Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Spike-TBR: a Noise Resilient Neuromorphic Event Representation

T0 review · 4 major / 6 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 Spiking filter keeps 94% accuracy where TBR collapses to 40%.

desk verdict Useful noise-robust encoding idea, but Algorithm 2 omits the decay that makes it work—so the headline result is not reproducible as written. read the letter →

arxiv 2506.04817 v2 pith:BVXV43TK submitted 2025-06-05 cs.CV

Gabriele Magrini , Federico Becattini , Luca Cultrera , Lorenzo Berlincioni , Pietro Pala , Alberto Del Bimbo This is my paper · ORCID
classification cs.CV
keywords eventcamerasneuromorphicvisionrepresentationspikingneuralnetworksnoiserobustnesstemporalbinaryLIFneuronsactionrecognition
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 Spike-TBR, a frame-based event-camera encoding that replaces the binary event/no-event slices of Temporal Binary Representation (TBR) with spikes produced by a layer of spiking neurons. The aim is to fix TBR's vulnerability to noise, where a single noise event in the most recent slice flips the most significant bit and drastically changes the representation. The claim is that because a neuron's membrane must accumulate enough events to cross a threshold before firing, isolated noise events are filtered out before they can corrupt the encoding. On clean data Spike-TBR matches or improves on TBR across four action and lip-reading datasets, and under injected noise it holds accuracy far better, for example 93.94% versus 40.50% at 1% noise on DVSGesture-128. If true, this gives a simple way to make event-based vision pipelines resilient to noisy sensors without changing the downstream architecture.

What carries the argument

The central object is the per-pixel spiking layer inserted between raw events and the TBR code: each pixel is assigned one neuron whose membrane accumulates event arrivals and fires a spike only when the accumulated potential reaches the threshold, after which the potential resets. The spike train sampled every $\Delta t$ replaces TBR's raw event/no-event indicator, and the same stacked binary-to-decimal conversion produces the final frame. The threshold-and-decay dynamics are the noise filter: isolated noise events rarely accumulate enough potential to fire, while real motion produces enough events to cross the threshold. The decay parameter $\beta$ (with $V_{th}=1.1$ and dataset-specific $\beta$ between 0.5 and 0.9) controls how much history the membrane keeps and is tuned per dataset, and the parametric LIF variant makes the time constant learnable.

What would settle it

Record a fixed scene with a neuromorphic camera whose noise level and temporal correlation are measured, then run Spike-TBR and TBR on the same stream using the paper's default $V_{th}=1.1$ and $\beta$ values; if TBR matches or beats Spike-TBR under real sensor noise, or if Spike-TBR's advantage disappears when $\beta$ is perturbed by a small amount, the claimed generality of the noise filter is unsupported.

Watch

Extended reading notes

Core claim

The central discovery is that interpreting each pixel's event history through a spiking neuron converts an all-or-nothing binary code into an accumulation-and-threshold code, and that this conversion is what buys noise resilience. TBR encodes the last N time slices as a binary string per pixel, so noise in the newest slice contaminates the most significant bit. Spike-TBR feeds events into a per-pixel leaky membrane; the membrane integrates events, decays with rate $\beta$, and emits a spike only when the potential crosses a threshold $V_{th}=1.1$; the spikes over N slices are then stacked and binary-to-decimal converted exactly as in TBR. With four neuron variants (LIF, recurrent LIF, light-refractory LIF, and parametric LIF), the encoding matches or exceeds TBR on clean data and degrades far more slowly under synthetic and recorded sensor noise. The paper also shows that training TBR with noise augmentation only helps at the exact noise level seen in training, whereas Spike-TBR generalizes across noise levels without being trained on noise.

Load-bearing premise

The noise resilience rests on hand-set parameters—threshold $V_{th}=1.1$ and dataset-specific decay $\beta$ between 0.5 and 0.9—remaining appropriate for the sensor and task; the paper shows $\beta$ has opposite effects on dense versus sparse event streams and notes the encoding may need fine-tuning for different neuromorphic cameras, but does not measure how fast performance degrades when these parameters are off.

Editorial extensions

If this is right

  • Spike-TBR can be dropped into existing TBR pipelines by replacing the TBR encoder with a spiking layer, leaving the downstream I3D classifier unchanged.
  • On clean data the representation matches or improves on TBR on DVSGesture-128 and DVSLip, while on NCaltech-101 it trails top augmented methods but remains close to TBR.
  • Under injected noise up to 1–2%, LIF variants keep useful accuracy where TBR collapses to near chance, and this holds when noise comes from recorded background activity of an event camera.
  • Training TBR with noise augmentation at a fixed level does not generalize: the model overfits that noise level and collapses at higher levels, whereas Spike-TBR shows resilience without noise-aware training.
  • The added spiking layer costs negligible computation, around 50K accumulated operations versus 194M multiply-accumulates for the I3D network, so the noise filter is nearly free.

Reading between the lines

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

  • A natural testable extension is to make $V_{th}$ itself learnable or adaptive per sensor, since the paper fixes $V_{th}=1.1$ and tunes $\beta$ per dataset; the authors' stated camera-dependence limitation suggests this is the first parameter to address.
  • Because the paper's noise model is uniform random injection, real sensor noise with spatial or temporal structure (hot pixels, bursty noise) may behave differently; a structured-noise benchmark would separate the benefit of accumulation from the benefit of the specific noise distribution.
  • The same spike-layer-before-encoding idea should transfer to other TBR-like representations and to tasks beyond classification, such as optical flow or tracking, where frame-level resilience could reduce the need for event preprocessing.
  • The opposite effect of $\beta$ on dense versus sparse event streams suggests that a single adaptive decay schedule, rather than one fixed value, could make the representation robust across a wider range of scene dynamics.
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 Spike-TBR, an event-stream encoding that replaces TBR's binary event/no-event slices with spikes emitted by a per-pixel spiking-neuron layer whose membrane integrates events and fires when a threshold is exceeded. Four neuron variants (LIF, RecLIF, LRLIF, PLIF) are evaluated on DVSGesture-128, DVSLip, NCaltech-101, and MICCGesture, using an I3D classifier. The central claim is that Spike-TBR preserves or improves clean-data accuracy while being substantially more robust than TBR under injected event-stream noise, including a real-noise test recorded with a Prophesee EVK4 camera.

Significance. If the central claim holds, Spike-TBR would be a simple, practical bridge between frame-based and spike-based event processing: it keeps TBR's compact frame representation while inheriting noise filtering from SNN temporal dynamics. The paper contains several strengths worth crediting: the TBR baseline is rerun under the same pipeline, the noise study includes a real-sensor noise test (Table 6), the β ablation (Table 7) studies the main hyperparameter, and the computational overhead of the SNN layer is quantified (Section 4.4). However, the formal algorithm as published does not implement the decay mechanism that the noise-robustness argument relies on, and the abstract's clean-data claim is contradicted by the paper's own tables. These issues are load-bearing and need to be fixed before the claims can be accepted.

major comments (4)
  1. [Section 3.2, Algorithm 2] Algorithm 2 omits the membrane leak/decay that is the stated mechanism of noise filtering. Eqs. (4)-(5) define the LIF dynamics with decay rate β = 1 - 1/τ_m, and the text explains that the membrane 'forgets' isolated events. However, lines 5-11 of Algorithm 2 perform only V(xi,yi) ← V(xi,yi) + w(pi), followed by a spike/reset when V ≥ Vthr = 1.1; there is no V ← βV update, no per-timestep leak, and no decay between events with different timestamps. With w=1 and Vthr=1.1, the published procedure is a per-pixel counter that fires after two events anywhere within the processed window, not a leaky integrate-and-fire neuron. This is an internal inconsistency between the mathematical model and the procedural definition, and it directly affects reproducibility of the headline result in Table 5 (93.94% at 1% noise on DVSGesture-128) and the noise curves in Fig. 4. Please correct Algorithm 2 to include the leakage term and clarify how continuous event timestamps are mapped to the discrete time steps used in the decay update.
  2. [Abstract, Section 4.1, Tables 1-4] The abstract's claim that Spike-TBR shows 'superior performance in noise-affected scenarios while improving the results on clean data' is not supported by the reported clean-data results. On NCaltech-101 all Spike-TBR variants are below TBR (68.08-69.28 vs. 72.50 in Table 3), on MICCGesture only the LIF variant ties TBR at 73.16 while the other three variants are clearly lower (Table 4), and on DVSGesture-128 the variants tie TBR at 99.62 (Table 1). Only on DVSLip do all variants improve on TBR (Table 2). The honest summary is 'comparable or better on three datasets, worse on NCaltech-101.' Please revise the abstract and conclusion to match the tables.
  3. [Section 4.2, Fig. 4 and Table 5] The training protocol for the noise-robustness experiments is underspecified. The paper states that Spike-TBR demonstrates resilience 'even when training on clean data,' and Table 5 is presented as a control study where TBR is trained with noise augmentation. However, the text does not explicitly state for Fig. 4 and Table 5 whether all Spike-TBR variants and the vanilla TBR baseline were trained on clean data only, or whether any noise augmentation or validation-set noise tuning was used. Please state the training condition for each curve, because the comparison between 'TBR + noise augmentation' and 'Spike-TBR' in Table 5 is only meaningful if the Spike-TBR models were not exposed to the same augmentation.
  4. [Section 4.3, Table 7, and Section 5] The robustness claim is strongly dependent on the hand-set decay rate β, which is tuned per dataset on a held-out validation set. Table 7 shows that the optimal β differs sharply between DVSGesture-128 and DVSLip, and that a poor β choice can reduce clean accuracy from 75.91% to 45.99% (DVSLip). The authors acknowledge in Section 5 that the encoding 'may require fine-tuning when applied to different neuromorphic cameras.' Given that the paper's central contribution is a noise-resilient representation, the sensitivity of the method to β should be quantified for the noise levels used in Fig. 4, and the claim of resilience should be framed with this parameter-dependence made explicit.
minor comments (6)
  1. [Section 3.2] The sentence 'the input is left without decay' is ambiguous and appears to contradict Eq. (4) and the description of the membrane decaying over time; please clarify what quantity is not decayed.
  2. [Algorithm 2] Line 8 writes a spike at the continuous event timestamp ti, while lines 13-14 treat S as a slice-indexed binary sequence S(x,y,t1)...S(x,y,tN). The mapping from event timestamps to temporal slices is not specified; please define it explicitly.
  3. [Section 3.2, Eqs. (2)-(3)] The threshold is denoted Vth in Eqs. (2)-(3), Vthr in Algorithm 2, and Vth = 1.1 in the text; please unify the notation.
  4. [Table 5] Several entries in Table 5 are missing separators, e.g., '99.6280.30 40.50 9.469.09' and '98.86 98.4899.24 9.46 9.09'; these should be formatted as separate columns.
  5. [Section 4.2] The noise injection description 'we add an event every Δt ms with probability p for each pixel' should be stated as an independent Bernoulli draw at each Δt interval per pixel, to make the simulation exactly reproducible.
  6. [General] The paper refers to 'Algorithm 2' but no Algorithm 1 is present; please renumber or add the missing algorithm.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Spike-TBR is an empirical encoding evaluated against external benchmarks and a re-run TBR baseline; beta/threshold tuning is not a disguised prediction and the TBR self-citation is not load-bearing.

full rationale

The paper's central claim is empirical: Spike-TBR, an SNN-based encoding, retains accuracy under injected and real event-stream noise better than TBR. The evaluation is self-contained: four public datasets, an I3D classifier, a re-run TBR baseline, and comparisons to published numbers. The noise-resilience result is not derived from a fitted parameter renamed as a prediction. The only tuned parameter, the decay rate beta, is selected on a held-out validation set (Section 4.1: 'we tune the decay rate beta on a held-out validation set'), and the threshold is a fixed design choice (Vth = 1.1); neither is fit to the noise test conditions, and the paper additionally tests on real Prophesee EVK4 noise (Table 6) for which no fitting was performed. Self-citations to TBR [5] are used as a baseline to be improved, not as an unverified premise; the authors state 'we run all TBR experiments with our pipeline to establish a fair comparison', so the comparison does not reduce to the citation. The paper also self-identifies a genuine limitation in Section 5 ('the encoding strategy may require fine-tuning when applied to different neuromorphic cameras'), which is a generalization concern, not circularity. The skeptic's observation that Algorithm 2 omits the leak/decay term present in Eqs. 4-5 is a reproducibility/internal-consistency concern that could affect the reported numbers, but it is not a case of a derivation reducing to its own inputs, so it does not change the circularity score.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a handful of hand-set hyperparameters (threshold, decay rate, slice length) and on the assumption that synthetic random noise approximates real sensor noise. No new physical entities are introduced.

free parameters (6)
  • beta (membrane decay rate) = 0.5 (DVSGesture-128), 0.9 (DVSLip), 0.7 (NCaltech-101), 0.7 (MICCGesture)
    Tuned on a held-out validation set; controls how quickly the membrane potential decays and thereby how much noise is filtered.
  • Vth (spike threshold) = 1.1
    Set by hand to ensure a single event does not fire the neuron; isolated noise events are filtered.
  • delta t (time slice length) = 2.5 ms for DVSGesture-128, NCaltech-101, MICCGesture; 6.25 ms for DVSLip
    Chosen based on event-stream density; the paper says longer slices for sparser streams.
  • N (number of TBR bits) = 8
    Fixed for all datasets, determines the number of temporal slices per frame.
  • delta T (accumulation window) = 500 ms chunks for three datasets; full sequence for DVSLip
    Defines sample boundaries; non-overlapping chunks with majority voting at test time.
  • PLIF learnable membrane time constant = learned during training
    For the PLIF variant, tau_m is optimized by backpropagation rather than set by hand.
assumptions (3)
  • standard math LIF neuron subthreshold dynamics, Eq. 4: V[t] = V[t-1] - (1/tau_m)(V[t-1] - V_rest) + X[t]
    Taken from prior spiking neural network literature [24]; the paper modifies the reset rule for some variants.
  • domain assumption Event stream is modeled as (x,y,t,p) tuples with polarity p in {-1,+1}
    Standard event camera output model assumed in Section 3; no justification needed beyond the sensor's behavior.
  • domain assumption Noise injection model: for noise level p, an event is added at each pixel in each time slice with probability p, independently
    Used to evaluate robustness (Section 4.2); it assumes random i.i.d. noise, while real sensor noise may be spatially or temporally correlated. The authors partially address this with a real-noise recording.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spike-TBR: a Noise Resilient Neuromorphic Event Representation." pith.science (2026). https://pith.science/paper/BVXV43TK

@misc{pith2026250604817,
  author       = {Pith},
  title        = {Pith review of: Spike-TBR: a Noise Resilient Neuromorphic Event Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BVXV43TK}},
  note         = {Machine review of arXiv:2506.04817}
}
read the original abstract

Event cameras offer significant advantages over traditional frame-based sensors, including higher temporal resolution, lower latency and dynamic range. However, efficiently converting event streams into formats compatible with standard computer vision pipelines remains a challenging problem, particularly in the presence of noise. In this paper, we propose Spike-TBR, a novel event-based encoding strategy based on Temporal Binary Representation (TBR), addressing its vulnerability to noise by integrating spiking neurons. Spike-TBR combines the frame-based advantages of TBR with the noise-filtering capabilities of spiking neural networks, creating a more robust representation of event streams. We evaluate four variants of Spike-TBR, each using different spiking neurons, across multiple datasets, demonstrating superior performance in noise-affected scenarios while improving the results on clean data. Our method bridges the gap between spike-based and frame-based processing, offering a simple noise-resilient solution for event-driven vision applications.

Figures

Figures reproduced from arXiv: 2506.04817 by the authors.

Figure 1
Figure 1. Spike-TBR pipeline. As the raw event stream is generated by the sensor (left) the SNN membrane accumulates potential [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Pipeline experiment setup. Events are aggregated with Spike-TBR, generating a sequence of TBR frames. Then, a classifier as I3D is used. In this case, the output of each neuron is propagated back to itself, increasing the effect of spikes and further propagating them through time. This means that for an input X[t], the neu￾ron cell will get as actual input: i[t +1] = X[t] +y[t −1] (6) where y[t −1] is the neuron cel… view at source ↗
Figure 3
Figure 3. Impact of noise on binary event frames and filtering properties of Spik [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Noise-affected performances on the tested datasets based on the different spiking neuronal architecture. Plots (a), (b), (c), and (d) correspond to DVSGesture [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID 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. Drone Detection with Event Cameras

    cs.CV 2025-08 conditional novelty 2.0 of 10

    A survey of event camera-based drone detection that maps methods by data representation and covers tracking, forecasting, and propeller signature analysis.

Reference graph

Works this paper leans on

56 extracted references · 49 canonical work pages · cited by 1 Pith paper

  1. [1]

    Nguyen, T.-T

    A. Nguyen, T.-T. Do, D. G. Caldwell, N. G. Tsagarakis, Real-time 6dof pose relocalization for event cameras with stacked spatial lstm networks, in: Proceedings of the IEEE CVPR Workshops, 2019, pp. 0–0

    work page 2019

  2. [2]

    S. Miao, G. Chen, X. Ning, Y . Zi, K. Ren, Z. Bing, A. Knoll, Neuromor- phic vision datasets for pedestrian detection, action recognition, and fall detection, Frontiers in neurorobotics 13 (2019) 38

    work page 2019

  3. [3]

    Spatiotemporal Filtering for Event-Based Action Recognition

    R. Ghosh, A. Gupta, A. Nakagawa, A. Soares, N. Thakor, Spa- tiotemporal filtering for event-based action recognition, arXiv preprint arXiv:1903.07067 (2019). 8

    work page Pith review arXiv 2019

  4. [4]

    Cannici, M

    M. Cannici, M. Ciccone, A. Romanoni, M. Matteucci, A differentiable recurrent surface for asynchronous event-based data, in: ECCV , 2020

    work page 2020

  5. [5]

    S. U. Innocenti, F. Becattini, F. Pernici, A. Del Bimbo, Temporal binary representation for event-based action recognition, in: ICPR, IEEE, 2021, pp. 10426–10432

    work page 2021

  6. [6]

    Q. Liu, D. Xing, H. Tang, D. Ma, G. Pan, Event-based action recognition using motion information and spiking neural networks., in: IJCAI, 2021

    work page 2021

  7. [7]

    C. Lee, A. K. Kosta, A. Z. Zhu, K. Chaney, K. Daniilidis, K. Roy, Spike- flownet: event-based optical flow estimation with energy-efficient hybrid neural networks, in: ECCV , 2020, pp. 366–382

    work page 2020

  8. [8]

    Cuadrado, U

    J. Cuadrado, U. Ranc ¸on, B. R. Cottereau, F. Barranco, T. Masquelier, Optical flow estimation from event-based cameras and spiking neural net- works, Frontiers in Neuroscience 17 (2023) 1160034

    work page 2023

Show all 56 references
  1. [9]

    Gallego, T

    G. Gallego, T. Delbruck, G. Orchard, C. Bartolozzi, B. Taba, A. Censi, S. Leutenegger, A. Davison, J. Conradt, K. Daniilidis, et al., Event-based vision: A survey, arXiv preprint arXiv:1904.08405 (2019)

    2019 arXiv

  2. [10]

    Gehrig, A

    D. Gehrig, A. Loquercio, K. Derpanis, D. Scaramuzza, End-to-end learn- ing of representations for asynchronous event-based data, in: 2019 IEEE/CVF ICCV , IEEE Computer Society, Los Alamitos, CA, USA, 2019, pp. 5632–5642

    2019

  3. [11]

    Cannici, M

    M. Cannici, M. Ciccone, A. Romanoni, M. Matteucci, Asynchronous convolutional networks for object detection in neuromorphic cameras, in: Proceedings of the IEEE CVPR Workshops, 2019, pp. 0–0

    2019

  4. [12]

    Kaiser, A

    J. Kaiser, A. Friedrich, J. Tieck, D. Reichard, A. Roennau, E. Neftci, R. Dillmann, Embodied neuromorphic vision with event-driven random backpropagation, arXiv preprint arXiv:1904.04805 (2019)

    2019 arXiv

  5. [13]

    Maass, Networks of spiking neurons: the third generation of neural network models, Neural networks 10 (1997) 1659–1671

    W. Maass, Networks of spiking neurons: the third generation of neural network models, Neural networks 10 (1997) 1659–1671

    1997

  6. [14]

    W. S. McCulloch, W. Pitts, A logical calculus of the ideas immanent in nervous activity, bulletin of mathematical biophysics (1943) 115–133

    1943

  7. [15]

    C. Lee, A. K. Kosta, K. Roy, Fusion-flownet: Energy-efficient optical flow estimation using sensor fusion and deep fused spiking-analog net- work architectures, in: 2022 ICRA, IEEE, 2022, pp. 6504–6510

    2022

  8. [16]

    Cordone, B

    L. Cordone, B. Miramond, S. Ferrante, Learning from event cameras with sparse spiking convolutional neural networks, in: 2021 IJCNN, IEEE, 2021, pp. 1–8

    2021

  9. [17]

    Hagenaars, F

    J. Hagenaars, F. Paredes-Vall ´es, G. De Croon, Self-supervised learning of event-based optical flow with spiking neural networks, NeurIPS 34 (2021) 7167–7179

    2021

  10. [18]

    Z. Yang, Y . Wu, G. Wang, Y . Yang, G. Li, L. Deng, J. Zhu, L. Shi, Dash- net: A hybrid artificial and spiking neural network for high-speed object tracking, arXiv preprint arXiv:1909.12942 (2019)

    2019 arXiv

  11. [19]

    S. Negi, D. Sharma, A. K. Kosta, K. Roy, Best of both worlds: Hy- brid snn-ann architecture for event-based optical flow estimation, arXiv preprint arXiv:2306.02960 (2023)

    2023 arXiv

  12. [20]

    G. K. Cohen, Event-based feature detection, recognition and classifica- tion, Ph.D. thesis, Western Sydney University (Australia), 2015

    2015

  13. [21]

    S. S. Chowdhury, C. Lee, K. Roy, Towards understanding the effect of leak in spiking neural networks, Neurocomputing 464 (2021) 83–94

    2021

  14. [22]

    S. Park, D. Lee, S. Yoon, Noise-robust deep spiking neural networks with temporal information, in: 2021 58th ACM/IEEE DAC, IEEE, 2021

    2021

  15. [23]

    T. K. T. Thieu, R. Melnik, Effects of noise on leaky integrate-and-fire neuron models for neuromorphic computing applications, in: Computa- tional Science and Its Applications, Cham, 2022, pp. 3–18

    2022

  16. [24]

    Maass, C

    W. Maass, C. M. Bishop, Pulsed neural networks, MIT press, 2001

    2001

  17. [25]

    B. Yin, F. Corradi, S. M. Boht ´e, Effective and efficient computation with multiple-timescale spiking recurrent neural networks, in: International Conference on Neuromorphic Systems 2020, 2020, pp. 1–8

    2020

  18. [26]

    W. Fang, Z. Yu, Y . Chen, T. Masquelier, T. Huang, Y . Tian, Incorporating learnable membrane time constant to enhance learning of spiking neural networks, in: Proceedings of the IEEE/CVF ICCV , 2021, pp. 2661–2671

    2021

  19. [27]

    Carreira, A

    J. Carreira, A. Zisserman, Quo vadis, action recognition? a new model and the kinetics dataset, in: proceedings of the IEEE CVPR, 2017

    2017

  20. [28]

    Berlincioni, L

    L. Berlincioni, L. Cultrera, C. Albisani, L. Cresti, A. Leonardo, S. Pic- chioni, F. Becattini, A. Del Bimbo, Neuromorphic event-based facial expression recognition, in: Proceedings of the IEEE/CVF CVPR, 2023

    2023

  21. [29]

    Berlincioni, L

    L. Berlincioni, L. Cultrera, F. Becattini, A. Del Bimbo, Neuromorphic valence and arousal estimation, arXiv preprint arXiv:2401.16058 (2024)

    2024 arXiv

  22. [30]

    Becattini, L

    F. Becattini, L. Cultrera, L. Berlincioni, C. Ferrari, A. Leonardo, A. Del Bimbo, Neuromorphic facial analysis with cross-modal super- vision, arXiv preprint arXiv:2409.10213 (2024)

    2024 arXiv

  23. [31]

    M. Xiao, Q. Meng, Z. Zhang, D. He, Z. Lin, Online training through time for spiking neural networks, NeurIPS 35 (2022) 20717–20730

    2022

  24. [32]

    Kaiser, H

    J. Kaiser, H. Mostafa, E. Neftci, Synaptic plasticity dynamics for deep continuous local learning, Frontiers in Neuroscience 14 (2020) 424

    2020

  25. [33]

    Sch ¨one, N

    M. Sch ¨one, N. M. Sushma, J. Zhuge, C. Mayr, A. Subramoney, D. Kap- pel, Scalable event-by-event processing of neuromorphic sensory signals with deep state-space models, arXiv preprint arXiv:2404.18508 (2024)

    2024 arXiv

  26. [34]

    Subramoney, K

    A. Subramoney, K. K. Nazeer, M. Sch ¨one, C. Mayr, D. Kappel, Ef- ficient recurrent architectures through activity sparsity and sparse back- propagation through time, arXiv preprint arXiv:2206.06178 (2022)

    2022 arXiv

  27. [35]

    X. She, S. Dash, S. Mukhopadhyay, Sequence approximation using feed- forward spiking neural network for spatiotemporal learning: Theory and optimization methods, in: ICLR, 2022

    2022

  28. [36]

    Arjmand, Y

    C. Arjmand, Y . Xu, K. Shidqi, A. F. Dobrita, K. Vadivel, P. Detterer, M. Sifalakis, A. Yousefzadeh, G. Tang, Trip: Trainable region-of- interest prediction for hardware-efficient neuromorphic processing on event-based vision, arXiv preprint arXiv:2406.17483 (2024)

    2024 arXiv

  29. [37]

    C. Liu, X. Qi, E. Y . Lam, N. Wong, Fast classification and action recog- nition with event-based imaging, IEEE access 10 (2022) 55638–55649

    2022

  30. [38]

    Bulzomi, M

    H. Bulzomi, M. Schweiker, A. Gruel, J. Martinet, End-to-end neuromor- phic lip-reading, in: Proceedings of the IEEE/CVF CVPR, 2023

    2023

  31. [39]

    S. Yoo, E. Y .-J. Lee, Z. Wang, X. Wang, W. D. Lu, Rn-net: Reservoir nodes-enabled neuromorphic vision sensing network, Advanced Intelli- gent Systems (2024) 2400265

    2024

  32. [40]

    Z. Wang, Q. She, A. Smolic, Action-net: Multipath excitation for action recognition, in: Proceedings of the IEEE/CVF CVPR, 2021

    2021

  33. [41]

    Mesquida, M

    T. Mesquida, M. Dampfhoffer, T. Dalgaty, P. Vivet, A. Sironi, C. Posch, G2n2: Lightweight event stream classification with gru graph neural net- works, in: BMVC, 2023, p. 660

    2023

  34. [42]

    Zhang, J

    W. Zhang, J. Wang, Y . Luo, L. Yu, W. Yu, Z. He, Mtga: Multi-view tem- poral granularity aligned aggregation for event-based lip-reading, arXiv preprint arXiv:2404.11979 (2024)

    2024 arXiv

  35. [43]

    Dampfhoffer, T

    M. Dampfhoffer, T. Mesquida, Neuromorphic lip-reading with signed spiking gated recurrent units, in: Proceedings of the IEEE/CVF CVPR, 2024, pp. 2141–2151

    2024

  36. [44]

    C. Duan, J. Ding, S. Chen, Z. Yu, T. Huang, Temporal effective batch normalization in spiking neural networks, NeurIPS 35 (2022)

    2022

  37. [45]

    D. Zhao, Y . Zeng, Y . Li, Backeisnn: A deep spiking neural network with adaptive self-feedback and balanced excitatory–inhibitory neurons, Neural Networks 154 (2022) 68–77

    2022

  38. [46]

    Z. Zhou, Y . Zhu, C. He, Y . Wang, S. Yan, Y . Tian, L. Yuan, Spik- former: When spiking neural network meets transformer, arXiv preprint arXiv:2209.15425 (2022)

    2022 arXiv

  39. [47]

    Y . Ding, L. Zuo, M. Jing, P. He, Y . Xiao, Shrinking your timestep: To- wards low-latency neuromorphic object recognition with spiking neural networks, in: Proceedings of AAAI, volume 38, 2024, pp. 11811–11819

    2024

  40. [48]

    Y . Li, Y . Kim, H. Park, T. Geller, P. Panda, Neuromorphic data augmenta- tion for training spiking neural networks, in: ECCV , 2022, pp. 631–649

    2022

  41. [49]

    G. Shen, D. Zhao, Y . Zeng, Eventmix: An efficient data augmentation strategy for event-based learning, Information Sciences (2023) 119170

    2023

  42. [50]

    A. Amir, B. Taba, D. Berg, T. Melano, J. McKinstry, C. Di Nolfo, T. Nayak, A. Andreopoulos, G. Garreau, M. Mendoza, et al., A low power, fully event-based gesture recognition system, in: Proceedings of the IEEE CVPR, 2017

    2017

  43. [51]

    G. Tan, Y . Wang, H. Han, Y . Cao, F. Wu, Z.-J. Zha, Multi-grained spatio- temporal features perceived network for event-based lip-reading, in: Pro- ceedings of the IEEE/CVF CVPR, 2022, pp. 20094–20103

    2022

  44. [52]

    Orchard, A

    G. Orchard, A. Jayawant, G. K. Cohen, N. Thakor, Converting static image datasets to spiking neuromorphic datasets using saccades, Frontiers in Neuroscience 9 (2015)

    2015

  45. [53]

    Fei-Fei, R

    L. Fei-Fei, R. Fergus, P. Perona, Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories, in: CVPR Workshop, 2004, pp. 178–178

    2004

  46. [54]

    W. Fang, Y . Chen, J. Ding, Z. Yu, T. Masquelier, D. Chen, L. Huang, H. Zhou, G. Li, Y . Tian, Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence, Science Advances 9 (2023) eadi1480

    2023

  47. [55]

    G. Lenz, K. Chaney, S. B. Shrestha, O. Oubari, S. Picaud, G. Zarrella, Tonic: event-based datasets and transformations., 2021

    2021

  48. [56]

    G. Chen, P. Peng, G. Li, Y . Tian, Training full spike neural networks via auxiliary accumulation pathway, arXiv preprint arXiv:2301.11929 (2023)

    2023 arXiv

Pith tools

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