pith. sign in

arxiv: 2605.18003 · v1 · pith:4T56I574new · submitted 2026-05-18 · 💻 cs.NE · cs.AI

Spiker-LL: An Energy-Efficient FPGA Accelerator Enabling Adaptive Local Learning in Spiking Neural Networks

Alessio Caviglia , Filippo Marostica , Alessandro Savino , Stefano Di Carlo This is my paper

Pith reviewed 2026-05-20 00:06 UTC · model grok-4.3

classification 💻 cs.NE cs.AI
keywords FPGA acceleratorSpiking Neural NetworksSTSF local learningEnergy efficiencyEdge deploymentOnline learningSNN accelerator
0
0 comments X

The pith

SPIKER-LL extends an FPGA SNN accelerator with on-device learning via the STSF rule.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper presents SPIKER-LL, which adds support for the STSF local learning rule to the existing Spiker+ FPGA architecture for spiking neural networks. The extensions allow both inference and online learning to run on the same hardware with only modest added cost. Results on MNIST, Fashion-MNIST, and DIGITS reach up to 93 percent accuracy, sub-millisecond latency, and less than 0.1 millijoules per inference. The design stays free of DSP blocks and scales for edge FPGA use, which matters for placing adaptive models on low-power devices without constant cloud retraining.

Core claim

Through targeted microarchitectural extensions to the Spiker+ inference architecture, SPIKER-LL enables efficient support for the STSF local learning rule. It performs inference and online learning with minimal overhead. Across MNIST, F-MNIST, and DIGITS, the design achieves up to 93% accuracy, sub-millisecond latency, and less than 0.1 mJ per inference while remaining DSP-free and highly scalable for edge-FPGA deployments.

What carries the argument

Microarchitectural extensions to the Spiker+ datapath that map the STSF local learning rule for online weight updates with low resource overhead.

If this is right

  • Enables on-device adaptation for spiking networks on edge FPGAs without external training hardware.
  • Maintains sub-millisecond inference while supporting continuous learning at under 0.1 mJ per sample.
  • Remains DSP-free, allowing deployment on resource-constrained FPGA families.
  • Delivers competitive accuracy on standard image datasets using a single unified accelerator.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same microarchitectural approach could support other local learning rules in SNNs beyond STSF.
  • This points toward fully on-chip training for dynamic edge environments where data changes over time.
  • It may reduce reliance on periodic cloud-based retraining for deployed neuromorphic systems.

Load-bearing premise

The STSF local learning rule can be added to the existing Spiker+ hardware path with only modest resource overhead and without degrading accuracy or energy efficiency beyond the reported bounds.

What would settle it

An FPGA implementation of SPIKER-LL that measures energy consumption above 0.1 mJ per inference or accuracy below 93% on MNIST when using the STSF rule would show the efficiency claims do not hold.

Figures

Figures reproduced from arXiv: 2605.18003 by Alessandro Savino, Alessio Caviglia, Filippo Marostica, Stefano Di Carlo.

Figure 1
Figure 1. Figure 1: Global view of the SPIKER-LL architecture. SPIKER-LL extends the Spiker+ inference-oriented architecture with additional local and global structures that implement and coordinate the on-device learning process. synapse, the weight update combines three quantities available in hardware: (i) the pre-synaptic spike spre(t), (ii) the post￾synaptic spike spost(t), and (iii) a global modulatory signal Φ(t) distr… view at source ↗
Figure 2
Figure 2. Figure 2: Envisioned deployment of SPIKER-LL in the field with different operating modes. B. SPIKER-LL deployment environment Incorporating local learning into a hardware accelerator is not sufficient without a deployment workflow, illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
read the original abstract

Deploying adaptive intelligence at the edge remains challenging due to the high computational and energy cost of training neural models. Spiking Neural Networks (SNNs) offer a promising alternative, but enabling on-device learning requires hardware-algorithm co-design. This paper presents SPIKER-LL, an FPGA-based SNN accelerator that extends the open-source Spiker+ inference architecture with efficient support for the STSF local learning rule. Through targeted microarchitectural extensions, SPIKER-LL performs inference and online learning with minimal overhead. Across MNIST, F-MNIST, and DIGITS, it achieves up to 93% accuracy, sub-millisecond latency, and less than 0.1 mJ per inference, while remaining DSP-free and highly scalable for edge-FPGA deployments.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 3 minor

Summary. The paper presents SPIKER-LL, an FPGA-based SNN accelerator extending the open-source Spiker+ inference architecture with microarchitectural support for the STSF local learning rule. It enables combined inference and online learning with low overhead, reporting up to 93% accuracy, sub-millisecond latency, and <0.1 mJ per inference on MNIST, Fashion-MNIST, and DIGITS while remaining DSP-free and scalable for edge deployments.

Significance. If the empirical results hold, the work provides a concrete hardware demonstration of energy-efficient on-device learning for SNNs, which is valuable for edge-AI applications. Strengths include the DSP-free implementation, reported scalability, and internal consistency between resource tables, latency, and energy measurements. The focus on minimal-overhead mapping of a local learning rule to an existing datapath addresses a practical gap in SNN hardware.

major comments (1)
  1. [Experimental results] Experimental results section: Accuracy, latency, and energy figures are reported without error bars, number of independent runs, or explicit dataset split details (e.g., train/test ratios or validation strategy). This makes it difficult to assess statistical robustness of the 'up to 93%' claim and the overhead comparison to the baseline Spiker+ design.
minor comments (3)
  1. [Figure 3] Figure 3 (resource utilization): The bar chart for LUT/BRAM usage in learning vs. inference mode would benefit from explicit numerical labels and a direct side-by-side comparison column for the unmodified Spiker+ on the same FPGA device.
  2. [Section 3.2] Section 3.2: The description of STSF rule integration could include a short pseudocode snippet or timing diagram showing the additional control signals and memory accesses required for weight updates.
  3. [Related work] Related work: A brief quantitative comparison table against other recent FPGA SNN accelerators (e.g., those supporting online learning) would better contextualize the energy and DSP-free advantages.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive assessment of SPIKER-LL and the recommendation for minor revision. We address the major comment on the experimental results section below.

read point-by-point responses

  1. Referee: Experimental results section: Accuracy, latency, and energy figures are reported without error bars, number of independent runs, or explicit dataset split details (e.g., train/test ratios or validation strategy). This makes it difficult to assess statistical robustness of the 'up to 93%' claim and the overhead comparison to the baseline Spiker+ design.

    Authors: We agree that additional details would improve the clarity and allow better assessment of statistical robustness. In the revised manuscript we will explicitly state the dataset splits employed (standard 60,000/10,000 train/test partitions for MNIST and Fashion-MNIST together with the predefined split for DIGITS). We will also report the number of independent training runs performed with different random seeds and include mean accuracy with standard-deviation error bars. For latency and energy we will clarify that measurements are deterministic for a given configuration on the FPGA and present averages over repeated inferences. These changes will strengthen both the 'up to 93%' claim and the overhead comparison against Spiker+. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The manuscript is an empirical hardware implementation and measurement study that extends the open-source Spiker+ architecture with microarchitectural support for the STSF local learning rule. All headline results (accuracy up to 93%, sub-millisecond latency, <0.1 mJ per inference, DSP-free operation) are obtained from direct FPGA synthesis, place-and-route, and runtime measurements on MNIST, F-MNIST, and DIGITS; no equations, fitted parameters, or derivation steps are presented that reduce by construction to the reported numbers or to prior self-citations. The central claims rest on independently verifiable resource utilization tables, power measurements, and accuracy benchmarks rather than any self-referential mapping or uniqueness theorem.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The paper introduces no new mathematical axioms, free parameters, or invented physical entities; its claims rest on standard digital design assumptions and the correctness of the underlying Spiker+ open-source components.

pith-pipeline@v0.9.0 · 5671 in / 1225 out tokens · 45873 ms · 2026-05-20T00:06:29.629659+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages · 1 internal anchor

  1. [1]

    Green ai,

    R. Schwartz, J. Dodge, N. A. Smith, and O. Etzioni, “Green ai,” Communications of the ACM, vol. 63, no. 12, pp. 54–63, 2020

    work page 2020

  2. [2]

    Networks of spiking neurons: the third generation of neural network models,

    W. Maass, “Networks of spiking neurons: the third generation of neural network models,”Neural networks, vol. 10, no. 9, pp. 1659–1671, 1997

    work page 1997

  3. [3]

    Energy-efficient digital design: A comparative study of event-driven and clock-driven spiking neurons,

    F. Marostica, A. Carpegna, A. Savino, and S. D. Carlo, “Energy-efficient digital design: A comparative study of event-driven and clock-driven spiking neurons,” in2025 IEEE Computer Society Annual Symposium on VLSI (ISVLSI), vol. 1, 2025, pp. 1–6

    work page 2025

  4. [4]

    Backpropagation through time: what it does and how to do it,

    P. Werbos, “Backpropagation through time: what it does and how to do it,”Proceedings of the IEEE, vol. 78, no. 10, pp. 1550–1560, 1990

    work page 1990

  5. [5]

    Training spiking neural networks using lessons from deep learning,

    J. K. Eshraghian, M. Ward, E. O. Neftci, X. Wang, G. Lenz, G. Dwivedi, M. Bennamoun, D. S. Jeong, and W. D. Lu, “Training spiking neural networks using lessons from deep learning,”Proceedings of the IEEE, vol. 111, no. 9, pp. 1016–1054, 2023

    work page 2023

  6. [6]

    Spiker+: a framework for the generation of efficient Spiking Neural Networks FPGA accelerators for inference at the edge,

    A. Carpegna, A. Savino, and S. D. Carlo, “Spiker+: a framework for the generation of efficient Spiking Neural Networks FPGA accelerators for inference at the edge,”IEEE Transactions on Emerging Topics in Computing, pp. 1–15, 2024

    work page 2024

  7. [7]

    Spiker: an fpga-optimized hardware accelerator for spiking neural networks,

    A. Carpegna, A. Savino, and S. Di Carlo, “Spiker: an fpga-optimized hardware accelerator for spiking neural networks,” in2022 IEEE Com- puter Society Annual Symposium on VLSI (ISVLSI), 2022, pp. 14–19

    work page 2022

  8. [8]

    Sfatti: Spiking fpga accelerator for temporal task-driven inference - a case study on mnist,

    A. Caviglia, F. Marostica, A. Carpegna, A. Savino, and S. Di Carlo, “Sfatti: Spiking fpga accelerator for temporal task-driven inference - a case study on mnist,” in2025 IEEE International Conference on Image Processing Workshops (ICIPW), 2025, pp. 59–64

    work page 2025

  9. [9]

    STSF: Spiking Time Sparse Feedback Learning for Spiking Neural Networks,

    P. He, R. Xiao, C. Tang, S. Huang, J. Lv, and H. Tang, “STSF: Spiking Time Sparse Feedback Learning for Spiking Neural Networks,”IEEE Transactions on Neural Networks and Learning Systems, pp. 1–14, 2025

    work page 2025

  10. [10]

    Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based opti- mization to spiking neural networks,

    E. O. Neftci, H. Mostafa, and F. Zenke, “Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based opti- mization to spiking neural networks,”IEEE Signal Processing Magazine, vol. 36, no. 6, pp. 51–63, 2019

    work page 2019

  11. [11]

    A solution to the learning dilemma for recurrent networks of spiking neurons,

    G. Bellec, F. Scherr, A. Subramoney, E. Hajek, D. Salaj, R. Legenstein, and W. Maass, “A solution to the learning dilemma for recurrent networks of spiking neurons,”Nature communications, vol. 11, no. 1, p. 3625, 2020

    work page 2020

  12. [12]

    Synaptic plasticity dynamics for deep continuous local learning (decolle),

    J. Kaiser, H. Mostafa, and E. Neftci, “Synaptic plasticity dynamics for deep continuous local learning (decolle),”Frontiers in Neuroscience, vol. V olume 14 - 2020, 2020

    work page 2020

  13. [13]

    Loihi: A neuromorphic manycore processor with on-chip learning,

    M. Davies, N. Srinivasa, T.-H. Lin, G. Chinya, Y . Cao, S. H. Choday, G. Dimou, P. Joshi, N. Imam, S. Jain, Y . Liao, C.-K. Lin, A. Lines, R. Liu, D. Mathaikutty, S. McCoy, A. Paul, J. Tse, G. Venkataramanan, Y .-H. Weng, A. Wild, Y . Yang, and H. Wang, “Loihi: A neuromorphic manycore processor with on-chip learning,”IEEE Micro, vol. 38, no. 1, pp. 82–99, 2018

    work page 2018

  14. [14]

    ReckOn: A 28nm Sub-mm2 Task-Agnostic Spiking Recurrent Neural Network Processor Enabling On-Chip Learn- ing over Second-Long Timescales,

    C. Frenkel and G. Indiveri, “ReckOn: A 28nm Sub-mm2 Task-Agnostic Spiking Recurrent Neural Network Processor Enabling On-Chip Learn- ing over Second-Long Timescales,” in2022 IEEE International Solid- State Circuits Conference (ISSCC), vol. 65, 2022, pp. 1–3

    work page 2022

  15. [15]

    Morphbungee-tiny: An ultra-low-cost fpga implementation of neuromorphic architecture with high-accuracy on-chip e-prop learn- ing,

    Z. Zhong, J. Yang, H. Wang, Y . Jiang, Y . Lin, P. Gan, L. Liu, L. Liang, and C. Shi, “Morphbungee-tiny: An ultra-low-cost fpga implementation of neuromorphic architecture with high-accuracy on-chip e-prop learn- ing,”IEEE Transactions on Circuits and Systems II: Express Briefs, pp. 1–1, 2025

    work page 2025

  16. [16]

    An edge neuromorphic processor with high-accuracy on-chip aggregate-label learning,

    G. Chen, M. Chen, T. Wang, H. Wang, X. Fu, Y . Lin, L. Liu, and C. Shi, “An edge neuromorphic processor with high-accuracy on-chip aggregate-label learning,”IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 72, no. 3, pp. 509–513, 2025

    work page 2025

  17. [17]

    Triplebrain: A compact neuromorphic hardware core with fast on-chip self-organizing and reinforcement spike-timing dependent plasticity,

    H. Wang, Z. He, T. Wang, J. He, X. Zhou, Y . Wang, L. Liu, N. Wu, M. Tian, and C. Shi, “Triplebrain: A compact neuromorphic hardware core with fast on-chip self-organizing and reinforcement spike-timing dependent plasticity,”IEEE Transactions on Biomedical Circuits and Systems, vol. 16, no. 4, pp. 636–650, 2022

    work page 2022

  18. [18]

    Sc-plr: An approximate spiking neural network accelerator with on-chip predictive learning rule,

    W. Liu, S. Xiao, Y . Liu, and Z. Yu, “Sc-plr: An approximate spiking neural network accelerator with on-chip predictive learning rule,”IEEE Transactions on Biomedical Circuits and Systems, vol. 18, no. 5, pp. 1156–1165, 2024

    work page 2024

  19. [19]

    An FPGA-Based Co- Processor for Spiking Neural Networks with On-Chip STDP-Based Learning,

    N. N. Thao N., B. Veeravalli, and X. Fong, “An FPGA-Based Co- Processor for Spiking Neural Networks with On-Chip STDP-Based Learning,” in2022 IEEE International Symposium on Circuits and Systems (ISCAS), 2022, pp. 2157–2161

    work page 2022

  20. [20]

    Spike-Train Level Direct Feedback Alignment: Sidestepping Backpropagation for On-Chip Training of Spiking Neural Nets,

    J. Lee, R. Zhang, W. Zhang, Y . Liu, and P. Li, “Spike-Train Level Direct Feedback Alignment: Sidestepping Backpropagation for On-Chip Training of Spiking Neural Nets,”Frontiers in Neuroscience, vol. 14, 2020

    work page 2020

  21. [21]

    A low cost neuromorphic learning engine based on a high performance supervised SNN learning algorithm,

    A. Siddique, M. I. Vai, and S. H. Pun, “A low cost neuromorphic learning engine based on a high performance supervised SNN learning algorithm,”Scientific Reports, vol. 13, no. 1, p. 6280, 2023

    work page 2023

  22. [22]

    The mnist database of handwritten digit images for machine learning research [best of the web],

    L. Deng, “The mnist database of handwritten digit images for machine learning research [best of the web],”IEEE Signal Processing Magazine, vol. 29, no. 6, pp. 141–142, 2012

    work page 2012

  23. [23]

    Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms

    H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,”arXiv preprint arXiv:1708.07747, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  24. [24]

    Optical Recognition of Handwritten Digits,

    E. Alpaydin and C. Kaynak, “Optical Recognition of Handwritten Digits,” UCI Machine Learning Repository, 1998

    work page 1998