Pith. sign in

REVIEW 3 major objections 5 minor 40 references

ADSEQ: A delay-aware autograd-compatible framework for spike-event delivery in SNNs

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A spike-event queue can propagate exact gradients through delayed spikes, extending exact delay training beyond LIF-specific neuron models.

desk verdict Solid benchmark study undermined by a sign error in the core gradient derivation and missing code/validation. read the letter →

arxiv 2512.05906 v2 pith:WJLK7ST6 submitted 2025-12-05 cs.NE

Lennart P. L. Landsmeer , Amirreza Movahedin , Said Hamdioui , Christos Strydis This is my paper
classification cs.NE
keywords spikingneuralnetworkssynapticdelayseventqueuesexactgradientsautodifferentiationsurrogateneuromorphichardwaredelaylearning
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 sets out to show that synaptic delays in spiking neural networks can be trained with exact gradients using a sparse spike-event queue rather than dense surrogate signals. The authors derive a gradient-transfer identity — ∂x⁺/∂θ = ∂x⁻/∂θ − (1/τ) ∂t_post/∂θ, where t_post is the delayed arrival time — and argue it generalizes the EventProp jump condition to arbitrary first-order synaptic variables. They implement the identity in an autodifferentiable ML library with custom derivative definitions, and benchmark several queue data structures on CPUs, GPUs, tensor-processing units, and dataflow accelerators. A sympathetic reader would take the contribution as a new, more general route to delay learning in SNNs, with performance strongly shaped by queue implementation and hardware architecture.

What carries the argument

The central object is the autodifferentiable spike-event queue: a data structure that holds delayed spike deliveries together with their gradient messages. The identity ∂x⁺/∂θ = ∂x⁻/∂θ − (1/τ) ∂t_post/∂θ is the load-bearing piece — it converts an arrival-time derivative into a state gradient jump at delivery. The queue implementations (ring, FIFO ring, sorted array, binary heap, single-spike, etc.) are the testbed for how this gradient propagation performs on different hardware.

What would settle it

Simulate a single presynaptic spike traveling through a known delay to a postsynaptic first-order synapse; compute the gradient of the postsynaptic current at a fixed observation time with respect to the delay using the queue's custom derivative, and compare against a finite-difference of the discretized forward simulation at fine temporal resolution. Any mismatch beyond numerical tolerance (especially a sign flip) would rule out Eq. (15).

Watch

Extended reading notes

Core claim

The central claim is that the gradient of a postsynaptic state variable through a delayed spike can be computed by a simple jump correction: the derivative before arrival plus minus (1/τ) times the derivative of the arrival time. Because the arrival time is the presynaptic spike time plus the trainable delay, this couples delay gradients directly into the postsynaptic state. The paper asserts that this identity, implemented as custom jacobian-vector products on queue operations, lets an event queue carry exact gradient information without densifying spikes in time. It also claims this works for any first-order synapse model, going beyond the simplified LIF cells that existing exact-gradient

Load-bearing premise

The derivation rests on the jump condition (ẋ⁻ − ẋ⁺) = −1/τ and on transferring this continuous-time jump to a discrete-time queue whose delivery times are quantized to integer timesteps.

Editorial extensions

If this is right

  • Synaptic delays can be trained directly and exactly in a sparse, event-based manner, without surrogate gradients or dense temporal traces.
  • The gradient propagation applies to general first-order synaptic models, not just the LIF cells treated by earlier exact-gradient methods.
  • Queue data-structure choice dominates runtime: tree-based and FIFO queues win on CPUs, ring buffers win on GPUs when memory fits, and sorted arrays win on TPU-like accelerators.
  • Lossy queues that deliberately drop spikes offer a simple performance-accuracy trade-off, with drop rates predictable from delay and firing-rate pressure.

Reading between the lines

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

  • If the gradient identity survives discrete-time quantization, it could be applied to any threshold-crossing event in continuous-time neuron models, including subthreshold dynamics and conductance-based synapses, opening a path to exact delay learning in biophysical networks.
  • A key testable extension is a correction term for sub-timestep arrival jitter: real spike events happen between integer timesteps, and the queue's quantization may bias gradients; comparing against fine-step simulations would quantify this.
  • The benchmark pattern suggests future autodifferentiable libraries could explicitly permit primal and tangent data structures to diverge — e.g., bit-packed spike storage for the forward pass and lossy buffers for gradients — which would be a natural next step for the performance-accuracy frontier.
  • The identity itself might be derivable from a more general adjoint or variational principle; if so, it would connect to a wider class of event-based gradient methods.
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

3 major / 5 minor

Summary. The paper presents ADSEQ, an autodifferentiable spike-event queue framework implemented in JAX, aimed at training SNNs with synaptic delays while preserving the sparsity of spike events. The central theoretical section (§3.1.2) derives a gradient-propagation rule through event queues, culminating in Eq. (15): ∂x⁺/∂θ = ∂x⁻/∂θ − (1/τ)∂t_post/∂θ. The paper also describes several queue data structures (Ring, FIFORing, SortedArray, BinaryHeap, lossy variants, etc.) and benchmarks them on CPU, GPU, TPU, and Groq LPU, concluding that queue-choice effects are architecture-dependent and that lossy queues offer a performance/accuracy trade-off.

Significance. If correct and properly discretized, the proposed queue-gradient rule would be a useful bridge between exact event-based SNN gradients (EventProp/DelGrad) and general autodiff frameworks, because it would allow trainable delays in a sparse, model-agnostic way. The benchmark study is broad and contains useful empirical data on data-structure/hardware interactions. Strengths are the clear motivation, the range of implementations, and the cross-platform comparison. These strengths are, however, conditional on the theoretical derivation being sound; as written the derivation has a sign error and the transfer to discrete-time queues is unsubstantiated. The paper claims an open-source implementation and randomized tests, but no code or test results are included in the reviewed manuscript, so I could not verify the implementations.

major comments (3)
  1. [§3.1.2, Eqs. (12)–(15)] There is a sign error in the jump-condition derivative. From Eq. (12) x⁺=x⁻+1 and Eq. (13) ẋ=−x/τ, we have ẋ⁻=−x⁻/τ and ẋ⁺=−(x⁻+1)/τ, hence ẋ⁻−ẋ⁺=+1/τ. Substituting into Eq. (14) yields ∂x⁺/∂θ = ∂x⁻/∂θ + (1/τ)∂t_post/∂θ, opposite to Eq. (15). Since every queue JVP/VJP in the implementation is based on this equation, the sign of all delay-related gradient terms is flipped. The error must be corrected in the derivation, in the custom gradient definitions, and in the validation; a single sign mistake in the central equation is load-bearing.
  2. [§3.1.2–§3.1.3] The derivation is performed for continuous-time event times, but the implemented queues operate on a discrete time grid with delivery times quantized to integer timesteps. The paper does not state how t_post is computed from t_pre and d in the forward pass (rounding? flooring?), nor does it prove that the continuous-time jump condition remains valid for the discretized, queued delivery. This is not a formal detail: if delivery happens at integer steps, the derivative of t_post w.r.t. d is not the smooth 1 used in Eq. (11), and the exactness claim for the discrete-time autodiff implementation is unsupported. A discrete-time adjoint derivation or finite-difference verification is required.
  3. [§4.3 and Table 3] The claim that correctness was verified by randomized pytest-based testing is not backed by any test code, gradient check, or numerical comparison. For a paper whose contribution is exact gradients through queues, the absence of a finite-difference gradient check is a serious omission. The benchmarks cannot distinguish a correct implementation of Eq. (15) from one that implements the wrong sign. Please include the test suite or, at minimum, a table of gradient checks (analytic vs. numerical) for several queue implementations.
minor comments (5)
  1. [Title/Abstract] The arXiv title is 'ADSEQ: ...' while the full text title is 'EventQueues: ...'; unify these for consistency.
  2. [Table 3] Table 3 contains `?` and `M` entries with no legend. Explain what these mean (not measured? out of memory?) and consider reporting failure modes explicitly.
  3. [References] References [8] and [10] are duplicates (Neftci et al. 2018). Several related-work table entries lack clear citation pointers or are ambiguous; please check the table against the reference list.
  4. [Eq. (17)] Eq. (17) uses d y_post/dx, but the differentiation variable appears to be a parameter θ, not x. Align the notation with the surrounding equations.
  5. [Figures 2 and 3] The text refers to panels A/B in Figures 2 and 3, but the captions and figure panels are hard to match as rendered. Please make panel labels explicit in the captions.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the queue-gradient derivation is an application of EventProp's external jump-condition identity to delayed events, and the benchmark conclusions are direct measurements.

full rationale

The central derivation in Sec. 3.1.2 starts from EventProp's published jump-condition equation (Eq. 7, citing [17]) and the standard spike-time derivative (Eq. 6, citing [34]), then applies them to t_post = t_pre + d. Eq. (15) is not assumed as an input; it is derived from Eqs. (12)-(14). One can question the algebra: with x^+ = x^- + 1 and xdot = -x/tau, the paper's claim that (xdot^- - xdot^+) = -1/tau appears inconsistent, since the equations imply +1/tau. This is a mathematical/correctness issue, not circularity: the result is not equivalent to its own inputs by construction, and no fitted parameter is renamed as a prediction. The self-citations in the paper ([5], [27], [38], [39]) are used for hardware background, previously observed compilation artifacts, and prior cell-level gradient-insertion work; none is load-bearing for Eq. (15) or for the queue benchmark conclusions. The queue-vs-platform performance findings are direct measured comparisons, not predictions forced by the derivation. Therefore no significant circularity is present; score 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the EventProp spike-time derivative and an exponential-decay jump condition; no new physical entities are introduced. The discrete-time queue application is an unstated extra assumption, and benchmark workloads introduce hand-chosen hyperparameters.

free parameters (2)
  • benchmark workload settings = λ=400, delay=80 timesteps, 1000 queues, 10k recurrent neurons
    Chosen as representative defaults in Section 4.2; results are specific to these settings and no sensitivity analysis is reported.
  • queue capacities in benchmark variants = FIFORing[2/3/4], LossyRing[2/3/4], BinaryHeap[7], SortedArray[4]
    Capacities are hand-picked benchmark variants; performance conclusions depend on these choices.
assumptions (4)
  • domain assumption Spike-time derivative formula dt_pre/dθ = −(1/v̇_pre) dv_pre/dθ (Eq. 10)
    Used without proof; requires v̇_pre ≠ 0 at threshold and a smooth voltage crossing. Cited to EventProp and reference [34].
  • domain assumption Postsynaptic state variables obey first-order exponential decay ẋ = −x/τ with a constant jump x⁺ = x⁻ + 1 on spike arrival (Eqs. 12–13)
    The derivation of Eq. (15) depends entirely on this restrictive form; it does not cover arbitrary biophysical synapses despite claiming 'the general case'.
  • ad hoc to paper Continuous-time delay derivatives can be applied unchanged to a discrete-time queue with timestep-quantized delivery times
    The queue delivers at integer timesteps, so ∂ceil(t_post/Δt)/∂θ is not generally equal to ∂t_post/∂θ; no analysis or correction is provided.
  • domain assumption Spike trains are sparse enough that dropping/merging events under LTI preserves correctness for the studied workloads
    Lossy queues drop spikes; drop rates are computed for Poisson input only, not for the recurrent training benchmarks where the impact is unquantified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ADSEQ: A delay-aware autograd-compatible framework for spike-event delivery in SNNs." pith.science (2026). https://pith.science/paper/WJLK7ST6

@misc{pith2026251205906,
  author       = {Pith},
  title        = {Pith review of: ADSEQ: A delay-aware autograd-compatible framework for spike-event delivery in SNNs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WJLK7ST6}},
  note         = {Machine review of arXiv:2512.05906}
}
read the original abstract

Spiking neural networks (SNNs), central to computational neuroscience and neuromorphic machine learning (ML), require efficient simulation and gradient-based training. While AI accelerators offer promising speedups, gradient-based SNNs typically implement sparse spike events using dense, memory-heavy data-structures. Existing exact gradient methods lack generality, and current simulators often omit or inefficiently handle delayed spikes. We address this by deriving gradient computation through spike event queues, including delays, and implementing this into memory-efficient, autodifferentiable spike event queues (ADSEQ). These are benchmarked across CPU, GPU, TPU, and LPU platforms. We find that queue design strongly shapes performance. CPUs, as expected, perform well with traditional tree-based or FIFO implementations, while GPUs excel with ring buffers for smaller simulations, yet under higher memory pressure prefer more sparse data-structures. TPUs seem to favor an implementation based on sorting intrinsics. Selective spike dropping provides a simple performance-accuracy trade-off, which could be enhanced by future autograd frameworks adapting diverging primal/tangent data-structures.

Figures

Figures reproduced from arXiv: 2512.05906 by the authors.

Figure 1
Figure 1. Study setup. We implement spike event queues that support gradients and bench [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Scaling of event queues with batch size (queue count) [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. A) Scaling of event queues with queue capacity. B) Architecture Effect [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Spike drop rates for different lossy queue implementations given an incoming [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 1 canonical work pages

  1. [1]

    Cambridge University Press, 2006

    Nicholas T Carnevale and Michael L Hines.The NEURON book. Cambridge University Press, 2006

    2006

  2. [2]

    Supervised learning in spiking neural networks with synaptic delay-weight plasticity.Neurocomputing, 409:103–118, 2020

    Malu Zhang, Jibin Wu, Ammar Belatreche, Zihan Pan, Xiurui Xie, Yansong Chua, Guoqi Li, Hong Qu, and Haizhou Li. Supervised learning in spiking neural networks with synaptic delay-weight plasticity.Neurocomputing, 409:103–118, 2020

    2020

  3. [3]

    Homeostatic bidirec- tional plasticity in upbound and downbound micromodules of the olivocerebellar system

    El´ ıas M Fern´ andez Santoro, Lennart PL Landsmeer, Said Hamdioui, Christos Stry- dis, Chris I De Zeeuw, Aleksandra Badura, and Mario Negrello. Homeostatic bidirec- tional plasticity in upbound and downbound micromodules of the olivocerebellar system. bioRxiv, pages 2025–01, 2025

    2025

  4. [4]

    The elements of differentiable programming

    Mathieu Blondel and Vincent Roulet. The elements of differentiable programming. arXiv preprint arXiv:2403.14606, 2024

    arXiv 2024

  5. [5]

    Trick- ing ai chips into simulating the human brain: A detailed performance analysis.Neuro- computing, 598:127953, 2024

    Lennart PL Landsmeer, Max CW Engelen, Rene Miedema, and Christos Strydis. Trick- ing ai chips into simulating the human brain: A detailed performance analysis.Neuro- computing, 598:127953, 2024

    2024

  6. [6]

    A differentiable brain simulator bridging brain simulation and brain- inspired computing.arXiv preprint arXiv:2311.05106, 2023

    Chaoming Wang, Tianqiu Zhang, Sichao He, Yifeng Gong, Hongyaoxing Gu, Shangyang Li, and Si Wu. A differentiable brain simulator bridging brain simulation and brain- inspired computing.arXiv preprint arXiv:2311.05106, 2023

    arXiv 2023

  7. [7]

    Kadhim, Matthijs Pals, Jonas Beck, Ziwei Huang, Manuel Gloeckler, Janne K

    Michael Deistler, Kyra L. Kadhim, Matthijs Pals, Jonas Beck, Ziwei Huang, Manuel Gloeckler, Janne K. Lappalainen, Cornelius Schr¨ oder, Philipp Berens, Pedro J. Gon¸ calves, and Jakob H. Macke. Differentiable simulation enables large-scale train- ing of detailed biophysical models of neural dynamics.bioRxiv, 2024. doi: 10.1101/2024.08.21.608979

    work page doi:10.1101/2024.08.21.608979 2024

  8. [8]

    Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks.IEEE Signal Processing Magazine, 36(6):51–63, 2019. 20

    2019

Show all 40 references
  1. [9]

    Saunders, Hassaan Khan, Devdhar Patel, Darpan T

    Hananel Hazan, Daniel J. Saunders, Hassaan Khan, Devdhar Patel, Darpan T. Sang- havi, Hava T. Siegelmann, and Robert Kozma. Bindsnet: A machine learning- oriented spiking neural networks library in python.Frontiers in Neuroinfor- matics, 12:89, 2018. ISSN 1662-5196. doi: 10.33...

    2018

  2. [10]

    Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks.IEEE Signal Processing Magazine, 36(6):51–63, 2019

    2019

  3. [11]

    Nengodl: Combining deep learning and neuromorphic modelling methods.Neuroinformatics, 17(4):611–628, 2019

    Daniel Rasmussen. Nengodl: Combining deep learning and neuromorphic modelling methods.Neuroinformatics, 17(4):611–628, 2019

    2019

  4. [12]

    Muir, Felix Bauer, and Philipp Weidel

    Dylan R. Muir, Felix Bauer, and Philipp Weidel. Rockpool documentaton, September

  5. [13]

    Spikingjelly

    Wei Fang, Yanqi Chen, Jianhao Ding, Ding Chen, Zhaofei Yu, Huihui Zhou, Timoth´ ee Masquelier, Yonghong Tian, and other contributors. Spikingjelly. https://github.com/fangwei123456/spikingjelly, 2020

    2020

  6. [14]

    Norse - A deep learning library for spiking neural networks, January 2021

    Christian Pehle and Jens Egholm Pedersen. Norse - A deep learning library for spiking neural networks, January 2021. URLhttps://doi.org/10.5281/zenodo.4422025. Documentation: https://norse.ai/docs/

    2021 doi

  7. [15]

    Training spik- ing neural networks using lessons from deep learning.arXiv preprint arXiv:2109.12894, 2021

    Jason K Eshraghian, Max Ward, Emre Neftci, Xinxin Wang, Gregor Lenz, Girish Dwivedi, Mohammed Bennamoun, Doo Seok Jeong, and Wei D Lu. Training spik- ing neural networks using lessons from deep learning.arXiv preprint arXiv:2109.12894, 2021

    2021 arXiv

  8. [16]

    Spyx: A library for just-in-time compiled opti- mization of spiking neural networks.arXiv preprint arXiv:2402.18994, 2024

    Kade M Heckel and Thomas Nowotny. Spyx: A library for just-in-time compiled opti- mization of spiking neural networks.arXiv preprint arXiv:2402.18994, 2024

    2024 arXiv

  9. [17]

    Event-based backpropagation can compute exact gradients for spiking neural networks.Scientific Reports, 11(1):12829, 2021

    Timo C Wunderlich and Christian Pehle. Event-based backpropagation can compute exact gradients for spiking neural networks.Scientific Reports, 11(1):12829, 2021

    2021

  10. [18]

    Delgrad: Exact gradients in spiking networks for learning transmission delays and weights.arXiv preprint arXiv:2404.19165, 2024

    Julian G¨ oltz, Jimmy Weber, Laura Kriener, Peter Lake, Melika Payvand, and Mihai A Petrovici. Delgrad: Exact gradients in spiking networks for learning transmission delays and weights.arXiv preprint arXiv:2404.19165, 2024

    2024 arXiv

  11. [19]

    jaxsnn: Event-driven gradient estimation for analog neuromorphic hardware.arXiv preprint arXiv:2401.16841, 2024

    Eric M¨ uller, Moritz Althaus, Elias Arnold, Philipp Spilger, Christian Pehle, and Jo- hannes Schemmel. jaxsnn: Event-driven gradient estimation for analog neuromorphic hardware.arXiv preprint arXiv:2401.16841, 2024

    2024 arXiv

  12. [20]

    mlgenn: accelerating snn inference using gpu-enabled neural networks.Neuromorphic Computing and Engineering, 2(2):024002, 2022

    James Paul Turner, James C Knight, Ajay Subramanian, and Thomas Nowotny. mlgenn: accelerating snn inference using gpu-enabled neural networks.Neuromorphic Computing and Engineering, 2(2):024002, 2022. 21

    2022

  13. [21]

    Cam- bridge University Press, 2009

    Sanjeev Arora and Boaz Barak.Computational complexity: a modern approach. Cam- bridge University Press, 2009

    2009

  14. [22]

    Arbor—a morphologically-detailed neural net- work simulation library for contemporary high-performance computing architectures

    Nora Abi Akar, Ben Cumming, Vasileios Karakasis, Anne K¨ usters, Wouter Klijn, Alexander Peyser, and Stuart Yates. Arbor—a morphologically-detailed neural net- work simulation library for contemporary high-performance computing architectures. In2019 27th euromicro internationa...

    2019

  15. [23]

    The groq software- defined scale-out tensor streaming multiprocessor: From chips-to-systems architectural overview

    Dennis Abts, John Kim, Garrin Kimmell, Matthew Boyd, Kris Kang, Sahil Parmar, Andrew Ling, Andrew Bitar, Ibrahim Ahmed, and Jonathan Ross. The groq software- defined scale-out tensor streaming multiprocessor: From chips-to-systems architectural overview. In2022 IEEE Hot Chips ...

    2022

  16. [24]

    Eden: a high-performance, general-purpose, neuroml-based neural simula- tor.Frontiers in neuroinformatics, 16:724336, 2022

    Sotirios Panagiotou, Harry Sidiropoulos, Dimitrios Soudris, Mario Negrello, and Chris- tos Strydis. Eden: a high-performance, general-purpose, neuroml-based neural simula- tor.Frontiers in neuroinformatics, 16:724336, 2022

    2022

  17. [25]

    Efficient event-based delay learning in spiking neural networks.arXiv preprint arXiv:2501.07331, 2025

    Bal´ azs M´ esz´ aros, James C Knight, and Thomas Nowotny. Efficient event-based delay learning in spiking neural networks.arXiv preprint arXiv:2501.07331, 2025

    2025 arXiv

  18. [26]

    MIT press, 2007

    Eugene M Izhikevich.Dynamical systems in neuroscience. MIT press, 2007

    2007

  19. [27]

    Gradient diffusion: Enhancing multicompartmental neuron models for gradient-based self-tuning and homeostatic control.arXiv preprint arXiv:2412.07327, 2024

    Lennart PL Landsmeer, Mario Negrello, Said Hamdioui, and Christos Strydis. Gradient diffusion: Enhancing multicompartmental neuron models for gradient-based self-tuning and homeostatic control.arXiv preprint arXiv:2412.07327, 2024

    2024

  20. [28]

    Fast and energy-efficient neuromorphic deep learning with first-spike times.Nature machine intelligence, 3(9):823–835, 2021

    Julian G¨ oltz, Laura Kriener, Andreas Baumbach, Sebastian Billaudelle, Oliver Bre- itwieser, Benjamin Cramer, Dominik Dold, Akos Ferenc Kungl, Walter Senn, Johannes Schemmel, et al. Fast and energy-efficient neuromorphic deep learning with first-spike times.Nature machine int...

    2021

  21. [29]

    Nest: An environment for neural systems simulations.Forschung und wisschenschaftliches Rechnen, Beitr¨ age zum Heinz-Billing- Preis, 58:43–70, 2001

    Markus Diesmann and Marc-Oliver Gewaltig. Nest: An environment for neural systems simulations.Forschung und wisschenschaftliches Rechnen, Beitr¨ age zum Heinz-Billing- Preis, 58:43–70, 2001

    2001

  22. [30]

    Brian 2- the second coming: spiking neural network simulation in python with code generation

    Marcel Stimberg, Dan FM Goodman, Victor Benichoux, and Romain Brette. Brian 2- the second coming: spiking neural network simulation in python with code generation. BMC neuroscience, 14:1–1, 2013

    2013

  23. [31]

    Genn: a code generation framework for accelerated brain simulations.Scientific reports, 6(1):18854, 2016

    Esin Yavuz, James Turner, and Thomas Nowotny. Genn: a code generation framework for accelerated brain simulations.Scientific reports, 6(1):18854, 2016. 22

    2016

  24. [32]

    Loss shaping enhances exact gradient learning with eventprop in spiking neural networks.Neuromorphic Computing and Engineering, 5(1):014001, 2025

    Thomas Nowotny, James P Turner, and James C Knight. Loss shaping enhances exact gradient learning with eventprop in spiking neural networks.Neuromorphic Computing and Engineering, 5(1):014001, 2025

    2025

  25. [33]

    Fast simulations of highly-connected spiking cortical models using gpus.Frontiers in Computational Neuroscience, 15:627620, 2021

    Bruno Golosio, Gianmarco Tiddia, Chiara De Luca, Elena Pastorelli, Francesco Simula, and Pier Stanislao Paolucci. Fast simulations of highly-connected spiking cortical models using gpus.Frontiers in Computational Neuroscience, 15:627620, 2021

    2021

  26. [34]

    A proof of a key formula in the error- backpropagation learning algorithm for multiple spiking neural networks

    Wenyu Yang, Dakun Yang, and Yetian Fan. A proof of a key formula in the error- backpropagation learning algorithm for multiple spiking neural networks. In Zhigang Zeng, Yangmin Li, and Irwin King, editors,Advances in Neural Networks – ISNN 2014, pages 19–26, Cham, 2014. Spring...

    2014

  27. [35]

    Bgpq: A heap-based priority queue design for gpus

    Yanhao Chen, Fei Hua, Yuwei Jin, and Eddy Z Zhang. Bgpq: A heap-based priority queue design for gpus. InProceedings of the 50th International Conference on Parallel Processing, pages 1–10, 2021

    2021

  28. [36]

    Tpu v4: An optically reconfigurable supercomputer for machine learning with hardware support for embeddings

    Norm Jouppi, George Kurian, Sheng Li, Peter Ma, Rahul Nagarajan, Lifeng Nai, Nis- hant Patil, Suvinay Subramanian, Andy Swing, Brian Towles, et al. Tpu v4: An optically reconfigurable supercomputer for machine learning with hardware support for embeddings. InProceedings of the...

    2023

  29. [37]

    Think fast: A tensor streaming processor (tsp) for accelerating deep learning workloads

    Dennis Abts, Jonathan Ross, Jonathan Sparling, Mark Wong-VanHaren, Max Baker, Tom Hawkins, Andrew Bell, John Thompson, Temesghen Kahsai, Garrin Kimmell, et al. Think fast: A tensor streaming processor (tsp) for accelerating deep learning workloads. In2020 ACM/IEEE 47th Annual ...

    2020

  30. [38]

    Huma: Hetero- geneous, ultra low-latency model accelerator for the virtual brain on a versal adaptive soc

    Amirreza Movahedin, Lennart PL Landsmeer, and Christos Strydis. Huma: Hetero- geneous, ultra low-latency model accelerator for the virtual brain on a versal adaptive soc. InProceedings of the 2025 ACM/SIGDA International Symposium on Field Pro- grammable Gate Arrays, pages 223...

    2025

  31. [39]

    Efficient and realistic brain simulation: A review and design guide for memristor-based approaches.Advanced Materials Tech- nologies, page e01587, 2025

    Lennart Paul Liong Landsmeer, Muhammad Ali Siddiqi, Heba Abunahla, Mario Ne- grello, Said Hamdioui, and Christos Strydis. Efficient and realistic brain simulation: A review and design guide for memristor-based approaches.Advanced Materials Tech- nologies, page e01587, 2025. 23

    2025

  32. [2019]

    URLhttps://doi.org/10.5281/zenodo.3773845

    doi

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.