Yu, Y . Tian, and T. Huang, “Optimal ann-snn conversion for fast and accurate inference in deep spiking neural networks,” arXiv preprint arXiv:2105.11654, 2021

    2021 arXiv

  15. [21]

    Dnn-snn co-learning for sustainable symbol detection in 5g systems on loihi chip,

    S. Liu, Y . Liang, and Y . Yi, “Dnn-snn co-learning for sustainable symbol detection in 5g systems on loihi chip,” IEEE Transactions on Sustainable Computing, vol. 9, no. 2, pp. 170–181, 2024

    2024

  16. [22]

    Attention spiking neural networks,

    M. Yao, G. Zhao, H. Zhang, Y . Hu, L. Deng, Y . Tian, B. Xu, and G. Li, “Attention spiking neural networks,”IEEE transactions on pattern analysis and machine intelligence , vol. 45, no. 8, pp. 9393–9410, 2023

    2023

  17. [23]

    Event-based multimodal spiking neural network with attention mechanism,

    Q. Liu, D. Xing, L. Feng, H. Tang, and G. Pan, “Event-based multimodal spiking neural network with attention mechanism,” in ICASSP 2022- 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 8922–8926, IEEE, 2022

    2022

  18. [24]

    A bio-inspired elderly action recognition system for ambient assisted living,

    K. M. Oikonomou, I. Kansizoglou, I. T. Papapetros, and A. Gasteratos, “A bio-inspired elderly action recognition system for ambient assisted living,” in 2023 18th International Workshop on Cellular Nanoscale Networks and their Applications (CNNA) , pp. 1–6, IEEE, 2023

    2023

  19. [25]

    Enabling spike-based backpropagation for training deep neural network architec- tures,

    C. Lee, S. S. Sarwar, P. Panda, G. Srinivasan, and K. Roy, “Enabling spike-based backpropagation for training deep neural network architec- tures,” Frontiers in neuroscience, vol. 14, p. 497482, 2020

    2020

  20. [26]

    Going deeper in spiking neural networks: Vgg and residual architectures,

    A. Sengupta, Y . Ye, R. Wang, C. Liu, and K. Roy, “Going deeper in spiking neural networks: Vgg and residual architectures,” Frontiers in neuroscience, vol. 13, p. 95, 2019

    2019

  21. [27]

    Spikingresformer: Bridging resnet and vision transformer in spiking neural networks,

    X. Shi, Z. Hao, and Z. Yu, “Spikingresformer: Bridging resnet and vision transformer in spiking neural networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 5610–5619, 2024

    2024

  22. [28]

    Sensory modalities are not separate modal- ities: plasticity and interactions,

    S. Shimojo and L. Shams, “Sensory modalities are not separate modal- ities: plasticity and interactions,” Current opinion in neurobiology , vol. 11, no. 4, pp. 505–509, 2001

    2001

  23. [29]

    Event-based multimodal spiking neural network with attention mechanism,

    Q. Liu, D. Xing, L. Feng, H. Tang, and G. Pan, “Event-based multimodal spiking neural network with attention mechanism,” in ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 8922–8926, 2022

    2022

  24. [30]

    Stdp based unsupervised multimodal learning with cross-modal processing in spiking neural networks,

    N. Rathi and K. Roy, “Stdp based unsupervised multimodal learning with cross-modal processing in spiking neural networks,” IEEE Transactions on Emerging Topics in Computational Intelligence , vol. 5, no. 1, pp. 143–153, 2021

    2021

  25. [32]

    Advancing neuromorphic com- puting with loihi: A survey of results and outlook,

    M. Davies, A. Wild, G. Orchard, Y . Sandamirskaya, G. A. F. Guerra, P. Joshi, P. Plank, and S. R. Risbud, “Advancing neuromorphic com- puting with loihi: A survey of results and outlook,” Proceedings of the IEEE, vol. 109, no. 5, pp. 911–934, 2021

    2021

  26. [33]

    An efficient spiking neural network for recognizing gestures with a dvs camera on the loihi neuromorphic processor,

    R. Massa, A. Marchisio, M. Martina, and M. Shafique, “An efficient spiking neural network for recognizing gestures with a dvs camera on the loihi neuromorphic processor,” in 2020 International Joint Conference on Neural Networks (IJCNN) , pp. 1–9, IEEE, 2020

    2020

  27. [34]

    Simplified spiking neural network architecture and stdp learning algorithm applied to image classification,

    T. Iakymchuk, A. Rosado-Mu ˜noz, J. F. Guerrero-Mart´ınez, M. Bataller- Mompe´an, and J. V . Franc´es-V´ıllora, “Simplified spiking neural network architecture and stdp learning algorithm applied to image classification,” EURASIP Journal on Image and Video Processing, vol. 201...

    2015

  28. [35]

    A biologically plausible supervised learning method for spiking neural networks using the symmetric stdp rule,

    Y . Hao, X. Huang, M. Dong, and B. Xu, “A biologically plausible supervised learning method for spiking neural networks using the symmetric stdp rule,” Neural Networks, vol. 121, pp. 387–395, 2020

    2020

  29. [36]

    Stdp-based spiking deep convolutional neural networks for object recognition,

    S. R. Kheradpisheh, M. Ganjtabesh, S. J. Thorpe, and T. Masquelier, “Stdp-based spiking deep convolutional neural networks for object recognition,” Neural Networks, vol. 99, pp. 56–67, 2018

    2018

  30. [37]

    Optimal ann- snn conversion for high-accuracy and ultra-low-latency spiking neural networks,

    T. Bu, W. Fang, J. Ding, P. Dai, Z. Yu, and T. Huang, “Optimal ann- snn conversion for high-accuracy and ultra-low-latency spiking neural networks,” arXiv preprint arXiv:2303.04347 , 2023

    2023 arXiv

  31. [38]

    A universal ann-to- snn framework for achieving high accuracy and low latency deep spiking neural networks,

    Y . Wang, H. Liu, M. Zhang, X. Luo, and H. Qu, “A universal ann-to- snn framework for achieving high accuracy and low latency deep spiking neural networks,” Neural Networks, vol. 174, p. 106244, 2024

    2024

  32. [39]

    A new ann- snn conversion method with high accuracy, low latency and good robustness.,

    B. Wang, J. Cao, J. Chen, S. Feng, and Y . Wang, “A new ann- snn conversion method with high accuracy, low latency and good robustness.,” in IJCAI, pp. 3067–3075, 2023

    2023

  33. [40]

    Temporal efficient training of spiking neural network via gradient re-weighting,

    S. Deng, Y . Li, S. Zhang, and S. Gu, “Temporal efficient training of spiking neural network via gradient re-weighting,” arXiv preprint arXiv:2202.11946, 2022

    2022 arXiv

  34. [41]

    Membrane potential batch normalization for spiking neural networks,

    Y . Guo, Y . Zhang, Y . Chen, W. Peng, X. Liu, L. Zhang, X. Huang, and Z. Ma, “Membrane potential batch normalization for spiking neural networks,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , pp. 19420–19430, 2023

    2023

  35. [42]

    Towards memory-and time-efficient backpropagation for training spiking neural networks,

    Q. Meng, M. Xiao, S. Yan, Y . Wang, Z. Lin, and Z.-Q. Luo, “Towards memory-and time-efficient backpropagation for training spiking neural networks,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , pp. 6166–6176, 2023

    2023

  36. [43]

    Temporal-coded spiking neural networks with dynamic firing threshold: Learning with event-driven backpropagation,

    W. Wei, M. Zhang, H. Qu, A. Belatreche, J. Zhang, and H. Chen, “Temporal-coded spiking neural networks with dynamic firing threshold: Learning with event-driven backpropagation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , pp. 10552– 10562, 2023

    2023

  37. [44]

    Going deeper with directly-trained larger spiking neural networks,

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

    2021

  38. [45]

    Deep residual learning in spiking neural networks,

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

    2021

  39. [46]

    Spikingformer: Spike-driven residual learning for transformer-based spiking neural network,

    C. Zhou, L. Yu, Z. Zhou, Z. Ma, H. Zhang, H. Zhou, and Y . Tian, “Spikingformer: Spike-driven residual learning for transformer-based spiking neural network,” arXiv preprint arXiv:2304.11954 , 2023

    2023

  40. [47]

    Spiking deep residual networks,

    Y . Hu, H. Tang, and G. Pan, “Spiking deep residual networks,” IEEE Transactions on Neural Networks and Learning Systems , vol. 34, no. 8, pp. 5200–5205, 2021

    2021

  41. [48]

    Multimodal feature fusion for robust event detection in web videos,

    P. Natarajan, S. Wu, S. Vitaladevuni, X. Zhuang, S. Tsakalidis, U. Park, R. Prasad, and P. Natarajan, “Multimodal feature fusion for robust event detection in web videos,” in 2012 IEEE Conference on Computer Vision and Pattern Recognition, pp. 1298–1305, IEEE, 2012

    2012

  42. [49]

    Recurrent neural networks for emotion recognition in video,

    S. Ebrahimi Kahou, V . Michalski, K. Konda, R. Memisevic, and C. Pal, “Recurrent neural networks for emotion recognition in video,” in Pro- ceedings of the 2015 ACM on international conference on multimodal interaction, pp. 467–474, 2015

    2015

  43. [50]

    Context-sensitive learning for enhanced audiovisual emotion classification,

    A. Metallinou, M. Wollmer, A. Katsamanis, F. Eyben, B. Schuller, and S. Narayanan, “Context-sensitive learning for enhanced audiovisual emotion classification,” IEEE Transactions on Affective Computing , vol. 3, no. 2, pp. 184–198, 2012

    2012

  44. [51]

    Audiovisual classification of vocal outbursts in human conversation using long-short-term memory networks,

    F. Eyben, S. Petridis, B. Schuller, G. Tzimiropoulos, S. Zafeiriou, and M. Pantic, “Audiovisual classification of vocal outbursts in human conversation using long-short-term memory networks,” in 2011 IEEE International Conference on Acoustics, Speech and Signal Processing (ICA...

    2011

  45. [52]

    An active learning paradigm for online audio-visual emotion recognition,

    I. Kansizoglou, L. Bampis, and A. Gasteratos, “An active learning paradigm for online audio-visual emotion recognition,” IEEE Transac- tions on Affective Computing , vol. 13, no. 2, pp. 756–768, 2022

    2022

  46. [53]

    Audio visual speech recognition with multimodal recurrent neural networks,

    W. Feng, N. Guan, Y . Li, X. Zhang, and Z. Luo, “Audio visual speech recognition with multimodal recurrent neural networks,” in 2017 International Joint Conference on neural networks (IJCNN) , pp. 681– 688, IEEE, 2017

    2017

  47. [54]

    Recurrent neural network transducer for audio-visual speech recognition,

    T. Makino, H. Liao, Y . Assael, B. Shillingford, B. Garcia, O. Braga, and O. Siohan, “Recurrent neural network transducer for audio-visual speech recognition,” in 2019 IEEE automatic speech recognition and understanding workshop (ASRU) , pp. 905–912, IEEE, 2019. JOURNAL OF LAT...

    2019

  48. [55]

    Stdp based unsupervised multimodal learning with cross-modal processing in spiking neural networks,

    N. Rathi and K. Roy, “Stdp based unsupervised multimodal learning with cross-modal processing in spiking neural networks,” IEEE Transactions on Emerging Topics in Computational Intelligence , vol. 5, no. 1, pp. 143–153, 2018

    2018

  49. [56]

    Deep learning face representa- tion by joint identification-verification,

    Y . Sun, Y . Chen, X. Wang, and X. Tang, “Deep learning face representa- tion by joint identification-verification,” Advances in neural information processing systems, vol. 27, 2014

    2014

  50. [57]

    A discriminative feature learning approach for deep face recognition,

    Y . Wen, K. Zhang, Z. Li, and Y . Qiao, “A discriminative feature learning approach for deep face recognition,” in Computer vision–ECCV 2016: 14th European conference, amsterdam, the netherlands, October 11–14, 2016, proceedings, part VII 14 , pp. 499–515, Springer, 2016

    2016

  51. [58]

    Semi-supervised discriminative classification robust to sample-outliers and feature-noises,

    E. Adeli, K.-H. Thung, L. An, G. Wu, F. Shi, T. Wang, and D. Shen, “Semi-supervised discriminative classification robust to sample-outliers and feature-noises,” IEEE transactions on pattern analysis and machine intelligence, vol. 41, no. 2, pp. 515–522, 2018

    2018

  52. [59]

    Large-margin softmax loss for convolutional neural networks,

    W. Liu, Y . Wen, Z. Yu, and M. Yang, “Large-margin softmax loss for convolutional neural networks,” arXiv preprint arXiv:1612.02295, 2016

    2016 arXiv

  53. [60]

    Deep hyperspherical learning,

    W. Liu, Y .-M. Zhang, X. Li, Z. Yu, B. Dai, T. Zhao, and L. Song, “Deep hyperspherical learning,” Advances in neural information processing systems, vol. 30, 2017

    2017

  54. [61]

    Arcface: Additive angular margin loss for deep face recognition,

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou, “Arcface: Additive angular margin loss for deep face recognition,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 4690–4699, 2019

    2019

  55. [62]

    Sphereface: Deep hypersphere embedding for face recognition,

    W. Liu, Y . Wen, Z. Yu, M. Li, B. Raj, and L. Song, “Sphereface: Deep hypersphere embedding for face recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 212– 220, 2017

    2017

  56. [64]

    Automatic differentiation in pytorch,

    A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer, “Automatic differentiation in pytorch,” in NIPS-W, 2017

    2017

  57. [65]

    Adam: A method for stochastic optimization,

    D. P. Kingma, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  58. [66]

    Incorporating learnable membrane time constant to enhance learning of spiking neural networks,

    W. Fang, Z. Yu, Y . Chen, T. Masquelier, T. Huang, and Y . Tian, “Incorporating learnable membrane time constant to enhance learning of spiking neural networks,” in Proceedings of the IEEE/CVF international conference on computer vision , pp. 2661–2671, 2021

    2021

  59. [67]

    Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation,

    N. Rathi, G. Srinivasan, P. Panda, and K. Roy, “Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation,” arXiv preprint arXiv:2005.01807 , 2020

    2005 arXiv

  60. [68]

    Diet-snn: Direct input encoding with leakage and threshold optimization in deep spiking neural networks,

    N. Rathi and K. Roy, “Diet-snn: Direct input encoding with leakage and threshold optimization in deep spiking neural networks,” arXiv preprint arXiv:2008.03658, 2020

    2008 arXiv

  61. [69]

    Differentiable spike: Rethinking gradient-descent for training spiking neural networks,

    Y . Li, Y . Guo, S. Zhang, S. Deng, Y . Hai, and S. Gu, “Differentiable spike: Rethinking gradient-descent for training spiking neural networks,” Advances in Neural Information Processing Systems, vol. 34, pp. 23426– 23439, 2021

    2021

  62. [70]

    Spikformer: When spiking neural network meets transformer,

    Z. Zhou, Y . Zhu, C. He, Y . Wang, S. Yan, Y . Tian, and L. Yuan, “Spikformer: When spiking neural network meets transformer,” arXiv preprint arXiv:2209.15425, 2022. Oikonomou Katerina Maria is a Ph.D. Candidate at the Democritus University of Thrace, Xanthi, Greece, working w...

    2022 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.